doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pcbi.1004176 | Extensive Decoupling of Metabolic Genes in Cancer | Tumorigenesis requires the re-organization of metabolism to support malignant proliferation. We examine how the altered metabolism of cancer cells is reflected in the rewiring of co-expression patterns among metabolic genes. Focusing on breast and clear-cell kidney tumors, we report the existence of key metabolic genes which act as hubs of differential co-expression, showing significantly different co-regulation patterns between normal and tumor states. We compare our findings to those from classical differential expression analysis, and counterintuitively observe that the extent of a gene's differential co-expression only weakly correlates with its differential expression, suggesting that the two measures probe different features of metabolism. Focusing on this discrepancy, we use changes in co-expression patterns to highlight the apparent loss of regulation by the transcription factor HNF4A in clear cell renal cell carcinoma, despite no differential expression of HNF4A. Finally, we aggregate the results of differential co-expression analysis into a Pan-Cancer analysis across seven distinct cancer types to identify pairs of metabolic genes which may be recurrently dysregulated. Among our results is a cluster of four genes, all components of the mitochondrial electron transport chain, which show significant loss of co-expression in tumor tissue, pointing to potential mitochondrial dysfunction in these tumor types.
| The metabolism of malignant tumors is deranged. The transition from healthy to cancerous state involves, among other factors, the transcriptional coordination of genes spread throughout the cell’s metabolic pathways. An examination of this multivariate regulatory effort can offer insights which may remain hidden from analyses focusing on a single gene in isolation. Such an analysis is particularly relevant for metabolic networks, whose constituent enzymes are fundamentally linked through their common utilization of a limited pool of substrates. Here, we examine the extent to which altered metabolism is reflected in the co-expression patterns of genes, shedding light on the differential regulation of metabolic genes within tumors. We study patterns of differential co-expression across metabolic pathways in both breast and kidney tumors, and integrate regulatory information to study the drivers of these changes. Among the results of our analysis is the apparent dsyregulation of genes controlled by HNF4A in clear-cell kidney tumors. Finally, by combining the results of our analyses across seven different tissues, we identify the recurrent decoupling of a set of mitochondrial genes, pointing to possible mitochondrial dysfunction in these cancers.
| All cellular events, from the transduction of signals to the translation of nucleic acids, rely on the interaction of molecular entities. Indeed, one may argue that the fundamental unit of a biological network is not its constituent components (e.g. proteins or genes), but rather the edges representing the interactions between them. Then, it follows that the manifestation of disease, of a deranged phenotype of this network, should be evident by observing changes in the wiring and activity of these edges.
Here, we study the interactions between pairs of genes encoding metabolic enzymes, and how these interactions change in the course of transformation of normal cells to malignant tumor. This notion of studying “interactions” is particularly important for understanding the network of coupled enzymatic reactions which constitute metabolism. It is well-known that tumors, which are under strong selection for proliferative capacity, must re-organize their metabolism in order to deliver the precursors and energy needed to grow as quickly as possible. Otto Warburg published a series of key findings highlighting a fundamental dysregulation in glycolytic metabolism in cancer, whereby cancer cells metabolized high levels of glucose to lactate [1]. Some of the earliest chemotherapies (e.g. methotrexate) targeted a metabolic phenotype which distinguished tumor from normal tissue. In recent years, an invigorated field has identified a number of distinct “metabolic lesions” in various tumors, including, for example, the preferential expression of PKM2 [2] and the presence of an oncometabolite, 2-hydroxyglutarate, in cells with activating IDH1 and IDH2 mutations [3].
Our use of the term “interaction” above is loose: for the purposes of our study, which focuses on the analysis of gene expression data, we say that two metabolic genes putatively interact if we observe they are co-expressed. This co-expression may occur by chance, or as a result of co-regulation by a set of common factors. Furthermore, while strong co-expression is more likely to occur between proteins which physically interact with each other, the highly connected structure of the metabolic network suggests that even genes residing in opposing corners of metabolism may be coupled to each other. Regardless of the source of co-expression, our goal is to identify regions of the metabolic network whose co-expression patterns appear fundamentally different between normal and cancerous tissue samples. Put another way, we intentionally search for cases where two genes are co-expressed in one manner in normal tissue, and then co-expressed in an entirely different manner in the tumor tissue. Our approach follows other studies employing techniques to detect so-called “differential co-expression” of genes [4–11].
Differential expression analysis is the standard method for comparing the expression patterns of genes across conditions. Aside from its ubiquitous use in research, several large-scale surveys of differential expression focusing exclusively on metabolic genes in cancer have been completed [12, 13]. In contrast, while a handful of publications have examined differential co-expression in various cancer settings (for example, [9, 14–17]), differential co-expression analysis remains largely absent in most studies of gene expression and (to our knowledge), no survey of differential co-expression among metabolic genes in cancer ahs been undertaken. This is, at least in part, due to the requirement for large sample sizes in order to detect statistically significant differential co-expression patterns. Here, we embark on such a large-scale analysis of RNA-Seq data from 3000 samples of primary tumor and adjacent normal samples from seven distinct tissues, and focus our attention squarely on the expression patterns of 1789 metabolic genes. Among our main findings is the (previously known, see [18], but potentially under-appreciated) observation that genes with strong differential co-expression patterns are not necessarily differential expressed. A relatively large fraction of the genes we identify in our study show no substantial difference in their absolute expression between tumor and normal tissue, but nevertheless exhibit recurrent differential co-expression.
The results to be presented will encompass a variety of analyses, studying differential co-expression patterns first across two cancer types for which we have the most data available (breast and clear cell renal cell carcinomas, (KIRC)), and then expanding to include five other cancer types (lung, thyroid, prostate, liver, and head and neck), as described in Table 1. In the course of doing so, we propose two simple, but novel, analyses which integrate pathway information to assess the functional role of differentially co-expressed gene pairs. We examine the association between differential co-expression and differential expression, and identify genes which are strongly enriched for one measure but not the other. By leveraging our findings against regulatory (i.e. transcription factor binding) data, we identify transcription factors whose targets are highly enriched for differential co-expression. Among our findings is a previously unreported loss of co-expression between HNF4A, a transcription factor, and its regulatory targets in KIRC. Finally, we leverage the scale of our study to complete a “Pan-Cancer” analysis of differential co-expression, searching for those pairs of metabolic genes which are recurrently differentially co-expressed across multiple cancer types. Our results highlight a small group of four mitochondrial electron transport chain (ETC) genes which are recurrently differentially co-expressed, hinting at a fundamental alteration in the function of the ETC in tumors.
All TCGA expression data were accessed using the Broad Institute Firehose. RSEM-normalized expression was used for the co-expression calculations. Entrez IDs of metabolic genes were extracted from the Recon2 genome scale metabolic network reconstruction [19], and used to extract the corresponding metabolic gene expression data from the TCGA datasets.
We begin by describing the methodology, broadly illustrated in Fig 1, to detect changes in co-expression patterns between normal and tumor samples. After obtaining RNA-Seq data, we calculate the Spearman correlation (a non-parametric measure of the correlation of two random variables employing ranks) of each pair of genes i, j, and record the p-value pij associated with this correlation. These calculations are performed separately for tumor and normal samples. To account for multiple hypothesis testing, we apply the conservative Bonferonni correction [20], yielding corresponding adjusted p-values p ^. The results of these correlation calculations are stored in two matrices, CT and CN (corresponding to tumor and normal samples, respectively), with entries
C i j = { r i j , if p ^ i j < τ 0 , otherwise (1)
Here, τ is a significance threshold for our Bonferroni-corrected p-values. Throughout the manuscript unless otherwise stated, we employ a threshold τ = 1×10−2.
Our goal is to identify significant differences between the strength of co-expression (as quantified by the correlation coefficients) in tumor and normal samples. Such a comparison of sample correlation coefficients must be done with care. In fact, the difference between two correlation coefficients is not sufficient information to determine how often such a difference would appear by chance. We offer an example to illustrate this phenomenon. Very small correlation coefficients (say, r1 = 0.1, r2 = −0.1) may appear in random, uncorrelated data simply by chance. In this case, the difference between the two correlation coefficients (r1−r2 = 0.2) should be categorized as statistically insignificant because it is quite likely to happen by chance. On the other hand, the same difference for two very large correlation coefficients (say, r1 = 0.99, r2 = 0.79) appears less likely to happen by chance; instead, this difference is more likely to arise via the corruption of a nearly perfect correlation by a confounding factor or noise.
The basis of this intuition is that very large correlation coefficients are observed quite rarely by chance. More importantly, the variance of the correlation coefficient estimated from the data (referred to as the sample correlation coefficient, r) depends on the value of the true correlation coefficient underlying the data (referred to as the population correlation coefficient, ρ). In particular, the variance of sample correlation coefficient is approximately [21, 22]
Var ( r ) ∝ ( 1 − ρ 2 ) 2
Thus, as the population correlation coefficient tends to ±1, the variance of the sample correlation coefficient asymptotically approaches zero. This dependence of the variance of r on ρ itself makes it very difficult to carry out hypothesis tests comparing two sample correlation coefficients. A standard method for testing for a difference between correlation coefficients is to employ a transformation to stabilize the variances, making them independent of ρ. Here, we use the Fisher r to z transformation:
z = 1 2 log ( 1 + r 1 − r ) . (2)
The change of variables in Eq (2) is well-known, and has been used in prior work on differential co-expression [7]. When applied to data drawn from a bivariate normal distribution, this transformation yields a quantity which is approximately normally distributed with variance σ 2 = 1 N − 3 independent of the population mean, with N equal to the size of the population. By applying this transformation to our measured correlation coefficients in normal tissue and tumor samples, we are able to apply a Z-test to determine if the correlation coefficients r i j T and r i j N are significantly different. In particular, the quantity (3), which measures the difference between the two transformed correlation coefficients, is approximately normally distributed with mean zero and variance one:
Δ z i j = z T − z N 1 N T − 3 + 1 N N − 3 , (3)
where NT is the number of tumor samples, NN is the number of normal samples, zT is the Fisher-transformed tumor sample correlation coefficient, and zN is the Fisher-transformed normal sample correlation coefficient. Python code for the differential co-expression test is included in S1 Code. Thus, we can associate p-values p i , j z with the Z-test in (3) for each pair of genes i, j. After again correcting pz for multiple hypothesis testing using the Bonferonni correction, we stored the results of our calculations in a matrix D with entries
D i j = { Δ r i j , if p ^ i j z < τ and ( p ^ i j T < τ or p ^ i j N < τ ) 0 , otherwise (4)
where p ^ i j z is the Bonferonni adjusted value of p i j z. The entries of the matrix D correspond to the change in gene co-expression between tumor and normal samples, and will be our main object of study. We emphasize one final, but important, feature of Eq 4: an entry of D is nonzero if and only if that gene pair shows both (1) a significant change in co-expression between tumor and normal samples, and (2) the genes were co-expressed at a statistically significant level in tumor or normal samples (or both). This ensures that those gene pairs which we call differentially co-expressed are also co-expressed at a statistically significant level in at least one group of samples.
We assigned each gene in our study to one or more pathways using the subsystem assignments in the Recon2 human metabolic reconstruction [19]. Then, for each TCGA study, we calculated a score for each pathway i, Ei, using:
E i = ∑ j ∈ P i S j 0 , (5)
where Pi is the set of all genes in pathway i. Thus, Ei counts the total number of dysregulations for all genes in pathway i. We then divided each Ei by the number of genes in pathway i to obtain a normalized pathway score E ^ i. Thus, E ^ i quantifies the differential co-expression of all genes in a pathway, averaged over the number of genes in that pathway. We excluded from our analysis pathways composed of fewer than five genes.
We obtained data on transcription factor targets from the Broad Institute’s MSigDB website [23].
Assuming that a particular regulatory factor has m targets, we calculate the total number of differential co-expression edges in the sub-network composed of only these m gene targets. In this subnetwork, there are t = ( m 2 ) = m × ( m − 1 ) 2 total possible edges. If we see e edges in the true subnetwork, we can calculate the probability that these edges would appear by chance. Given that the probability of a random“differential co-expression edge” in the network is p (e.g. for a Bonferonni-corrected p-value threshold of 1 × 10−2 for detecting differential co-expression, p ≈ 8 × 10−3), the probability of seeing at least e edges by chance is
P = ∑ i = e t ( t i ) i p ( t − i ) 1 − p (6)
A Bonferonni correction is then applied to the vector of p-values for all transcription factor motifs.
With our analytical framework established, our first aim was to assess how pervasive differential co-expression was among metabolic genes in cancer samples. We used the Recon2 human metabolic network reconstruction [19] to identify metabolic genes, and extracted expression corresponding to these genes from the TCGA datasets. We applied the differential co-expression analysis described above to two TCGA studies (breast, BRCA; and clear cell renal cell carcinoma, KIRC) with large numbers of both tumor and normal RNA-Seq samples (106 and 71 normal samples, 914 and 480 primary tumor samples, respectively). Using the list of metabolic genes from Recon2, we were able to extract data for 1,789 unique metabolic genes. We used a strict Bonferonni corrected p-value threshold of 1 × 10−2 to identify pairs of genes which we called differentially co-expressed. Across the total number of pairs of metabolic genes in our dataset (approximately (2 × 103)2/2 = 2 × 106 distinct pairs), we calculated (for each of the two studies) that approximately 2.5 percent of gene pairs were differentially co-expressed. The top differentially co-expressed gene pairs are reported in Tables 2 and 3.
To independently test the extent of differential co-expression in our data, we followed the protocol presented in [17] and completed a permutation test to assess how frequently we would expect the observed changes in correlation coefficients by chance (S1 Fig). In this analysis, we shuffled the labels (e.g. tumor or normal) of all samples, and calculated the difference in correlation coefficients and transformed correlation coefficients in the new, permuted data. This process was repeated 10000 times, and the results aggregated to form a distribution. Inspection of the results confirmed that for a large number of gene pairs, the differences in correlation coefficients were larger in the real data than in the permuted data (S1 Fig). Although it was computationally intractable to complete enough permutations of the data to generate robust p-values (because of the large correction for multiple hypothesis testing), we nevertheless found that 12% of gene pairs showed a higher difference in both (1) tumor and normal correlation coefficients and (2) transformed correlation coefficients than in any of the 10000 permuted data sets. These findings supported our observation of extensive differential co-expression in metabolic genes.
Naturally, we were interested in identifying those genes which were enriched for membership in differentially co-expressed gene pairs. To find these genes, we calculated two “scores” for each gene:
S i 0, the number of differentially co-expressed gene pairs which gene i participates in
S i = − ∑ j : p ^ i j z < τ ln ( p ^ i j z ), a weighted sum of the number of differentially co-expressed pairs gene i participates in
The score S, based on Fisher’s method for combining p-values from independent statistical tests [24], accounted for both the frequency of a gene’s membership in differentially co-expressed pairs, as well as the confidence with which we could claim the gene pair was differentially co-expressed (i.e. by the magnitude of p ^ i − z ). It is important to note that each test of differential co-expression in our dataset is not independent, so we cannot use S as a formal test statistic. However, its use as a measure of the recurrence and magnitude of a gene’s overall differential co-expression is nevertheless useful.
In breast cancer, the top-ranked gene was ACAT1 (Acetyl-CoA acetyltransferase, not be confused with the enzyme acyl-Coenzyme A: cholesterol acyltransferase 1, which is encoded by the gene SOAT1). The enzyme translated from ACAT1 catalyzes the formation of acetoacetyl-CoA, which along with acetyl-CoA is the precursor to 3-hydroxy-3-methylglutaryl-CoA. These two metabolites lie at the beginning of the mevalonate pathway, which generates precusors for cholesterol and steroid biosynthesis. Intriguingly, Freed-Pastor and colleagues [25] recently reported that upregulation of the mevalonate pathway is sufficient and necessary for mutant p53 to have phenotypic effects on cell architecture in mammary tissue. Overexpression of various genes in the mevalonate pathway has also been shown to associate with poor prognosis in breast cancer [26]. Interestingly, ACAT1 is differentially coexpressed with 11 genes for which it is a catalytic partner: ACAA2, DLD, MLYCD, HADHB, HADH, OXCT1, PCCA, PDHA1, PDHB, and ACSS1. A plot of the differences in the correlation of these genes with ACAT1 is in S6 Fig. In many cases, the co-expression patterns show remarkably tight correlations in normal tissue, and these correlations are partially or completely eroded in the tumor samples. Functionally, many of these genes are part of the terminal reactions in glycolysis, lipid biosynthesis and fatty acid oxidation. This loss of co-expression suggests that the flux generated by these pathways is no longer coupled to the flux through ACAT1 in tumor cells.
For KIRC, the highest-scoring differentially co-expressed gene was PSAT1 (phosphoserine aminotransferase 1), a key enzyme in the serine biosynthesis pathway which has already been associated with breast and colorectal cancers before [27, 28], but has not yet been associated with kidney cancer. PSAT1 was differentially co-expressed with 492 other metabolic genes in the dataset, with the strongest signals coming from genes like GATM (glycine aminotransferase), GBA3 (a beta-glucosidase), and SLC10A2 (a bile transporter) (S4 Fig). Because nearly all of the strongest signals came from loss of positive correlation in normal samples, we further identified those genes with which PSAT1 was more strongly co-expressed in tumor samples than in normal samples (S5 Fig). These genes included several galactosidases (GLA, GLB1), glycogen phosphorylase (PYGB), and SLC35A2, which transfers nucleotide sugars into the Golgi body for the purposes of glcosylation. Neither the substrates (3-phosphonoxypyruvate,glutamate) nor the products (phosphoserine, 2-oxoglutarate) of PSAT1 participate in the glycogenolysis pathway, suggesting that the positive correlation between PSAT1 and glycogen breakdown in tumors may be the result of indirect couplings. In particular, it is possible that the overexpression of glycogen phosphorylase may liberate carbon units to be shunted from glycolysis into the serine biosynthesis pathway through PSAT1, as well as into the Golgi body for glycosylation in tumor cells.
Following our analysis of PSAT1, we reasoned that a particularly interesting set of genes were those showing a higher degree of co-expression (as quantified by the magnitude of the Spearman correlation coefficient) in tumor samples relative to normal samples. For both BRCA and KIRC, we isolated pairs of genes exhibiting this property, and scored each metabolic gene based on how many such interactions it participated in. Interestingly, in both studies the highest-scoring gene was associated with the metabolism of lipids. In KIRC, the highest scoring gene was mevalonate kinase, MVK, a key gene in the cholesterol pathway described above for BRCA. In breast tissue, the highest scoring gene was LIPG, an endothelial lipase which catalyzes the hydrolysis of lipids. The products of this hydrolysis can then be used for the production of signaling lipids as well as cell membrane components.
We decided to investigate more comprehensively whether differential co-expression patterns were similar between BRCA and KIRC. To probe whether common, “global” patterns of differential co-expression existed between the two studies, we completed a principal components analysis (PCA, Fig 1A). We assembled a concatenated differential co-expression matrix:
D C = [ D BRCA D KIRC ] (7)
with dimension 2m × m, where m is the number of genes under study. For a given index i < m, row i corresponded to the differential co-expression pattern of that gene in BRCA, while row m+i corresponded to the differential co-expression pattern in KIRC. Thus, each column of DC corresponded to a metabolic gene, and stored the differential co-expression of that gene with all other metabolic genes in both breast and kidney studies.
Our expectation was that PCA would identify patterns of differential co-expression which breast and kidney cancers might share in common. Instead, we found that genes in the two studies displayed completely distinct patterns of differential co-expression (Fig 2A). While a large portion of the variance in the data was captured by the first two principal components (33 and 19 percent of the total variance in the data, respectively), most genes from breast cancer had nearly no loading on component 2, while most genes from kidney cancer had nearly no loading on component 1. The result was the cross pattern evident in Fig 2A.
Despite the results above, we still found a small positive correlation (Spearman ρ 0.28, p-value < 1e−30, Fig 2C) between differential co-expression in the two cancer types, suggesting that many genes showed high (or low) levels of differential co-expression in both studies. In general, most genes participated in relatively few differential co-expression interactions, while a small subset of “hub” genes participated in hundreds (Fig 2C, histograms). A particularly interesting example was ASS1, an enzyme involved in arginine synthesis and the synthesis of nitric oxide and polyamines. It is known that several tumor types exhibit an arginine auxotrophy phenotype, and are unable to proliferate in the absence of arginine [29]. Intriguingly, Qiu and colleagues recently reported the killing of triple-negative breast cancer cell lines under arginine deprivation, identifying it as a lucrative therapeutic target [30]. It is not clear from our analysis whether differential co-expression of ASS1 is associated with such a vulnerability, but its recurrent differential co-expression in both studies suggests that its activity may play an important role in malignancy.
The results of the PCA analysis above reflected the large number of cases of high differential co-expression in one tumor type, but none in the other. We explicitly identified such cases by calculating the mean and standard deviation of S0 (the number of differentially co-expressed gene pairs a gene participates in) for each study. We then searched for genes with S0 greater than two standard deviations above the mean S0 in one study, but with S0 = 0 in the other study (Fig 2C, blue and green points). KIRC-specific differentially co-expressed genes were highly enriched for SLC and ABC transporters. A particularly interesting kidney-specific gene was DPEP1 (a dipeptidase) in light of the recently observation of elevated dipeptide levels in a subset of clear cell renal carcinoma tumors (manuscript in preparation). In contrast, BRCA-specific genes included CDO1 (cysteine dioxygenase Type 1, whose inactivation was recently reported to contribute to survival and drug resistance in breast cancer [31]) and a number of genes involved in glycerolipid/lipid biosynthesis and associated with malignancy in breast cancer (GPAM [32] and MGLL [33]).
We also made special note of those pairs of differentially co-expressed genes which took part in a known, previously reported biological interaction. To do so, we extracted from the Pathway Commons database [34] a list of pairs of genes known to interact in either of two ways: 1) through the formation of a complex with each other (“In-Complex-With” interactions), and 2) through the production of a metabolite by the enzyme encoded by one gene in the pair, and subsequent use of that metabolite as a substrate for the enzyme encoded by the other gene in the pair (“Catalysis-Precedes” interactions) [35]. We then identified which pairs of differentially co-expressed genes participated in either of these kinds of interactions. These results were summarized in two additional gene-level statistics, S i C o m p and S i C a t, indicating the number of differentially co-expressed catalysis-precedes and in-complex-with interactions, respectively, a gene i participates in.
We compared the incidence of differential co-expression among pairs of genes participating in the binary interactions described above, to genes not participating in such interactions. To do so, we compared the distribution of transformed correlation coefficients (defined in 3) for the two groups of genes. We found a striking drop in co-expression among metabolic genes participating in a common molecular complex, an effect that was evident in both BRCA and KIRC (t-test, p-value < 1 × 10−200 BRCA, < 1 × 10−75 KIRC, S2 Fig, S3 Fig). A much weaker, but statistically significant, effect was also observed for catalytically adjacent genes (t-test, p-value < 1 × 10−18 BRCA, .0002 KIRC). Together, the results suggest a disruption of metabolic complexes in these two cancers. A more detailed future investigation is required to determine if this phenomenon is limited to metabolism, or is evident across all molecular complexes.
Finally, we analyzed the pattern of differential co-expression across metabolic pathways, as annotated in the Recon2 metabolic network [19] (see Methods). The results of our analysis are highlighted in Fig 2B, where we compared the score of each pathway in BRCA and KIRC, respectively. In breast cancer, among the most enriched pathways is peroxisomal transport genes, including the peroxisomal transporters ABCD1, ABCD2, and ABCD3, which transport fatty acids and acyl-CoAs and have been shown to be markers of tumor progression and response to therapy [36]. Notably, genes in the vitamin C pathway were enriched for differential co-expression in both cancers, possibly as an indirect consequence of high oxidative stress within the tumors.
A common first step in the analysis of gene expression data across samples is the identification of differentially expressed transcripts. The underlying rationale behind differential expression analysis of metabolic genes is intuitive: higher expression of genes in one condition over another suggests a difference in metabolic flux through those sets of genes. In this study, we are more concerned with the coupling of genes together: since metabolic genes are components of a network, different co-expression patterns may lead to differences in metabolic flux. Naturally, one may ask whether the two measures are in agreement; in other words, do genes which are up- or down-regulated in tumor (compared to normal tissue) also exhibit large differences in co-expression patterns in tumor (compared to normal) samples?
To explicitly test the connection between differential co-expression and differential expression, we compared the two measures for metabolic genes in BRCA and KIRC (Fig 3). We assessed differential expression using the limma voom package [37]. We found that the magnitude of differential expression (as quantified by the log2 ratio of tumor to normal expression) was weakly associated with the frequency of differential co-expression of a gene (BRCA, Spearman ρ 0.21, p-value 3 × 10−17; KIRC, Spearman ρ 0.11, p-value 4 × 10−6). In spite of this weak association, many of the most differentially expressed genes were members of very few dysregulated gene pairs, and conversely many genes which exhibited no substantial change in expression levels nevertheless were found to be frequent members of dysregulated gene pairs (S7 Fig).
The most intriguing observation we made was that a number of genes showed no measurable change in absolute expression levels, but nevertheless were among the most differentially co-expressed genes in the entire dataset (green dots, Fig 3). To find exceptional cases like these, we identified genes with S0greater than 2 standard deviations above the mean S0 for the study, but with an absolute log2 ratio of less than 0.2. For breast cancer, these genes included PLOD2 (procollagen lysyl hydroxylase 2 [38], recently reported to be essential for hypoxia-induced breast cancer metastasis), and LDHA, a key enzyme in the terminal end of glycolysis. In KIRC, several of the genes we identified (RENBP, GNE, and CTSA) were members of the glycoprotein sialyation pathway, which has also been associated with metastasis [39].
The presence of genes with exceptionally high differential co-expression and eseentially no differential expression (and the converse) deserves further discussion. It is possible that, depending on how the activity of a metabolic pathway is modulated, either differential expression or differential co-expression may be a more suitable technique for identifying such modulation. In one case, a gene may change in synchrony with its regulatory partners; that is, regardless of whether the gene is over- or under-expressed relative to normal tissue, it exhibits precisely the same co-expression patterns. Such an effect may be observed, for example, following the over-expression of a transcription factor common to all the genes in a co-expressed cluster. As we suggested earlier, synchronous regulation of a metabolic pathway may serve as a mechanism for increasing flux through the pathway, and would be detected through standard differential expression analysis. In contrast, a gene’s expression may correlate with different sets of genes in different conditions. In our case, the control over expression wielded by one transcription factor in normal tissue TFN would be ceded to a different transcription factor in tumor tissue TFT. The consequence is that the gene of interest is co-expressed with a completely distinct set of genes under the control of TFT. The differential co-expression of such a gene provides indirect evidence that the source or destination of metabolic flux through the enzyme encoded by this gene may be changing from normal to tumor tisues.
As alluded to above, the expression of genes is fundamentally orchestrated by regulatory factors such as transcription factors and microRNAs. Thus, the differential co-expression patterns we observe are likely due, at least in part, to differential regulatory activity by these molecules. Inspired by prior work linking transcription factors with observations of differential co-expression [18, 40], we examined our differential co-expression networks for an enrichment of targets associated with particular transcription factor motifs annotated in MSigDB [23]. To detect such enrichment, we isolated metabolic genes which were reported targets of a particular transcription factor. Then, we applied a binomial test (see Methods) to quantitatively assess whether the number of differential co-expression edges existing between only these target genes was higher than would be expected by chance. We used only highly significant differentially co-expressed edges, with a p-value threshold of 1 × 10−10.
Among the 556 transcription factor motifs we examined, only a handful were enriched in either kidney or breast cancer. In breast cancer, 21 transcription factors were identified as enriched in differentially co-expressed gene targets. The most enriched transcription factors (reported in Table 4) included SP1, NFAT, and ERR1. Several of these transcription factors have already been reported to play important roles in breast cancer throughout the literature. SP1 is known to be involved in cell proliferation, apoptosis, and cell differentiation and transformation, and has been reported as a prognostic marker for breast cancer [41, 42]. Both NFAT and SP1 have been shown to induce invasion of breast tissue via the transcriptional modulation of downstream genes [42, 43]. Perhaps most interesting is the identification of ERR1 (estrogen-related-receptor 1, also known as as ERR-α), an orphan receptor known to interact with PGC1-α to regulate a number of metabolism-related genes. ERR-α is regulated by ErbB2/Her2 signaling [44], and is associated with poor outcomes in breast cancer patients [45].
For kidney cancer, the pattern was far more unanimous: several of the most enriched transcription factor target sites were targets of HNF4A (Table 5). Out of the 15 transcription factors identified as enriched for differentially co-expressed gene targets, 5 were associated with HNF4. HNF4A is known to control cell proliferation in kidney cancer cell lines, and regulates a number of well-known cancer-associated genes to do so (e.g. CDKN1A and TGFA) [46–48]. Interestingly, HNF4A (one of the two isomers of HNF4, which was most enriched for differential co-expression targets) shows no clear differential expression pattern between KIRC tumor samples and adjacent normal tissue samples (Fig 4A), but does seem to exhibit more variation in tumor samples than in normal samples.
On the other hand, the co-expression of HNF4A and its metabolic gene targets is markedly different in normal and tumor samples (Fig 4B). A number of these genes (including PIK3R3, a member of the PI3K pathway, and PKLR, an isoform of pyruvate kinase) showed exceptionally strong co-expression with HNF4A in normal samples, only to have this co-expression abrogated in tumor samples (S8 Fig). Similarly, many of the strong co-expression patterns existent between the targets of HNF4A and HNF4A itself in normal samples wre also abrogated in tumor samples (Fig 4B). Together, these findings suggest that the regulatory program associated with HNF4A in normal tissue is disrupted in tumor tissue, a hypothesis in line with previous findings implicating its dysregulation with increased cell proliferation [46]. Given its high score in our enrichment analysis, we tested whether the expression of HNF4A was associated with patient survival in the TCGA data. After stratifying patients into groups with high and low expression (relative to the mean expression of HNF4A in the tumor samples), we found that low HNF4A expression is associated with shorter survival in KIRC patients (Fig 4D, log-rank p-value 0.007).
Taken together, our observations above suggest that HNF4A’s control over the expression of its targets changes in at least a subset of clear cell kidney tumors when compared to normal kidney tissue. It is possible that this loss of control occurs via under-expression of HNF4A itself. It is also possible that (as we proposed in the prior section) other transcription factors exert a more dominant control over HNF4A’s targets. In either case, this leads to the loss of co-expression among HNF4A’s targets, and between HNF4A itself and its targets.
This final section of our work strikes out into more difficult territory: we ask whether some patterns of differential co-expression may exist throughout different cancer types, regardless of their tissue of origin. While we have found a number of apparently dysregulated metabolic genes specific (and in some cases, common) to breast and clear cell renal cell carcinoma tumors, we have made little effort to search for common patterns across many different types of tumors. Such a search is necessarily complicated by the fact that our analytical method requires large numbers of normal and tumor samples for sufficient statistical power. The TCGA features few studies with large numbers of normal RNA-Seq samples. In order to balance the need for statistical power with our desire to detect so-called “PanCancer” patterns of differential co-expression, we included five more studies (lung adenocarcinoma, LUAD; hepatocellular carcinoma, LIHC; prostate adenocarcinoma, PRAD; head and neck squamous cell cancer, HNSC; and thyroid cancer, THCA) with at least 30 normal RNA-Seq samples, in our analysis. To increase the confidence of our predictions, we used a stricter p-value threshold of τ = 1 × 10−4 to call statistically significant differential co-expression.
The results of the PanCan analysis are shown in Fig 5. We retained only those genes which were members of a gene pair which was differentially co-expressed in at least three of the seven studies. Out of the 1789 metabolic genes under study, only 50 genes satisfied this criteria. Interestingly, many of these genes encode key enzymes in central metabolism (for example, PC, pyruvate carboxylase; LDHD, D-lactate dehydrogenase; IDH1, isocitrate dehydrogenase 1; ALDOA, aldolase A), pointing to apparently recurrent dysregulations of core pathways.
Among the many individual results of our PanCan analysis, perhaps the most interesting was the recurrent dysregulation of four genes in the mitochondrial electron transport chain (ETC): two genes associated with mitochondrial ATP synthase complex V (ATP5F1 and ATP5L), COX7B (part of the complex IV cytochrome c oxidase), and NDUFV2 (complex I). A number of other mitochondrial ETC genes are also differentially co-expressed (but to a lesser extent), including UQCR10, UQCRC2, UQCRC1, ATP5A1, and NDUFS3. Given how critical these protein complexes are to energy production and proliferation, we examined in detail the co-expression patterns of ATP5F1 and ATP5L. We found an exceptionally strong correlation in the expression of both genes in normal tissue. Across all seven studies, the expression of both genes was almost precisely equal (Fig 5B, blue dots). However, in tumor samples, the strength of the co-expression (as measured by the correlation coefficient) was substantially weaker. Notably, ATP5F1 and ATP5L were not differentially expressed; instead, their co-expression simply appeared “noisier” in tumor samples. To quantify whether this “noisier” co-expression may be occuring by chance, we fit each co-expression pattern in Fig 5B to a line, and then calculated the variance of the residuals of the fit. We used Levene’s test to test whether the variance of the residuals associated with tumor samples was larger than the variance of the residuals associated with normal samples. In all four tumor types, we confirmed that the tumor samples showed higher variance (p-value 7 × 10−17, 3 × 10−3, 6 × 10−9, 6 × 10−6, 3 × 10−11, 2 × 10−3, 5 × 10−3 for BRCA, HNSC, KIRC, LIHC, LUAD, PRAD, and THAC samples, respectively).
The functional consequences of these increasingly “noisy” co-expression patterns in ATP5F1 and ATP5L are unclear. It is known that stoichiometric imbalances of proteins (for example as a result of changes in gene dosage) in complex with each other can manifest phenotypically [49]. Given the recurrence of differential co-expression of three different gene pairs containing ATP5F1 and a second mitochondrial matrix member (ATP5L, COX7B, and NDUFV2), it is tempting to speculate that differential co-expression of ATP5F1 may lead to an altered mitochondrial phenotype. In particular, an imbalance in the levels of ATP5F1 and ATP5L may cause defects in the ability of mitochondria to efficiently conduct oxidative phosphorylation via the electron transport chain. Further experiments are required to evaluate this hypothesis.
In this work, we have searched for signals of differential co-expression in tumors. Among our findings, the most relevant is simply the prevalence of differential co-expression throughout metabolism. Gene expression studies are frequently the “first-step” analytical method of choice for understanding the consequences of a perturbation on an organism, or for the comparison of two distinct subsets of samples. While standard methods for differential expression analysis offer useful insights into the differential regulation of genes, our findings here (and the prior findings of others studying differential co-expression) suggest that a great deal of information remains to be culled from the study of “second-order” co-expression patterns between pairs of genes. We have shown that these two measures (differential expression and differential co-expression) are not interchangeable, and in many cases point to distinct regions of the metabolic network that may be dysregulated. Of course, it is important to remember that while the statistical power of both approaches relies on large sample sizes, differential co-expression is significantly more sensitive to sample size upon multiple hypothesis correction because of the large number of independent statistical tests (equal to the square of the number of genes) under evaluation. It will be interesting in the future to compare the results of our work to other methods for calculating differential interactions (e.g. partial correlations).
The orthogonality of differential expression and differential co-expression described above suggests that, to detect changes in the activity of a pathway, one must separately investigate the unilateral increase/decrease of enzyme levels, as well changes in their coordinated co-expression. In the first case, the expression of a large set of genes (for example, those in a long, linear metabolic pathway) may be synchronously upregulated. This coordinated up-regulation of transcription may, for example, enable the pathway to carry substantially more metabolic flux. In the second, perhaps more subtle case, the characteristic pattern of flux through a pathway may be re-wired (as illustrated in Fig 6). In Fig 6, the mechanism for this re-wiring is transcriptional, but in principle this type of coupling may arise through a variety of distinct mechanisms (such, as, for example, post-translational modification). In both cases, changes in intra- or extra-cellular conditions across a set of samples induces variation in the expression of genes. However, the manifestation of these changes may be hidden from either differential expression or differential co-expression analysis. Thus, we argue that both differential expression and differential co-expression analysis should play central, complementary roles in the analysis of gene expression data [11].
Our findings here are a small, first step in applying such a second-order analysis to cancer data, and in particular to the study of cancer metabolism. We have made a number of assumptions in order to make progress in the analysis, and these assumptions should be re-visited in future work. In particular, we have repeatedly assumed that the expression of a gene roughly correlates with the abundance of its translated protein product, and that this abundance correlates with enzyme activity. An entire field of theoretical study (metabolic control analysis, [50]) and a number of experimental studies (e.g. [51]) have shown that metabolite abundances are equally, if not more, important for the control of fluxes. We note, given an adequately large number of samples, an analogous “differential correlation analysis” is possible for metabolomics data. It would be especially interesting to compare the results from such an analysis with the analogous results using expression data.
One major concern with our results are the confounding effects of (1) contamination by stromal and immune cells, and (2) existence of heterogeneous tumor subtypes in the data. Tumor samples are often contaminated with mixtures of normal adjacent tissue and immune cells. Deconvolving the contribution of non-cancerous cells from the total signal obtained from a tumor sample remains a major computational challenge, and it is unclear how the contribution of this non-cancerous signal affects our differential co-expression results. A separate but related concern is the existence of distinct molecular subtypes in a set of samples (e.g. ER+, ER− breast cancer samples). We have not made any efforts to tease apart the confounding effects of these distinct subtypes in our work. Interestingly, it possible that a significant portion of the differential co-expression signal we identify derives directly form these subtypes; in other words, the primary differences between subtypes may lie among the differentially co-expressed genes. Evaluating such a hypothesis will require substantially larger sample sizes. Nevertheless, we feel that a more careful analysis of such patterns after subtype separation and stromal deconvolution is a lucrative route for future studies.
Finally, we would like to comment on the complementarity of differential expression and co-expression which we have proposed. In the course of responding to environmental stresses and stresses, it is inevitable that some genes will be both differentially expressed as well as differentially co-expressed. We are not arguing that one measure is superior to the other; rather, each offers a different glimpse onto the response of a highly-connected network to a perturbation. Neither the over-expression of a single gene, nor an increase in the co-expression of a pair of genes, signals a change in a pathway’s activity. However, by monitoring both measures, one univariate and the other multivariate, one may obtain a more complete picture of the complex system under examination.
|
10.1371/journal.pbio.1002333 | Reproducible Research Practices and Transparency across the Biomedical Literature | There is a growing movement to encourage reproducibility and transparency practices in the scientific community, including public access to raw data and protocols, the conduct of replication studies, systematic integration of evidence in systematic reviews, and the documentation of funding and potential conflicts of interest. In this survey, we assessed the current status of reproducibility and transparency addressing these indicators in a random sample of 441 biomedical journal articles published in 2000–2014. Only one study provided a full protocol and none made all raw data directly available. Replication studies were rare (n = 4), and only 16 studies had their data included in a subsequent systematic review or meta-analysis. The majority of studies did not mention anything about funding or conflicts of interest. The percentage of articles with no statement of conflict decreased substantially between 2000 and 2014 (94.4% in 2000 to 34.6% in 2014); the percentage of articles reporting statements of conflicts (0% in 2000, 15.4% in 2014) or no conflicts (5.6% in 2000, 50.0% in 2014) increased. Articles published in journals in the clinical medicine category versus other fields were almost twice as likely to not include any information on funding and to have private funding. This study provides baseline data to compare future progress in improving these indicators in the scientific literature.
| There is increasing interest in the scientific community about whether published research is transparent and reproducible. Lack of replication and non-transparency decreases the value of research. Several biomedical journals have started to encourage or require authors to submit detailed protocols, full datasets, and disclose information on funding and potential conflicts of interest. In this study, we investigate the reproducibility and transparency practices across the full spectrum of published biomedical literature from 2000–2014. We identify an ongoing lack of access to full datasets and detailed protocols for both clinical and non-clinical biomedical investigation. We also map the availability of information on funding and conflicts of interest in this literature. The results from this study provide baseline data to compare future progress in improving these indicators in the scientific literature. We believe that this information may be essential to sensitize stakeholders in science about the need for improving reproducibility and transparency practices.
| The inability to replicate published research has been an ongoing concern in the scientific community [1]. There is clear evidence from basic molecular and animal modeling research that a large portion of published articles lack reproducibility [2], which could potentially be related to the increase in lack of efficacy in clinical trials [3,4]. It has been suggested that the inability to replicate findings is due to a lack of research transparency [5]. Recently, there has been a growing movement to encourage making protocols, analytical codes, and data openly available [6–8]. In this study, we aimed to assess the current status of reproducibility and transparency in a random sample of published biomedical journal articles and to derive empirical data on indicators that have been proposed as being important to monitor in this regard [9], i.e., the proportion of studies sharing protocols and raw data, undergoing rigorous independent replication and reproducibility checks, and reporting conflicts of interest and sources of public and/or private funding.
A total of 441 (88.2%), from the original randomly selected 500 articles were publications in eligible research fields directly related to biomedicine. Of these, two-thirds had some form of empirical data (n = 304 (68.9%)—n = 268 excluding case studies and case series, in which protocols and raw data sharing may not be pertinent, and n = 259 excluding also systematic reviews, meta-analyses and cost-effectiveness analyses where replication in studies with different data would not be pertinent). Among the 441 eligible studies, four (0.9%) were cost effectiveness or decision analyses, 36 (8.2%) were case studies or case series, 15 (3.4%) were randomized clinical trials, five (1.1%) were systematic reviews or meta-analyses, and 244 (55.3%) were other articles with empirical data (including cross-sectional, case-control, cohort, and various other uncontrolled human or animal studies). Just over 30% of the articles were classified as research without empirical data or models/modeling studies. Less than one in five articles (19.2%) had open full-text access from PubMed Central and about half (47.8%) of the papers belonged to the journal category of clinical medicine (Table 1).
Excluding case studies or case series (in which a protocol would not be relevant) 267 (99.6%) of the 268 papers with empirical data did not include a link to a full study protocol. Only one article had a protocol; in fact, the article was itself the protocol of a trial, and it was published in the open-access journal Trials (A221). Another five studies either referenced their clinical trials identifier and included a link to ClinicalTrials.gov (A281, A434), provided only Clinical Trials identifiers (A407, A477), or stated that a Clinical Trials repository link was available on the journal website (A261), but none of these articles or their links contained information about a full protocol.
There were seven other articles that had additional methods sections, figures, brief analytical plans and/or considerations, or supplementary materials either as a detailed appendix at the end of the paper (A434) or online (A25, A35, A174, A376, A290, A361 [contained an error message for page not found]). However, none of these supplementary materials fit our pre-specified definition of publicly available full or partial protocols.
Of the 268 articles with empirical data (excluding case studies and case series) none provided access to all the raw data involved in the study. One article contained information on how to request a complete dataset (A287), two papers listed a non-functioning online link for supplementary data, data elements, or findings (A330, A361), and another four articles had supplementary files or links to some absorption spectra figures and/or data (A35, A117, A130, A305), but not to the entire raw data used in the paper.
About half (51.7%) of the 441 biomedical articles did not include any information on funding and about a third (34.7%) were publically funded either alone or in combination with other funding sources. Of the 153 publically funded articles, 62 had National Institutes of Health (NIH) funding and four received National Science Foundation (NSF) support, alone or in combination with other sources of funding (Fig 1). There was no major change in the pattern of sources of funding over the 15-y period (S5 Fig).
Of the 259 biomedical articles with empirical data, excluding case studies and case series, systematic reviews/meta-analyses, and cost effectiveness/decision analysis studies, only four (1.5%) clearly claimed or were inferred to be replication efforts trying to validate previous knowledge. Over half (51.7%) of the studies claimed to present some novel findings and four (1.5%) had clear statements of both study novelty and some form of replication. There were 117 (45.2%) articles that either had no statement or an unclear statement in the abstract and introduction about whether there were any novel findings or replication efforts.
For the 259 biomedical articles with empirical data, excluding case studies and case series, systematic reviews/meta-analyses, and cost effectiveness/decision analysis studies, we also assessed whether any subsequently published papers had cited the article, mentioning that the authors were attempting to replicate part or all of their findings. Eight articles (3.1%) in the final dataset had at least some portion of their findings replicated (A11, A59, A129, A222, A278, A285, A407, A441), while the remaining 251 articles had no citing article that claimed to be a replication. Five of these eight articles were from the clinical medicine journal category. Of the replicating articles, one was unable to reproduce the results from the original article (A59), but mentioned that different definitions were used [10]. Three articles had their results replicated through different methodology [11] (A222), [12] (A278), [13] (A407). One article (A11) had several subsequent studies that either confirmed portions of the original study [14] or failed to validate certain previous findings [15]. Two studies developed new methodology that the replicating studies confirmed (A285), either by comparing to available methods [16] or a newly developed method [17] (A441). One article (A129) was cited by a subsequent study by the same first author [18] that stated that one of their aims was to test a hypothesis from earlier observations with longer observation and modeling techniques.
In order to measure whether empirical studies are eventually integrated in systematic reviews, the 259 articles with empirical data (again excluding case studies and case series, systematic reviews/meta-analyses, and cost effectiveness/decision analysis studies) were assessed on whether they had been cited at least once in subsequent systematic reviews and/or meta-analyses. Empirical data from 16 articles (6.2%) were utilized in a systematic review/meta-analysis (A89, A93, A105, A157, A190, A222, A261, A268, A270, A278, A338, A340, A374, A407, A421, A477). At least one systematic review/meta-analysis cited another three articles but provided reasons for not including any of their data in a quantitative synthesis for any outcome (A83, A129, A221). Yet another 19 articles were cited incidentally by systematic reviews/meta-analyses (e.g., in introduction or discussion, but without having data considered in quantitative syntheses for any outcome) (A5, A28, A31, A112, A203, A207, A224, A256, A274, A319, A322, A327, A377, A400, A413, A433, A435, A453, A463). Lastly, there were 221 articles (85.3%) that were not cited in any systematic reviews/meta-analyses.
The large majority of the 441 articles had no conflict of interest statement (305 [69.2%]). Of the remaining, 110 (24.9%) did not report any conflicts of interest and 26 (5.9%) reported conflicts of interest. For the 15 randomized controlled trials, eight articles (53.3%) reported no conflicts of interest, four (26.7%) articles had no statement of conflict, and three (20.0%) articles had a clear statement of conflict.
Between 2000 and 2014, the percentage of articles with no statement of conflict decreased substantially (94.4% in 2000 to 34.6% in 2014), whereas the number of articles reporting statements of conflicts (0% in 2000, 15.4% in 2014) or no conflicts (5.6% in 2000, 50.0% in 2014) increased (Fig 2).
A comparison of articles published in journals in the clinical medicine category versus other fields showed some distinctive patterns (Table 2). Articles in the clinical medicine journal category were almost twice as likely to not include any information on funding and to have private funding, while they were far less likely to have public funding or funding from different types of sources (public and/or private and/or other). Articles in the clinical medicine journal category were also more likely to contain no statement on novelty or replication and less likely to claim novel study findings than articles in the “other” journal category. Furthermore, articles in the clinical medicine journal category were less likely to have full open access compared to other fields of study. There were no significant differences between replication, article citation for systematic review and/or meta-analysis, and statements of conflict (Table 2).
When further limited to the 304 articles with empirical data, articles in the clinical medicine category journals were more likely to not mention funding (59.0% versus 26.3%) and less likely to have a PubMed Central reference number (PMCID) (16.0% versus 27.5%). There was no significant difference in the proportion of articles not including statements of conflicts of interest (65.3% [clinical medicine] versus 71.3% [other]).
Our empirical evaluation shows that the published biomedical literature lacks transparency in important dimensions. We found a full protocol only for one study in our sample. In basic science exploratory research, formal protocols may not be available ahead of time. However, a post-research detailed protocol should be provided. It is unclear how many biomedical papers have no protocols versus do have protocols but do not make them publicly available. During the earlier years of the sampling time frame, there may not have been many online protocol-sharing repositories, such as OpenWetWare [19], which was created in 2005. However, authors could have included a statement about the availability of their protocol either upon request or on a personal or laboratory website.
Previous evaluations have identified common inconsistencies between available protocols and final publications of randomized trials [20,21]. For other types of study designs, such comparisons are hampered by the rare availability of protocols. Public protocol sharing not only provides external researchers ways to find possible discrepancies between final publications and research plans [9], more importantly, it allows study designs and experiments to be reproduced by interested scientists.
Of the 268 biomedical articles with empirical data assessed (excluding case studies and case series), none had open access to all the raw data. A previous evaluation found 9% of articles published in the 50 journals with the highest impact factor in 2009 had deposited full primary data online [21]. Our results may differ due to the fact that we focused on the full spectrum of biomedical journals in PubMed (median impact factor 3.2). Data sharing requirements have changed over the last few years, especially in high-impact journals [21,22], but these represent only a small fraction of the journals studied here. Although one article in our study claimed that the complete dataset was available upon request (A287), a statement of willingness to share may not guarantee that the data will be available to independently requesting scientists [23]. Sponsor priorities, lack of resources, personal investigator opinions, and proprietary perceptions may influence data withholding [24,25]. Six other articles included supplementary files or links with some additional data, but for two of them, the links were nonfunctioning (A330, A361). Evidence exists that even the most prestigious journals have supplementary information that eventually became unavailable [26].
Although the NIH reaffirmed their support for the concept of data sharing in 2003 by stating that applications seeking $500,000 or more in direct costs for a single year are expected to include plans for sharing data or statements why data sharing would not be possible [27], there is no evidence of a major change in data sharing practices between 2003 and 2014 when the entire PubMed-indexed literature is considered. According to the NSF Data Sharing Policy in 2011, investigators have been expected to share the primary data and other supplementary materials created or gathered under NSF grants with other researchers within a reasonable time frame [28]. With only a few studies funded by NSF in this sample of biomedical articles, it is not possible to determine whether this policy has had any impact over time. The sharing of raw data and protocols will be facilitated by the emergence of more available options and repositories, but this plethora of choices may need to be streamlined at some point. Investigators may also continue to use their preferred method of sharing.
The majority of papers claimed to present some novel discoveries. However, we suspect that very few papers truly have totally, disruptively innovative findings. Instead they may be operating in knowledge space where other past studies may also have operated, but they still claim novelty. It is difficult to probe objectively how much innovation is needed to be able to claim novelty. Moreover, none of the subsequently published papers that cited the original articles and mentioned that the authors were attempting to replicate part or all of their findings were full study replication attempts. Replication has been accepted as a sine qua non in a few disciplines, such as human genome epidemiology, but those disciplines are the exception. When some effort at replication is done, investigators may still try to differentiate their replication study as being different from the original and, thus, also make a case for novelty. There are many different proposals on how reproducible research can be guaranteed. These include approaches at reproducible practices, i.e., making other investigators able to repeat the process and calculations [29]; re-analysis (as in the case of randomized trials [30]); and replication by independent investigators, as in genetics, psychology, and cancer biology [31–33]. We also demonstrated that very few primary data are currently included in systematic reviews and meta-analyses. Despite the advent of evidence-based medicine, these data syntheses still cover only small fractions of the available evidence.
Previous studies have found that between 29% and 69% of published clinical research articles had some type of financial conflict [34,35], and a survey of NIH-funded life science researchers found that 43% of 2,167 respondents reported receiving some research-related contribution, such as reagents, equipment, travel funds, etc. [36]. In our study, which covers a very wide range of research designs and types, only 5.9% of the articles had conflict of interest statements. This is likely an underestimate of the prevalence of conflicts in biomedical research and could be a result of the lack of conflict of interest disclosure policies among many journals [37]. However, we also found that the number of statements reporting no conflicts of interest increased and, conversely, the number of articles without any statements decreased over time, perhaps due to strengthening of certain journal disclosure policies [37]. The persisting high prevalence of no statement of conflict is, nevertheless, worrisome. Conflicted stakeholders can operate in a stealth mode and have a significant impact on the design, conduct, and analysis of biomedical studies [1,38,39].
Slightly over half of the analyzed papers reported no funding. It is possible that some of them simply did not mention existing sponsors. Still, a large share of the published literature occurs without any support, and this should cause some concern. Public funding is listed for about a third of the 441 biomedical papers and NIH accounts for a mere 14% of the total biomedical literature. The challenge is even greater for clinical research in particular: only 9.0% of published papers in journals in the clinical medicine category mentioned NIH funding, and more than 70% of papers in this category mentioned no funding or clearly state that they had no funding at all. Underfunding in combination with conflicted sponsored funding creates a difficult situation for clinical research.
Our evaluation is limited to published biomedical research information. In theory, sometimes one may be able to obtain additional raw data and protocols, and clarifications on conflicts or funding by communicating with the authors or sponsors. However, the yield would be uncertain, and personal communications should not replace the lack of transparency in the published scientific record. Furthermore, the fact that we only used the published records means that we could not correct any inaccuracies in the claims of the original authors. This may be particularly prominent in the case of claims for novelty, in which some authors may have tried to sell their paper as being more novel than it really is, so as to make it more attractive for publication. Although the two investigators (SAI and JDW) used their best judgment and discussed all eligible papers before agreeing upon a final classification, certain decisions may have been subjective. In particular, when determining study novelty and replication for articles from diverse biomedical fields, difficulty arose assessing whether study results were truly groundbreaking or being fully replicated. In order to account for these limitations, all ambiguous articles were discussed with a third reviewer (JPAI).
We hope that our survey will further sensitize scientists, funders, journals, and other stakeholders in science to the need to improve these indicators. There are several efforts to improve reproducibility [40–42]. By continuing to monitor these indicators in the future, it is possible to track any evidence of improvement in the design, conduct, analysis, funding, and independence of biomedical research over time.
A sample of 500 English-language journal articles published between 2000 and 2014 was chosen randomly based on PubMed identification (PMID) numbers. PMID numbers ranging from 10,000,000 to 25,000,000 were inputted into OpenEpi (version 3.02) random number generator to select a random sample of 750 PMID numbers. Beginning from the first number generated, each number was verified for eligibility in sequence until 500 eligible PMID numbers were chosen. Of the original 750 numbers, 742 were checked, with 242 being ineligible (54 did not have an article assigned, 100 were from before the year 2000, 35 were not in English, and 53 were not in English and before the year 2000). The selected article distribution of PMID numbers (by year) was compared to the overall distribution of PMID numbers by year for English articles. The sample was found to be representative of the overall distribution (χ2 (df = 14), p > 0.05). This sample size was chosen because given 500 articles and assuming that about half of them might have empirical data, if no article is found to fulfill the criterion for a transparency indicator, then the 95% confidence interval around that 0% estimate does not exceed 1%.
Two investigators independently characterized and then cross-compared all extractions in groups of 50 articles at a time. Any uncertainties were first discussed in detail, and a third reviewer (JPAI) reassessed articles with arbitration discrepancies.
The sample was characterized into seven study categories: (1) no research (items with no data such as editorials, commentaries, news, comments and non-systematic expert reviews), (2) models/modeling or software or script or methods without empirical data (other than simulations), (3) case report or series (humans only, with or without review of the literature), (4) randomized clinical trials (humans only), (5) systematic reviews and/or meta-analyses (humans only), (6) cost effectiveness or decision analysis (humans only), and (7) other (empirical data that includes uncontrolled study [human], controlled non-randomized study [human], or basic science studies).
InCites Essential Science Indicators (ESI) was used to determine the main scientific field of each article. The journal for each index paper was searched in ESI in order to find the scientific field to which its Highly Cited Papers are ascribed. If a journal had articles ascribed to more than one scientific field, we examined the first five cited journals referenced by the index article. The journal names for these articles were then searched in ESI. If the majority belonged to the same field, this field was used for the index paper. If there was no majority, a field was selected based on the best judgment of the reviewers (JPAI, SAI, and JDW). If a specific journal was not found on ESI, we searched Journal Citation Reports (JCR) and identified the scientific field to which the highest-cited journal in the same JCR category had been ascribed to in ESI.
Publications in scientific fields not directly related to biomedical research (chemistry, physics, computer science, economics and business, engineering, geosciences, material science, mathematics, physics, and space science) were further excluded from analysis. Even though these fields may sometimes have repercussions for biomedicine, their transparency practices may differ systematically, and their evaluation would require a separate, focused effort. Thus, 59/500 articles were excluded.
JCR was used to determine 2013 journal impact factor. No information was recorded for journals without an impact factor for 2013. Availability of free access in PubMed Central was based on assignment of a PCMID (yes/no).
|
10.1371/journal.pgen.1005149 | Spastin Binds to Lipid Droplets and Affects Lipid Metabolism | Mutations in SPAST, encoding spastin, are the most common cause of autosomal dominant hereditary spastic paraplegia (HSP). HSP is characterized by weakness and spasticity of the lower limbs, owing to progressive retrograde degeneration of the long corticospinal axons. Spastin is a conserved microtubule (MT)-severing protein, involved in processes requiring rearrangement of the cytoskeleton in concert to membrane remodeling, such as neurite branching, axonal growth, midbody abscission, and endosome tubulation. Two isoforms of spastin are synthesized from alternative initiation codons (M1 and M87). We now show that spastin-M1 can sort from the endoplasmic reticulum (ER) to pre- and mature lipid droplets (LDs). A hydrophobic motif comprised of amino acids 57 through 86 of spastin was sufficient to direct a reporter protein to LDs, while mutation of arginine 65 to glycine abolished LD targeting. Increased levels of spastin-M1 expression reduced the number but increased the size of LDs. Expression of a mutant unable to bind and sever MTs caused clustering of LDs. Consistent with these findings, ubiquitous overexpression of Dspastin in Drosophila led to bigger and less numerous LDs in the fat bodies and increased triacylglycerol levels. In contrast, Dspastin overexpression increased LD number when expressed specifically in skeletal muscles or nerves. Downregulation of Dspastin and expression of a dominant-negative variant decreased LD number in Drosophila nerves, skeletal muscle and fat bodies, and reduced triacylglycerol levels in the larvae. Moreover, we found reduced amount of fat stores in intestinal cells of worms in which the spas-1 homologue was either depleted by RNA interference or deleted. Taken together, our data uncovers an evolutionarily conserved role of spastin as a positive regulator of LD metabolism and open up the possibility that dysfunction of LDs in axons may contribute to the pathogenesis of HSP.
| Hereditary spastic paraplegia (HSP) is a genetically heterogeneous neurological disease characterized by weakness and spasticity of the lower limbs, caused by progressive retrograde degeneration of the corticospinal axons, the longest in the central nervous system. The most commonly mutated gene in autosomal dominant forms of HSP, SPAST, encodes for spastin, a microtubule-severing protein. Spastin has been implicated in several processes involving remodeling of membrane structures. We now show that the longest spastin form, spastin-M1, harbors a lipid droplet targeting sequence, which allows targeting of the protein to the surface of lipid droplets, the organelles where cells store neutral lipids. Furthermore, we demonstrate that depletion of the homologous spastin proteins in both flies and worms affects lipid droplet number and triacylglycerol content. Our study adds to recent discoveries that implicate other HSP proteins in lipid droplet and lipid metabolism, and strongly suggests that lipid droplet dysfunction in neurons should be investigated to understand pathogenesis of HSP.
| Lipid droplets (LDs) are complex and dynamic organelles whose function is to assemble, store, and supply neutral lipids, mainly sterol esters and triacylglycerols (TAGs) [1, 2]. Initially recognized in specialized cells, such as adipocytes, it is now clear that any cell has the ability to form LDs. Current models consider LDs as specialized compartments of the tubular endoplasmic reticulum (ER), from which they derive in a step-wise process. This involves the formation of a lipid globule that grows within the two leaflets of the ER membrane bilayer via sequential and controlled recruitment of enzymes, which catalyze the accumulation of lipids and stimulate the formation of the curvature of the outer leaflet of the ER membrane [3]. Once LDs are formed, they may remain attached to the ER membrane [4]. Dysfunctions of LDs have been implicated in several pathologic conditions, such as obesity, atherosclerosis, and lipodystrophies [1]. LDs are occasionally found in ultrastructural studies of neurons [5], however very little is known about their role in these cells. Notably, LDs appear to be increased in the brain of Alzheimer patients [6]. Moreover, α-synuclein, the major constituent of Lewy bodies in Parkinson’s disease, was shown to accumulate on the surface of LDs in cells loaded with lipids [7]. A link has recently emerged between LDs and axonopathies of the central nervous system, such as hereditary spastic paraplegia (HSP).
HSP is a genetically heterogeneous neurological disease, clinically defined by the association of weakness and spasticity of the lower limbs (pure HSP) [8, 9]. The disease is caused by progressive retrograde degeneration of the longest axons of the central nervous system, those composing the corticospinal tract [10]. Most cases of autosomal dominant HSP are caused by mutations in three genes, SPAST, ATL1, and REEP1 [11–13]. SPAST encodes spastin, a microtubule (MT)-severing protein belonging to the AAA (ATPases Associated with various cellular Activities) family [14–16]. Spastin is involved in several processes requiring a dynamic cytoskeletal network, such as midbody abscission, neurite branching formation, axonal stability, and endosomal trafficking and tubulation [17–21]. In several of these processes, spastin mediated MT-severing is coupled to specific membrane remodeling events.
We previously showed that mammalian cells produce two spastin isoforms, spastin-M1 and spastin-M87, depending on the usage of two alternative start codons and alternative promoters [22, 23]. These isoforms differ in their subcellular localization, MT-severing activity, and binding to known interactors [18, 21, 22, 24]. The long spastin isoform (spastin-M1) is predominantly expressed in neurons, and appears enriched in the early secretory pathway, while the shorter spastin-M87 isoform can be recruited to endosomes [18, 22]. Spastin-M1 is characterized by an N-terminal sequence extension containing a hydrophobic stretch required for association with the ER membrane and interaction with REEP1 and atlastin-1, the product of the ATL1 gene [24]. Atlastin-1 is a GTPase of the dynamin superfamily, which mediates fusion of the tubular ER, while REEP1 regulates the morphology of the ER, by affecting the curvature of the ER membranes and by mediating interaction with the MTs [25, 26]. These findings have fostered the hypothesis that abnormalities of tubular ER morphogenesis in long motor axons underlie the pathogenesis of HSP [27].
Remarkably, a recent study showed that atlastin GTPases have an evolutionary conserved role in regulating LD size in invertebrates [28]. Moreover, at least other three proteins encoded by HSP causative genes have been implicated in LD function. Spartin (SPG20) is mutated in Troyer syndrome, a complicated form of autosomal recessive HSP [29]. Knockdown or overexpression of spartin affects LD turnover in cells, and spartin knockout female mice show increased LDs in adipose tissue [30, 31]. Dominant mutations in the BSCL2 gene, encoding the ER-resident protein seipin, are found in families presenting with a broad range of neurological features, including HSP with amyotrophy (SPG17), and Charcot Marie Tooth disease [9]. The exact molecular function of seipin is unknown, however both in humans and in yeast loss of seipin impairs LD formation [32]. Recently, recessive forms of HSP genes have been linked to mutations in genes involved in fatty acid metabolism, such as the phospholipases DDHD1 and DDHD2 [33, 34]. DDHD2 was further found to encode for the principal TAG lipase in the brain, and Ddhd2 knockout mouse showed an increased number of LDs and TAG accumulation in the central nervous system [35].
Here, we show that human spastin-M1 localizes to the ER and mature LDs, where it is recruited via action of a hydrophobic domain interrupted by an arginine residue. We show evidence that endogenous spastin-M1 is detectable in the LD fraction. Moreover, modulation of spastin levels regulates LD number and TAG levels in Drosophila and C. elegans. Our data further support the notion that studying the role of LDs in neurons is of relevance to fully comprehend the pathogenesis of HSP.
Expression of spastin-M1 in a variety of cell lines labels vesicular compartments that were found to co-localize only partially with markers of the early secretory pathway or endosomes [14, 18, 36, 37]. However, the nature of the majority of these structures has remained elusive. When HeLa cells are permeabilized using saponin, many but not all spastin-M1 labeled vesicles appear as ring structures, suggesting that they may represent LDs. We therefore expressed spastin-M1 and stained LDs using a neutral lipid dye (BODIPY 493/503) or antibodies against the LD protein PLIN2, a member of the PAT (perilipin/PLIN1, ADRP/PLIN2, TIP47/PLIN3) family of proteins. The spastin signal decorated the LDs and co-localized with PLIN2 (Fig 1A). Besides forming ring structures, overexpressed spastin-M1 also labels calnexin-positive aggregates, consistent with accumulation in ER membranes (Fig 1A). We observed that overexpression of spastin affects the morphology of the ER (Fig 1A). Next, we induced LD formation by incubating HeLa cells with oleic acid (OA) for 16 h. In this condition, cells accumulated a larger number of bigger LDs. Remarkably, we found that spastin-M1 localized to LDs (Fig 1B and S1 Fig). PLIN3 is a cytosolic protein that is recruited to nascent LD membranes within minutes after OA administration [38]. Cells co-expressing GFP-PLIN3 and spastin-M1 showed a high-degree of co-localization of both proteins (Fig 1B and S1B Fig). The staining pattern was independent of tagging the protein at its C- or N-terminus and was observed also after transfection of an untagged construct (S1 Fig). Spastin-M1 targeting to LDs is not cell type specific, as it was observed in different cell lines (COS7, NSC34, SH-SY5Y) (S2 Fig). After OA loading, cells expressing spastin-M1 still show a morphologically altered ER (S3 Fig).
To gain further insights into the distribution of spastin to different cell compartments, we fractionated postnuclear supernatants of OA-treated HeLa cells transfected with spastin-M1 by sucrose gradient centrifugation to separate floating LDs from cytoplasm and other organelles. In agreement with immunofluorescence experiments, spastin-M1 was enriched in the LD fraction that contained PLIN3. Still, a significant amount of spastin-M1 was detectable in the bottom fractions and in the pellet, where microsomal membranes are found (Fig 1C). Finally, to follow the recruitment of spastin-M1 to LDs, we combined transfection into HeLa cells with time-lapse video microscopy over several hours. Within the first 60 min of detectable spastin-M1 expression (6 to 7 hours after starting imaging), the protein was restricted to BODIPY 493/503-positive LDs. In contrast, spastin-M1 aggregates, which were not stained for neutral lipids, emerged with increasing levels of expression (Fig 1D). These data suggest that spastin-M1 preferentially targets LDs when these are present in cells. To further explore the connection between spastin-M1 and LDs, we investigated the relationship with pre-existing LDs (pre-LDs). Pre-LDs have been recently defined in COS1 cells as restricted ER microdomains with a core of neutral lipids that are resistant to starvation [4]. Upon arrival, lipids are first deposited in pre-LDs that can be labeled using a model peptide (HPos) that shifts from the ER to LDs in response to fatty acids feeding [4]. In transfected cells, most spastin-M1 puncta were also HPos-positive (Fig 2B). After further incubation in presence of OA for 24 h, spastin-M1 and HPos co-localized on the surface of LDs, indicating that they follow an OA-promoted transport pathway from the ER to LDs (Fig 2B).
All together, these data suggest that spastin-M1 localizes to LDs. After starvation, spastin-M1 localization corresponds to pre-LDs defined by HPos labeling.
To establish whether targeting of spastin to LDs is specific to the M1 isoform, we transfected spastin lacking the N-terminal extension, and analyzed LD association after OA administration. We found that, in contrast to spastin-M1, spastin-M87 does not decorate pre-LDs or LDs (Figs 2C and 3A). Deletion of the first 50 amino acids of spastin instead does not impair LD targeting (Fig 3A). Analysis with TMMHM server of the region of spastin encompassing amino acids 50 through 86 predicts a transmembrane domain between amino acids 57 and 79. Since both N- and C-termini of spastin were mapped to the cytoplasm, this hydrophobic region has been previously proposed to adopt a hairpin configuration [24]. Several proteins can sort to the surface of LDs via hydrophobic stretches interrupted by basic residues resulting in a hairpin configuration [1, 2]. In caveolin, positively charged stretches were identified to act cooperatively with a hydrophobic domain to mediate LD sorting [39]. A positive charged sequence in combination with a hydrophobic domain also characterize the HPos peptide [4]. Similarly, the hydrophobic motif of spastin is interrupted by an arginine (R65) and is followed by two basic residues (R81 and R84) (Fig 3B). We fused the region from amino acids 57 to 86 to mCherry (TM-mCherry), and found that this reporter construct is successfully recruited to LDs when expressed in HeLa cells (Fig 3C). Remarkably, when we mutated arginine 65 to a glycine in this construct (TM-R65G-mCherry), LD targeting was abolished (Fig 3C), and the mCherry protein showed a reticular and punctate staining only partially positive for an ER marker (S4A Fig). In contrast, a mutant construct TM-R81/84G-mCherry still localized to ring-structures filled with neutral lipids (Fig 3C). We then checked whether mutation of arginine 65 to glycine abolished LD localization of full-length spastin. Targeting of mCherry-spastin-M1-R65G to LDs was largely impaired in transfected cells, and spastin aggregated in puncta that were co-labeled by antibodies against the ER marker REEP5 (S4B Fig). All together, these data point to an essential role of the hydrophobic domain and of arginine 65 for LD targeting of spastin.
We observed that HeLa cells overexpressing spastin-M1 showed bigger LDs compared to neighboring non-transfected cells (Fig 4A). We therefore quantified the total number of LDs, the total cell volume occupied by LDs, and the average LD volume in spastin-M1 transfected cells compared with cells transfected with the empty vector. Spastin-M1 expression significantly reduced the number of LDs per cells, while increasing their size (Fig 4). We then transfected spastin-M87, which severs MTs more efficiently than spastin-M1 [40] and cause similar ER alterations as spastin-M1 (S3 Fig). Under this condition, LD number was slightly reduced, but LD size was not affected, suggesting that targeting of spastin to LDs is necessary to regulate their size (Fig 4A–4D). Targeting of spastin-M1 to LDs is maintained when expressing a mutant deleted of the MT-binding domain (spastin-ΔMBD), which leaves intact both the ER and the MT-network (Fig 4A and S3 Fig). However, expression of spastin-ΔMBD resulted in a perturbation of the distribution of LDs, by enhancing their clustering in proximity of the cell nucleus (Fig 4A–4E). The tight clustering of LDs precluded measurement of their individual size, since it was difficult to distinguish them as individual objects in image quantifications. These data suggest that spastin-M1 may regulate the size of LDs, while spastin interaction with MTs is important for LD dispersion into the cytoplasm.
An important question is whether endogenous spastin is recruited to LDs. Spastin-M1 is almost undetectable in HeLa and other cell lines containing a significant amount of LDs. We decided to investigate the subcellular distribution of spastin in the motor neuron-like NSC34 cells, in which the anti-spastin antibody detects four bands by immunoblotting (Fig 5A). These correspond to spastin-M1 and spastin-M87, with and without the alternatively spliced exon 4. Consistently, all four bands are depleted by a specific siRNA designed to recognize all spastin isoforms, while a siRNA specific for a sequence in SPAST exon 4 depletes only the two forms containing this exon (Fig 5A). As already described in many tissues and cell lines [22, 40], spastin-M87 with and without exon 4 are the predominant isoforms, and staining with a spastin-specific antibody mainly detects spastin-M87 by immunofluorescence. We therefore separated LDs on a sucrose gradient, after treating NSC34 cells with OA overnight. All spastin forms were present in the pellet fraction (containing ER membranes) in the expected relative ratio (Fig 5B). However, spastin-M1 and spastin-M87 partitioned differently in the other fractions of the gradient. Spastin-M87 was mainly detected in bottom fractions (Fig 5B), and only traces were found in the LD fraction. In contrast, spastin-M1 was found in the LD fraction but not in the bottom fractions. While in the starting lysate, spastin-M1 is present in low amounts in respect to spastin-M87, in the LD fraction spastin-M1 is the most abundant isoform (Fig 5C). This result argues against contamination of the LD fraction by ER membranes, and strongly supports specific targeting of endogenous spastin-M1 to LDs. A likely explanation for the detection of spastin-M87 in the LD fraction is based on the ability of spastin to form mixed hexamers of spastin-M1 and spastin-M87. Indeed, Flag-spastin-M1 expressed in OA-treated HeLa cells can recruit mCherry-spastin-M87 to ring-like structures filled with neutral lipids (Fig 5D). In conclusion, these data demonstrate that endogenous spastin-M1 is recruited to LDs.
We then asked whether spastin downregulation in NSC34 cells has an impact on LDs size or number, and on TAG levels. However, no major differences were detected in cells treated with a spastin-specific siRNA compared to cells treated with a control siRNA or to mock-transfected cells (S5 Fig).
To investigate whether spastin exerts any physiological role in LD biogenesis or metabolism, we turned to Drosophila, which has been previously established as a model organism to investigate the pathogenic mechanism of neurodegeneration caused by spastin mutations [41]. Neural-specific knockdown of Dspastin or overexpression of a dominant-negative variant (DspastinK467R) caused adult-onset impairment of locomotor and neurodegeneration, and associated with excessive stabilization of the MTs in the neuromuscular junction [41].
Drosophila spastin (Dspastin) is 758 amino acid-long and diverges from mammalian spastin mainly in the N-terminal region. However, in this region a transmembrane helix containing an arginine and followed by a positively charged motif is predicted, reminiscent of that found in human spastin (S6A Fig). When expressed in COS7 cells, Dspastin severs MTs (S6B Fig), and shows a reticular pattern of expression, suggestive of ER localization. Consistently, we found co-localization between the Dspastin signal and the ER marker calnexin (S6C Fig). Recruitment of Dspastin to LDs was observed, although it was never as prominent as that observed for human spastin (S6D Fig).
We decided to manipulate spastin levels in the fly and analyze potential effects on LD number and size. Dspastin knockdown and overexpression were achieved in vivo using the Gal4/UAS binary expression system [42]. Firstly, we altered Dspastin levels with the ubiquitous promoter actin-Gal4, which target all tissues, including fat bodies, the main fat store fat in the fly, where spastin is normally expressed at low levels (FlyAtlas microarray data and modENCODE tissue expression data). In agreement with results in HeLa cells, we observed that overexpression of spastin drastically increased the size of LDs in the fat bodies while reducing their number (Fig 6A–6D). In contrast, when spastin was downregulated by RNAi (actin-Gal4/UAS-DspastinRNAi), LDs were less numerous and the total area stained by BODIPY 493/503 was reduced (Fig 6A–6D). To confirm this result, we expressed the pathogenic mutant DspastinK467R, which contains an amino acid change known to abrogate ATP binding and elicit dominant-negative effects in vivo thereby mimicking loss of function phenotypes [41]. Similar to expression of DspastinRNAi, ubiquitous expression of DspastinK467R (actin-Gal4/UAS-DspastinK467R) caused a decrease in LD number and area stained by BODIPY 493/503 (Fig 6A–6D). We then measured the total content of TAGs in Drosophila larvae with downregulated Dspastin as well as in larvae expressing the wild-type and the K467R mutant protein. We detected a drastic decrease in TAG levels when Dspastin function was reduced and significantly higher levels of TAGs in individuals overexpressing wild-type Dspastin (Fig 6E).
Dspastin is enriched in the nervous system during embryonic development [43], and was detected at the neuromuscular junction in adult flies [44]. We decided to analyze whether spastin overexpression or downregulation affects LDs in the skeletal muscle or in the peripheral nerves, which are more relevant for the disease pathogenesis. Tissues were labeled with BODIPY 493/503 to visualize LDs and with anti-acetylated α-tubulin to visualize the stable MT network. We found that depletion of Dspastin by using both RNAi and expression of the dominant-negative mutant causes an excessive stabilization of the MT network when induced specifically in larval body muscles (Mef2-Gal4/UAS-DspastinRNAi and Mef2-Gal4/UAS-DspastinK467R), as previously shown [41], and resulted in a significant decrease of the total number and size of LDs (Fig 7A–7C). Importantly, DspastinK467R localized at the LD surface (Fig 7D).
We then analyzed LDs in Drosophila peripheral nerves that comprise a central core of motor and sensory axons surrounded by peripheral glia and perineural glia [45]. To rule out the possibility that in the Drosophila nervous system the LD probes Nile red or BODIPY 493/503 may partition to degradative lysosomal compartments or accumulate in ER membranes, we expressed the ER marker GFP-KDEL or the lysosome marker GFP-Lamp in nerves labeled with LD dyes (S7A and S7B Fig). The absence of co-localization between Nile red and GFP-KDEL or GFP-Lamp indicated that lipid probes accumulate in the fat store organelles. To demonstrate that the majority of LDs visible in the nerves are found within neurons and not in glial cells, we labeled wild-type nerves by expressing UAS-mCD8-GFP under the control of the glia-specific driver repo-Gal4 to visualize glia cell membranes and anti-HRP antibodies to reveal axons (S7C Fig). We found that LDs are enriched in axonal projections compared to glia (S7C Fig), and their distribution and size are similar in animals of different control genotypes. Having established this, we selectively downregulated Dspastin in axons using the neuronal-specific promoter Elav-Gal4, and found that both Dspastin depletion (Elav-Gal4/UAS-DspastinRNAi) and expression of the dominant-negative mutation K467R (Elav-Gal4/UAS-DspastinK467R) reduced LD number in the nerves, without affecting LD size (Fig 7E–7G).
In contrast with results in fat bodies, overexpression of wild-type Dspastin in both the muscles (Mef2-Gal4/UAS-Dspastin-myc) and the nervous system (Elav-Gal4/UAS-Dspastin-myc) caused a remarkable increase in LD number, while LDs were slightly larger only in the muscle (Fig 7E–7G). As expected, upon Dspastin overexpression, bundled acetylated α-tubulin appeared thinned and shortened (Fig 7E). All together, we conclude that Dspastin affects LDs and lipid metabolism in vivo, although its effects can differ depending on the tissue examined.
The C. elegans spastin homologue SPAS-1 encodes a 512 amino acid protein, which lacks the N-terminal extension observed in spastin-M1 and Dspastin. SPAS-1 localizes to the cytoplasm and severs MTs efficiently [46]. To investigate whether the function of spastin in controlling LD metabolism in vivo is conserved in C. elegans, we downregulated spas-1 by RNA interference (RNAi) to about 40% of the control, as determined using quantitative real-time PCR (Fig 8A). SPAS-1 depleted worms showed normal life span and no visible defects on their motility similar to the deletion mutant reported [46]. We stained intestinal cells, the major fat stores in C. elegans, using oil red O, and found a significant reduction of signal intensity in spas-1 downregulated worms compared to the controls (Fig 8B and 8C). Consistently, the amount of TAGs was reduced at comparable levels (Fig 8D). Furthermore, we obtained a spas-1 mutant strain [spas-1 (tm683)], carrying an intragenic deletion that leads to the lack of the protein [46]. Oil red O staining, as well as TAG measurements confirmed a significant reduction of stored lipids (Fig 8B, 8E and 8F). Moreover, a reduction in LDs number was observed after crossing the spas-1 (tm683) strain with a reporter line expressing GFP-tagged DGAT-2, a marker of LDs (Fig 8G). These data support an evolutionary conserved role of spastin in regulating lipid metabolism, also independently from direct targeting of the protein to LDs.
Lipid composition of biological membranes affects several physiological processes of crucial relevance in neurons, such as endo- and exocytosis, trafficking, and dynamics of organelles. Moreover, lipid molecules may act as direct signaling effectors [47]. Here, we unravel a role of spastin in regulating LD formation and lipid metabolism in different model systems, and open the question of the significance of these findings for HSP pathogenesis.
We found that inhibition of Dspastin activity in Drosophila using either RNAi or the expression of a dominant-negative mutant reduces LD number in all the tissues that we examined and consistently decreased TAG levels. Similarly, depletion or deletion of SPAS-1 affected the amount of fatty acids and the number of LDs stored in the intestinal cells of C. elegans. Remarkably, these effects were observed in the fly also in the skeletal muscles and the nerves, two tissues that express Dspastin at very high level [43, 44]. Peripheral nerves are composed of several axons surrounded by the subperineurial glia [45]. Although we cannot totally exclude a non-autonomous contribution to the phenotype from the peripheral glial cells, our analysis suggests that LD alterations occur in axons, thus linking this phenotype to a tissue that is relevant for the human pathology. Downregulation of spastin in NSC34 cells did not visibly affect LD size, morphology, and number of LDs, or TAG levels. We posit that loss of spastin is compensated in vitro, in line with the remarkably late-onset and selective axonal phenotype of patients carrying mutations in the SPAST gene. Moreover, transient downregulation in vitro may not be sufficiently sustained to produce the same effects observed if RNAi is achieved during embryonic development in vivo. This notwithstanding, the results in vivo substantiate the hypothesis that spastin positively regulates LD formation.
Our data are of particular interest if considered together with previous results showing that deletion of the atlastin-1 homologue in both worms and flies decrease LD size and levels of TAGs [28]. Currently, there is no clue as to whether downregulation of atlastin-1 affects LD size and formation in mammalian cells. It has been proposed that atlastin-mediated fusion of ER membranes could either affect the formation of LDs or mediate direct fusion of LDs [28]. Since spastin-M1 interacts with both atlastin-1 and REEP1, we speculate that MT-severing exerted by spastin may be required in concert with atlastin-1 and REEP1 to shape not only the ER tubules, but also the specialized LD subcompartment that derives from these tubules. The tubular ER is aligned along MT tracks, and the positive bending of the outer leaflet of the ER membrane occurring during LD formation upon neutral lipid accumulation may be favored by concomitant local MT-severing. Such a model would require spastin to accumulate in regions of the ER where fatty acids are packaged to be delivered into LDs. Consistently, we found that after starvation exogenous spastin-M1 labels ER microdomains that are positive for markers of pre-LDs.
In agreement with a role of spastin as positive regulator of LD formation at the tubular ER, we found that overexpression of Dspastin in both muscles and nerves drastically increase the number of LDs. In HeLa cells and in Drosophila fat bodies, however, spastin overexpression led to a significant increase in the volume, but a decrease in the number of LDs. Fat bodies express Dspastin at low level, and HeLa cells mostly express spastin-M87. It is possible that overexpression of spastin in these tissues causes a dominant-negative effect, as often observed upon overexpression of AAA proteins, resulting in a defect in formation of LDs. These LDs could then become bigger either by direct transfer of lipids, or by fusing with neighboring LDs. A recent study found that overexpression of GRAF1a, a brain specific protein containing a BAR domain, also caused the appearance of bigger and less numerous LDs in HeLa cells [48]. Remarkably, cells overexpressing GRAF1a showed an enhanced clustering of LDs [48], a phenotype that we also observed when we expressed in HeLa cells a mutant spastin unable to bind MT. GRAF1a has been proposed to affect LD motility, by a yet unknown mechanism, and subsequently LD growth [48]. Future studies are required to investigate whether spastin levels affect LD motility by severing the MT tracks along which LDs move. Interestingly, previous studies have reported that LD expansion and formation can be inhibited by disrupting the MT network with nocodazole, or by depleting dynein, a retrograde MT motor [49, 50].
Our data highlight the fact that LD phenotypes may be affected by the cellular, tissue, or organismal context. Cell and tissue-specific differences in LD function and composition have been so far only marginally investigated. There are several other examples in the literature of discordant results in different experimental systems, complicating our understanding of LD physiology. Both overexpression and downregulation of spartin in cells increased the number of LDs [30], while in vivo studies in spartin knock-out mice have detected an increase of adipocyte number in female mice only [31]. Deletion of seipin in yeast led to few, supersized LDs [51, 52], while fibroblasts and lymphoblasts from patients carrying loss-of-function mutations in the seipin-encoding gene showed numerous small LDs [52]. In deletion mutants of Drosophila seipin, smaller LDs and reduced lipid storage were observed in the fat bodies, but larger LDs and increased fat storage were detected in the proventriculus and in the anterior gut, and ectopic LDs appeared in the salivary gland [53]. Therefore, a future challenge will be to assess a possible LD phenotype in a vertebrate model of spastin deficiency, most interestingly within the long cortico-spinal axons degenerating in HSP.
An important question that arises from our study is whether the LD phenotype that we observe upon manipulation of spastin levels in different models requires targeting of spastin to LDs. We show that human spastin-M1 and Drosophila Dspastin can be sorted to mature LDs. Efficient spastin-M1 targeting to LDs is observed under condition of overexpression and upon OA loading. At endogenous levels of expression most spastin-M1 co-fractionates with the ER membranes, however we demonstrate that spastin-M1 is present in the LD fraction. Since spastin-M1 is expressed at very low levels in several tissues and cell lines, it is not surprising that spastin was never detected in proteomic studies of LD protein content [54–58].
LD targeting is independent from the ability of spastin to mediate severing, and is not caused by the disruption of the ER morphology per se. Instead, we have identified a LD targeting signal in the N-terminal region exclusively present in the M1 spastin isoform. This motif comprises a hydrophobic domain interrupted by a crucial arginine residue, strongly suggesting that it is not mere hydrophobicity but the ability to acquire a hairpin topology that is critical for LD targeting of spastin. Our results identify spastin-M1 as belonging to a class of monotopic proteins, containing a hydrophobic stretch destabilized by positive residues, which are recruited efficiently to LDs after feeding cells with OA, while under normal conditions they reside in a different cell compartments, mostly the ER. Examples of such proteins, among others, are DGAT2, NSDHL, ALDI, and caveolins [59, 60]. In caveolin, targeting to LDs is further mediated by a positively charged sequence in combination to the hydrophobic stretch [39].
A hydrophobic domain interrupted by a positive residue is conserved in the N-terminus of Drosophila spastin, despite substantial sequence divergence with the human protein, while C. elegans SPAS-1 does not possess a clear hydrophobic region and whether it targets LDs should be investigated in more detail. All together, these data suggest that the evolutionary conserved role of spastin in LD metabolism is probably secondary to a primary role exerted at the level of the ER membranes and the MT-network, as it has been suggested in the case of atlastin-1 [28]. In mammals, however, spastin-M1 has acquired a bona fide LD targeting domain. What is the functional significance of spastin-M1 at the LD surface? Further studies are required to definitively answer this question. At this stage, we cannot exclude that LD targeting of spastin-M1 may be coupled to its degradation in vivo, as a mean to dispose of excess spastin that may become toxic at a certain concentration. Strikingly, the levels of spastin-M1 are very low and several mechanisms are in place to regulate the expression of this isoform [22, 23]. LDs have been proposed as sequestration sites to keep proteins inactive, or as hubs for protein degradation [61]. A prerequisite for the existence of such pathways is specific sorting of some proteins to these organelles, as we demonstrated for spastin. An attractive speculation is that LDs could be employed as sequestration or degradation platforms for proteins that are controlling biogenesis of these organelles, providing a feedback mechanism to regulate the activity or turnover of these proteins. It is remarkable that another HSP protein, spartin, seems to be implicated in degradation of specific targets on LDs, by recruiting E3-ubiquitin ligase to these organelles. In fact, there is evidence that one of these targets is PLIN2 [62].
Our results prompt further studies to investigate whether a function of spastin in LD metabolism is relevant in axons, which have limited capability to store and utilize neutral lipids. Virtually nothing is known on LD number, fate, and role in neurons. Since LDs have been proposed to distribute not only fatty acids, but also phospholipids to other membrane-bound organelles and to be enriched in proteins regulating membrane transport [63], it is conceivable that their role in neurons may be connected to inter-membranes lipid trafficking. Our findings add another piece of evidence to the emerging picture that an imbalance in lipid metabolism may contribute to the pathogenesis of HSP.
Spastin-M1-myc in pcDNA3 and the deletion construct Δ50-spastin-myc in pMT21 were previously described [14, 64]. The Flag-spastin construct contains the spastin-M1 coding sequence C-terminal to 3X Flag into the p3XFlag-CMV vector. Untagged spastin-M1 in pcDNA3 was previously described [22]. The coding region of human spastin-M1 or spastin-Δ86 was amplified by PCR using appropriate primers and subcloned in frame into mCherry-C3 vector (HindIII/BamHI). To generate spastin-ΔMBD lacking amino acids 270 to 328, an internal deletion was generated in pcDNA3-spastin-myc construct [14], using the strategy described in [65] (S1 Table). This clone was subsequently used as a template to amplify spastin-ΔMBD for subcloning in the mCherry-C3 vector. The spastin region between amino acids 57 and 86 (TM) was cloned in mCherry-N3 vector (XhoI/BamHI). All mutants were generated by site-directed mutagenesis and verified by DNA sequencing (S1 Table). The generation of GFP-PLIN2, GFP-PLIN3 and GFP-HPos is described elsewhere [4, 38]. Dspastin-myc was cloned in pMT21 vector.
Antibodies used for analysis by immunofluorescence or western blot were: mouse monoclonal anti-acetylated tubulin (Sigma-Aldrich); rabbit polyclonal anti-BiP (Cell Signaling); rabbit polyclonal anti-calnexin (Enzo Life Science); mouse monoclonal anti-Flag (Sigma-Aldrich); mouse monoclonal anti-myc (Santa Cruz); rabbit polyclonal anti-myc (Sigma-Aldrich); guinea pig polyclonal anti-PLIN2 [Progen); rabbit polyclonal against PLIN3 [38]; rabbit polyclonal anti-REEP5 (Proteintech); mouse monoclonal anti-spastin (6C6) (Sigma-Aldrich); rabbit polyclonal against residues 87–354 of human spastin [64].
HeLa and COS7 cells were grown in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS). NSC34 cells [66] were cultured in DMEM supplemented with 5% defined FBS. Stealth small interfering RNAs (siRNAs) were synthesized by Invitrogen with the following sequences: Spastin (Spast): 5´-CCAGUGAGAUGAGAAAUAUUCGAUU-3´; Exon4 (Ex4): 5´-CGGACGUCUAUAACGAGAGUACUAA-3´, Control (C): Stealth RNAi negative control LO GC. Transfection of DNA constructs or siRNA duplexes (100 nM) was performed with Lipofectamine 2000 (Invitrogen). To induce LD formation, 400 μM OA (Sigma-Aldrich) complexed to fatty acid-free BSA (Sigma-Aldrich) was added to the culture medium overnight. In case of GFP-HPos transfection, the cells were starved in DMEM, L-glutamine, pyruvate, and nonessential amino acids but in the absence of serum for 24 h beginning directly after transfection. Additional methods can be found in S1 Text.
72 hours after downregulation, NSC34 cells were lysed in a buffer containing 50 mM Tris/HCl pH 7.4, 150 mM NaCl, 1 mM EDTA, 1% NP-40, 0.25% deoxycholic acid sodium salt, freshly supplemented with protease inhibitor cocktail (Sigma-Aldrich). Lysates were centrifuged at 20,000 g for 30 min and supernatant was collected for further western blot analysis.
For indirect immunofluorescence, cells grown onto glass coverslips were fixed with 4% paraformaldehyde for 30 min, incubated with 50 mM NH4Cl for 10 min and permeabilized with 0.5% saponin in PBS for 10 min. After 10 min in blocking solution (0.1% saponin, 10% pig serum in PBS), primary antibodies diluted in 0.1% saponin, 1% pig serum in PBS were applied to cells for 3 h. Cells were washed three times with PBS and secondary antibodies were applied to cells for 1 h. Finally, cells were washed once with PBS containing DAPI, twice with PBS alone and then samples were mounted using FluorSave Reagent (Calbiochem). When LDs were stained, BODIPY 493/503 (5 μM, Invitrogen) was applied either in the washing step together with DAPI, or administered to living cells for 20 min before fixation. Fluorescent images were acquired using a 63x NA 1.4 oil objective and Axio-Imager M2 microscope equipped with Apotome 2 (Zeiss) and processed using AxioVision software. Photographs show individual Z-stacks or projected images, as indicated in figure legends. Brightness levels were adjusted for image presentation using AxioVision within the linear range.
HeLa cells grown on glass-bottom dishes (MatTek corporation) and transfected with mCherry-spastin-M1 were imaged 8 h post-transfection with a spinning disc confocal microscope (Ultraview Vox, Perkin Elmer) using a 63x NA 1.49 oil immersion objective. Images were acquired and processed with the Volocity software (version 6.1, Perkin Elmer). BODIPY 493/503 (5 μM, Invitrogen) was added to the medium 6 h and washed out 7 h post-transfection.
Quantification of LDs was performed on merged Z-stack images acquired using an Axio-Imager M2 microscope equipped with Apotome 2 (Zeiss). Individual transfected cells were selected as Region of interest (ROI) by cropping and ROIs were then thresholded using the automated function to find objects with a minimum volume of 0.6 μm³ from Volocity software (version 6.1, Perkin Elmer). The thresholded images were used to quantify the total LD volume per cell. Using the same images we counted manually the LD number per cell. To obtain the average LD volume per cell, total LD volume per cell was divided by number of LD per cell. LDs were classified as clustered when more than 50% of the LDs were clustered on one side of the cell, as intermediate, when at least 50% of the LDs showed a dispersed distribution, and as dispersed, when LD distribution was undistinguishable from that of non-transfected neighboring cells.
NSC34 and HeLa cells were treated with 400 μM OA overnight to induce LD formation. Cells were harvested, washed twice with PBS, and resuspended in cold lysis buffer A (20 mM Tris/HCl pH 7.4, 1 mM EDTA) containing freshly added protease inhibitors (1 mM sodium orthovanadate, 1 mM NaF, 1 μg/ml Leupeptin, 10 μg/ml Aprotinin, 1 mM PMSF and 1x complete Mini Protease Inhibitor Cocktail from Roche). Lysate was passed through a 24-gauge needle, centrifuged at 500 g for 5 min, and the supernatant was collected (Input). 12 mg of the input for NSC34, and 3 mg in the case of HeLa cells was fractionated on a 20%-5%-0% sucrose gradient via centrifugation at 40,000 rpm (SW41 rotor; Beckman Coulter) for 3 h. Then, the tube was placed in a tube slicer (Beckman Coulter) and cut 0.75 cm from the top of the gradient. LD fraction floating in the top and all following fractions (1 ml each) were collected, precipitated with 10% final concentration of trichloroacetic acid (TCA) on ice and washed three times with acetone. For NSC34 cells, dried TCA pellets were resuspended in 50 μl (for LD fraction) or 250 μl (all other fractions) sample buffer (125 mM Tris/HCl pH 6.8, 4% SDS, 20% glycerol, 0.02% bromophenol blue, 2% β-mercaptoethanol). Pellet resulting from centrifugation was directly resuspended in 250 μl sample buffer. 50 μl of each sample were analyzed via western blotting. For HeLa cells, dried TCA pellets were resuspended in 200 μl (for LD fraction) or 1 ml (all the other fractions) sample buffer. 200 μl of each sample were analyzed via western blotting.
The UAS-DspastinRNAi Drosophila line used in this study was described previously [41]. UAS-Dspastin-myc and UAS-DspastinK467R-myc were generated by adding a myc epitope tag to the C-terminus of Dspastin and DspastinK467R constructs previously reported [41]. Six independent transgenic lines were derived for the K467R mutant and 5 for wild-type Dspastin. All lines were tested for protein expression by immunohistochemistry using different Gal4 driver lines. The two lines with the highest expression levels were chosen. The Gal4 activator lines Elav-Gal4, Mef2-Gal4, actin-Gal4, repo-Gal4 and the transgenic lines UAS-GFP-KDEL, UAS-GFP-Lamp and UAS-mCD8-GFP were obtained from the Bloomington Stock Center, Indiana University. Experimental crosses were performed at 28°C.
Drosophila immunostaining was performed on wandering third instar larvae reared at 28°C as previously described [41]. To visualize and determine the number and size of LDs, BODIPY 493/503 or Nile red positive structures in proximal axons, muscles and fat bodies of third instar larvae were imaged using a Nikon EZ-C1 confocal microscope equipped with a Nikon Plan APO 60.0×/1.40 oil immersion objective. Z-stacks with a step size of 0.5 μm were taken using identical settings. Each stack consisted of 15 to 20 plane images of 10 animals per genotype. The area of nerves and muscles and the area and number of LDs were calculated with ImageJ particle analyzer tool. The data collected were analyzed using Microsoft Office Excel 2007. The diameter of LDs was classified into different classes: 0–0.50 μm, 0.51–1 μm and >1.01 μm for muscle quantification; 0–0.50 μm, 0.51–1 μm, 1.01–1.5 and >1.51 μm for neuronal analysis; 0–5 μm, >5.01 μm for fat bodies experiments. Statistical analysis was performed with GraphPad Prism 3.03 software. Unpaired t-test was used to assess the differences in the number and area of LDs, while Mann-Whitney U test was used to assess the differences in the size distribution of LDs. Differences were considered statistically significant at p<0.05 (*) and p<0.005 (**). To determine total triglyceride 20 third instar larvae were homogenized in 250 μl PBST (0.05% Tween 20), incubated at 70°C for 10 min and then centrifuged at 3500 g for 3 min. Triglyceride amount in the hemolymph supernatant was measured using Serum Triglyceride Determination Kit (Sigma-Aldrich) as described [67].
C. elegans animals were grown on 20°C using standard procedures [68]. The wild-type strain was N2 Bristol strain. The mutant FX683 spas-1(tm683) was outcrossed four times before performing experiments. The transgenic strain used in the study is VS29 hjSi56 [vha-6p::3xFLAG::TEV::GFP::dgat-2::let-858 3’UTR]. The strains were kindly provided by Caenorhabditis Genetics Center (University of Minnesota, Minneapolis, MN).
Worms were fed either E. coli (HT115) containing an empty vector or E. coli expressing dsRNA against spas-1 (C24B5.2) gene from Ahringer library, as previously described [69]. Briefly, an overnight culture of bacteria containing RNAi plasmids was resuspended and grown to OD of 0.5 and then induced with 1 mM IPTG. For each experiment, three independent cultures were prepared and regarded as individual replicates. Clones were verified by sequencing and unc-54 RNAi was used as RNAi control. Worms were exposed to RNAi from hatching and collected on the first day of adulthood for experiments.
Worms were collected from a 9 mm plate and total RNA was isolated with Trizol (Invitrogen). DNAse treatment was performed using DNA-freeTM, DNAse treatment & removal (Ambion, Life technologies) according to the manufacturer’s protocol. RNA was quantified by spectrophotometry and 0.8 μg of total RNA was reversely transcribed using High Capacity cDNA Reverse Transcription Kit (Applied Biosystems). For each condition eight independent samples were prepared. RT-PCR was performed by the Step One Plus Real-Time PCR Systems (Applied Biosystems) with the following PCR conditions: 3 min at 95°C, followed by 40 cycles of 5 sec at 95°C and 15 sec at 60°C. Amplified products were detected with SYBR Green (Brilliant III Ultra Fast SYBR Green qPCR Master Mix, Agilent Technologies). Relative quantification was performed against act-1. Primers used for analysis are as follows: spas-1 primer pair 1: 5´-CCCGGAGAAGTGAAATCAGA-3´ and 5´-TGGTGCTGTGGCTCTTGTAG-3´; spas-1 primer pair 2: 5´-TTTCCCGAAACGAATTATGC-3´ and 5´-TTCGATCTGTCGATTTCACG-3´; act-1 primer pair: 5´-TCGTCCTCGACTCTGGAGAT-3´ and 5´-GCCATTTCTTGCTCGAAGTC-3´.
200–300 day-1 adult animals synchronized by egg-laying were permeabilized with 2x MRWB (160 mM KCl, 40 mM NaCl, 14 mM Na2EGTA, 1 mM spermidine-HCl, 0.4 mM spermine, 30 mM Na-PIPES pH 7.4, 0.2% β-mercaptoethanol) buffer containing 2% paraformaldehyde and stained with oil red O overnight. Animals were mounted and imaged with Axio-Imager M2 microscope outfitted with DIC optics (Zeiss) and processed using AxioVision software. Quantification of oil red O signal was performed using ImageJ software as described [70]. To image GFP-DGAT-2 vesicles, on the first day of adulthood, worms were placed on 2% agarose pads and immobilized with 50 mM Na-azide in M9 buffer (42 mM Na2HPO4, 22 mM KH2PO4, 86 mM NaCl, 1 mM MgSO4). Images were obtained on a spinning disc confocal microscope (Ultraview Vox, Perkin Elmer) using a 63x NA 1.49 oil immersion objective. Image stacks of proximal part of the gut were captured. Maximum intensity projections were obtained using Volocity software (version 6.1, Perkin Elmer). The number of GFP positive vesicles in a single proximal cell of the gut was determined using ImageJ software.
200–300 day-1 adult animals synchronized by egg-laying were collected in M9 buffer (42 mM Na2HPO4, 22 mM KH2PO4, 86 mM NaCl, 1 mM MgSO4) and subjected to freeze thaw cycles in liquid nitrogen twice, followed by constant sonication for 3 min and debris precipitation at 1300 g for 1 min at 4°C. Quantification of triglyceride was performed using EnzyChrom Triglyceride Assay Kit (Bioassay Systems) according to the manufacturer’s protocol and were normalized to total protein content determined using Bradford assay (Bio-Rad).
|
10.1371/journal.pcbi.1003727 | Memory Capacity of Networks with Stochastic Binary Synapses | In standard attractor neural network models, specific patterns of activity are stored in the synaptic matrix, so that they become fixed point attractors of the network dynamics. The storage capacity of such networks has been quantified in two ways: the maximal number of patterns that can be stored, and the stored information measured in bits per synapse. In this paper, we compute both quantities in fully connected networks of N binary neurons with binary synapses, storing patterns with coding level , in the large and sparse coding limits (). We also derive finite-size corrections that accurately reproduce the results of simulations in networks of tens of thousands of neurons. These methods are applied to three different scenarios: (1) the classic Willshaw model, (2) networks with stochastic learning in which patterns are shown only once (one shot learning), (3) networks with stochastic learning in which patterns are shown multiple times. The storage capacities are optimized over network parameters, which allows us to compare the performance of the different models. We show that finite-size effects strongly reduce the capacity, even for networks of realistic sizes. We discuss the implications of these results for memory storage in the hippocampus and cerebral cortex.
| Two central hypotheses in neuroscience are that long-term memory is sustained by modifications of the connectivity of neural circuits, while short-term memory is sustained by persistent neuronal activity following the presentation of a stimulus. These two hypotheses have been substantiated by several decades of electrophysiological experiments, reporting activity-dependent changes in synaptic connectivity in vitro, and stimulus-selective persistent neuronal activity in delayed response tasks in behaving monkeys. They have been implemented in attractor network models, that store specific patterns of activity using Hebbian plasticity rules, which then allow retrieval of these patterns as attractors of the network dynamics. A long-standing question in the field is how many patterns (or equivalently, how much information) can be stored in such networks? Here, we compute the storage capacity of networks of binary neurons and binary synapses. Synapses store information according to a simple stochastic learning process that consists of transitions between synaptic states conditioned on the states of pre- and post-synaptic neurons. We consider this learning process in two limits: a one shot learning scenario, where each pattern is presented only once, and a slow learning scenario, where noisy versions of a set of patterns are presented multiple times, but transition probabilities are small. The two limits are assumed to represent, in a simplified way, learning in the hippocampus and neocortex, respectively. We show that in both cases, the information stored per synapse remains finite in the large limit, when the coding is sparse. Furthermore, we characterize the strong finite size effects that exist in such networks.
| Attractor neural networks have been proposed as long-term memory storage devices [1], [2], [3]. In such networks, a pattern of activity (the set of firing rates of all neurons in the network) is said to be memorized if it is one of the stable states of the network dynamics. Specific patterns of activity become stable states thanks to synaptic plasticity mechanisms, including both long term potentiation and depression of synapses, that create positive feed-back loops through the network connectivity. Attractor states are consistent with the phenomenon of selective persistent activity during delay periods of delayed response tasks, which has been documented in numerous cortical areas in behaving monkeys [4], [5], [6], [7]. A long standing question in the field has been the question of the storage capacity of such networks. Much effort has been devoted to compute the number of attractor states that can be imprinted in the synaptic matrix, in networks of binary neurons [8], [9], [10], [11]. Models storing patterns with a covariance rule [12], [1], [8], [11] were shown to be able to store a number of patterns that scale linearly with the number of synapses per neuron. In the sparse coding limit (in which the average fraction of selective neurons per pattern goes to zero in the large limit), the capacity was shown to diverge as . These scalings lead to a network storing on the order of 1 bit per synapse, in the large limit, for any value of the coding level. Elizabeth Gardner [10] computed the maximal capacity, in the space of all possible coupling matrices, and demonstrated a similar scaling for capacity and information stored per synapse.
These initial studies, performed on the simplest possible networks (binary neurons, full connectivity, unrestricted synaptic weights) were followed by a second wave of studies that examined the effect of adding more neurobiological realism: random diluted connectivity [9], neurons characterized by analog firing rates [13], learning rules in which new patterns progressively erase the old ones [14], [15]. The above mentioned modifications were shown not to affect the scaling laws described above. One particular modification however was shown to have a drastic effect on capacity. A network with binary synapses and stochastic on-line learning was shown to have a drastically impaired performance, compared to networks with continuous synapses [16], [17]. For finite coding levels, the storage capacity was shown to be on the order of , not stored patterns, while the information stored per synapse goes to zero in the large limit. In the sparse coding limit however (), the capacity was shown to scale as , and therefore a similar scaling as the Gardner bound, while the information stored per synapse remains finite in this limit. These scaling laws are similar to the Willshaw model [18], which can be seen as a particular case of the Amit-Fusi [17] rule. The model was then subsequently studied in greater detail by Huang and Amit [19], [20] who computed the storage capacity for finite values of , using numerical simulations and several approximations for the distributions of the ‘local fields’ of the neurons. However, computing the precise storage capacity of this model in the large limit remains an open problem.
In this article we focus on a model of binary neurons where binary synapses are potentiated or depressed stochastically depending on the states of pre and post synaptic neurons [17]. We first introduce analytical methods that allow us to compute the storage capacity in the large limit, based on a binomial approximation for the synaptic inputs to the neurons. We first illustrate it on the Willshaw model and to recover the well-known result on the capacity of this model [18], [21], [22]. We then move to a stochastic learning rule, in which we study two different scenarios: (i) in which patterns are presented only once - we will refer to this model as the SP (Single Presentation) model [17]; (ii) in which noisy versions of the patterns are presented multiple-times - the MP (Multiple presentations) model [23]. For both models we compute the storage capacity and the information stored per synapse in the large limit, and investigate how they depend on the various parameters of the model. We then study finite size effects, and show that they have a huge effect even in networks of tens of thousands of neurons. Finally we show how capacity in finite size networks can be enhanced by introducing inhibition, as proposed in [19], [20]. In the discussion we summarize our results and discuss the relevance of the SP and MP networks to memory maintenance in the hippocampus and cortex.
The capacity of the Willshaw model has already been studied by a number of authors [18], [21], [22]. Here, we present the application of the analysis described in the previous section to the Willshaw model, for completeness and comparison with the models described in the next sections. In this model, after presenting patterns to the network, the synaptic matrix is described as follows: if at least one of the presented patterns had neuron and co-activated, otherwise. Thus, after the learning phase, we have,(21)
Saturating the inequalities (19,20) with fixed, one obtains the information stored per synapse,(22)
The information stored per synapse is shown as a function of in Figure 1a. A maximum is reached for at , but goes to zero in both the and limits. The model has a storage capacity comparable to its maximal value, in a large range of values of (between and ). We can also optimize capacity for a given value of , as shown in Figure 1b. It reaches its maximum at , and goes to zero in the small and large limits. Again, the model has a large storage capacity for a broad range of , for between and .
Previous studies [18], [21] have found an optimal capacity of . Those studies focused on a feed-forward network with a single output neuron, with no fluctuations in the number of selective neurons per pattern, and required that the number of errors on silent outputs is of the same order as the number of selective outputs in the whole set of patterns. In the calculations presented here, we have used a different criteria, namely that a given pattern (not all patterns) is exactly a fixed point of the dynamics of the network with a probability that goes to one in the large limit. Another possible definition would be to require that all the patterns are exact fixed points with probability one. In this case, for patterns with fixed numbers of selective neurons, the capacity drops by a factor of , , as already computed by Knoblauch et al [22].
A drawback of the Willshaw learning rule is that it only allows for synaptic potentiation. Thus, if patterns are continuously presented to the network, all synapses will eventually be potentiated and no memories can be retrieved. In [17] Amit and Fusi introduced a new learning rule that maintains the simplicity of the Willshaw model, but allows for continuous on-line learning. The proposed learning rule includes synaptic depression. At each learning time step , a new pattern with coding level is presented to the network, and synapses are updated stochastically:
for synapses such that :
if , then is potentiated to 1 with probability ; and if it stays at .
for synapses such that :
if , then stays at ; and if it is depressed to with probability .
The evolution of a synapse during learning can be described by the following Markov process:(23)where is the probability that a silent synapse is potentiated upon the presentation of pattern and is the probability that a potentiated synapse is depressed. After a sufficient number of patterns has been presented the distribution of synaptic weights in the network reaches a stationary state. We study the network in this stationary regime.
For the information capacity to be of order 1, the coding level has to scale as , as in the Willshaw model, and the effects of potentiation and depression have to be of the same order [17]. Thus we define the depression-potentiation ratio as,(24)
We can again use Eq. (9) and the saturated inequalities (19,20) to compute the maximal information capacity in the limit . This requires computing and , defined in the previous section, as a function of the different parameters characterizing the network. We track a pattern that has been presented time steps in the past. In the following we refer to as the age of the pattern. In the sparse coding limit, corresponds to the probability that a synapse is potentiated. It is determined by the depression-potentiation ratio ,(25)and(26)where . Our goal is to determine the age of the oldest pattern that is still a fixed point of the network dynamics, with probability one. Note that in this network, contrary to the Willshaw model in which all patterns are equivalent, here younger patterns, of age , are more strongly imprinted in the synaptic matrix, , and thus also stored with probability one.
Choosing an activation threshold and a coding level that saturate inequalities (19) and (20), information capacity can be expressed as:(27)
The optimal information is reached for which gives .
The dependence of on the different parameters is shown in Figure 2. Panel a shows the dependence on the fraction of activated synapses in the asymptotic learning regime. Panels b, c and d show the dependence on , and . Note from panel c that there is a broad range of values of that give information capacities similar to the optimal one. One can also observe that the optimal information capacity is about times lower in the SP model than in the Willshaw model. This is the price one pays to have a network that is able to continuously learn new patterns. However, it should be noted that at maximal capacity, in the Willshaw model, every pattern has a vanishing basin of attraction while in the SP model, only the oldest stable patterns have vanishing basins of attraction. This feature is not captured by our measure of storage capacity.
In the SP model, patterns are presented only once. Brunel et al [23] studied the same network of binary neurons with stochastic binary synapses but in a different learning context, where patterns are presented multiple times. More precisely, at each learning time step , a noisy version of one of the prototypes is presented to the network,(28)
Here is a noise level: if , presented patterns are identical to the prototypes, while if , the presented patterns are uncorrelated with the prototypes. As for the SP model this model achieves a finite non-zero information capacity in the large limit if the depression-potentiation ratio is of order one, and if the coding level scales with network size as . If learning is slow, , and the number of presentations of patterns of each class becomes large the probabilities and are [23]:(29)and(30)
We inserted those expressions in Eqs. (19,20) to study the maximal information capacity of the network under this learning protocol. The optimal information bits/synapse is reached at for which gives . In this limit, the network becomes equivalent to the Willshaw model.
The maximal capacity is about times larger than for a network that has to learn in one shot. On Figure 3a we plot the optimal capacity as a function of . The capacity of the slow learning network with multiple presentations is bounded by the capacity of the Willshaw model for all values of , and it is reached when the depression-potentiation ratio . For this value, no depression occurs during learning: the network loses palimpsest properties, i.e. the ability to erase older patterns to store new ones, and it is not able to learn if the presented patterns are noisy. The optimal capacity decreases with , for instance at (as many potentiation events as depression events at each pattern presentation), . Figure 3c shows the dependence as a function of . In Figure 3d, we show the optimized capacity for different values of the noise in the presented patterns. This quantifies the trade-off between the storage capacity and the generalization ability of the network [23].
The results we have presented so far are valid for infinite size networks. Finite-size effects can be computed for the three models we have discussed so far (see Methods). The main result of this section is that the capacity of networks of realistic sizes is very far from the large N limit. We compute capacities for finite networks in the SP and MP settings, and we validate our finite size calculations by presenting the results of simulations of large networks of sizes , .
We summarize the finite size calculations for the SP model (a more general and detailed analysis is given in Methods). In the finite network setting, conditional on the tested pattern having selective neurons, the probability of no error is given bywith
(31)
where and is given by Eq. (13). In the calculations for discussed in the previous sections we kept only the dominant term in , which yields Eqs. (19) and (20).
In the above equations, the first order corrections scale as , which has a dramatic effect on the storage capacity of finite networks. In Figure 4a,b, we plot (where the bar denotes an average over the distribution of ) as a function of the age of the pattern, and compare this with numerical simulations. It is plotted for and for learning and network parameters chosen to optimize the storage capacity of the infinite-size network (see Section ‘Amit-Fusi model’). We show the result for two different approximations of the field distribution: a binomial distribution (magenta), as used in the previous calculations for infinite size networks; and a gaussian (red) approximation (see Methods for calculations) as used by previous authors [19], [20], [24]. For these parameters the binomial approximation gives an accurate estimation of , while the gaussian calculation overestimates it.
The curves we get are far from the step functions predicted for by Eq. (45). To understand why, compare Eqs. (15), and (31): finite size effects can be neglected when and . Because the finite size effects are of order , it is only for huge values of that the asymptotic capacity can be recovered. For instance if we choose an activation threshold slightly above the optimal threshold given in Section ‘Amit-Fusi model’ (), then , and for we only have . In Figure 4c we plot as a function of where is the value of that optimizes capacity in the large limit, and the other parameters are the one that optimizes capacity. We see that we are still far from the large limit for . Networks of sizes have capacities which are only between 20% and 40% of the predicted capacity in the large limit. Neglecting fluctuations in the number of selective neurons, we can derive an expression for the number of stored patterns that includes the leading finite size correction for the SP model,(32)where and are two constants (see Methods).
If we take fluctuations in the number of selective neurons into account, it introduces other finite-size effects as can be seen from Eqs. (43) and (44) in the Methods section. These fluctuations can be discarded if and . In Figure 4d we plot for different values of N. We see that finite size effects are even stronger in this case.
To plot the curves of Figure 4, we chose parameters to be those that optimize storage capacity for infinite network sizes. When is finite, those parameters are no longer optimal. To optimize parameters at finite , since the probability of error as a function of age is no longer a step function, it is not possible to find the last pattern stored with probability one. Instead we define the capacity as the pattern age for which . Using Eqs. (31) and performing an average over the distribution of , we find parameters optimizing pattern capacity for fixed values of . Results are shown on Figure 5a,b for and . We show the results for the different approximations used to model the neural fields: the blue line is the binomial approximation, the cyan line the gaussian approximation and the magenta one is a gaussian approximation with a covariance term that takes into account correlations between synapses (see Methods and [19], [20]). For the storage capacity of simulated networks (black crosses) is well predicted by the binomial approximation while the gaussian approximations over-estimates capacity. For , the correlations between synapses can no longer be neglected [17]. The gaussian approximation with covariance captures the drop in capacity at large .
For , the SP model can store a maximum of patterns at a coding level (see blue curve in figure 5c). As suggested in Figures 4c,d, the capacity of finite networks is strongly reduced compare to the capacity predicted for infinite size networks. More precisely, if the network of size had the same information capacity as the infinite size network (27), it would store up to patterns at coding level . Part of this decrease in capacity is avoided if we consider patterns that have a fixed number of selective neurons. This corresponds to the red curve in figure 4c. For fixed sizes the capacity is approximately twice as large. Note that finite-size effects tend to decrease as the coding level increases. In Figure 5c, , and the capacity is of the value predicted by the large limit calculation. The ratio of actual to asymptotic capacities increases to at and at . In Figure 5d, we do the same analysis for the MP model with . Here we have also optimized all the parameters, except for the depression-potentiation ratio which is set to , ensuring that the network has the palimpsest property and the ability to deal with noisy patterns. For , the MP model with can store up to patterns, at (versus at for the SP model). One can also compute the optimized capacity for a given noise level. At , for and or at , for and .
So far, we have defined the storage capacity as the number of patterns that can be perfectly retrieved. However, it is quite common for attractor neural networks to have stable fixed point attractors that are close to, but not exactly equal to, patterns that are stored in the connectivity matrix. It is difficult to estimate analytically the stability of patterns that are retrieved with errors as it requires analysis of the dynamics at multiple time steps. We therefore used numerical simulations to check whether a tested pattern is retrieved as a fixed point of the dynamics at a sufficiently low error level. To quantify the degree of error, we introduce the overlap between the network fixed point and the tested pattern , with selective neurons(33)
In Figure 6a we show , the number of fixed-point attractors that have an overlap larger than with the corresponding stored pattern, for , and . Note that only a negligible number of tested patterns lead to fixed points with smaller than , for neurons. Considering fixed points with errors leads to a substantial increase in capacity, e.g. for the capacity increases from to . In Figure 6b, we quantify the information capacity in bits stored per synapse, defined as in Eq. (6), . Note that in the situation when retrieval is not always perfect this expression is only an approximation of the true information content. The coding level that optimizes the information capacity in bits per synapse is larger () than the one that optimizes the number of stored patterns (), since the information content of individual patterns decreases with . Finally, note that the information capacity is close to its optimum in a broad range of coding levels, up to .
As we have seen above, the fluctuations in the number of selective neurons in each pattern lead to a reduction in storage capacity in networks of finite size (e.g. Figure 5c,d). The detrimental effects of these fluctuations can be mitigated by adding a uniform inhibition to the network [19]. Using a simple instantaneous and linear inhibitory feed-back, the local fields become(34)
For infinite size networks, adding inhibition does not improve storage capacity since fluctuations in the number of selective neurons vanish in the large N limit. However, for finite size networks, minimizing those fluctuations leads to substantial increase in storage capacity. When testing the stability of pattern , if the number of selective neurons is unknown, the variance of the field on non-selective neurons is , and for selective neurons (for small ). The variance for non-selective neurons is minimized if , yielding the variance obtained with fixed size patterns. The same holds for selective neurons at . Choosing a value of between and brings the network capacity towards that of fixed size patterns. In Figure 7a, we show the storage capacity as a function of for these three scenarios. Optimizing the inhibition increases the maximal capacity by (green curve) compared to a network with no inhibition (blue curve). Red curve is the capacity without pattern size fluctuations. Inhibition increases the capacity from at to . In Figure 7b, information capacity measured in bits per synapse is shown as a function of in the same three scenarios. Note again that for , the capacity is quite close to the optimal capacity.
We have presented an analytical method to compute the storage capacity of networks of binary neurons with binary synapses in the sparse coding limit. When applied to the classic Willshaw model, in the infinite limit, we find a maximal storage capacity of , the same than found in previous studies, although with a different definition adapted to recurrent networks, as discussed in the section ‘Willshaw model’. We then used this method to study the storage capacity of a network with binary synapses and stochastic learning, in the single presentation (SP) scenario [17]. The main advantage of this model, compared to the Willshaw model, is its palimpsest property, that allows it to do on-line learning in an ever changing environment. Amit and Fusi showed that the optimal storage capacity was obtained in the sparse coding limit, and with a balance between the effect of depression and potentiation. The storage capacity of this network has been further studied for finite size networks in [19], [20]. We have complemented this work by computing analytically the storage capacity in the large limit. The optimal capacity of the SP model is , which is about times lower than the one of the Willshaw model. This decrease in storage capacity is similar to the decrease seen in palimpsest networks with continuous synapses - for example, in the Hopfield model the capacity is about , while in a palimpsest version the capacity drops to about . The reason for this decrease is that the most recently seen patterns have large basins of attraction, while older patterns have smaller ones. In the Willshaw model, all patterns are equivalent, and therefore they all have vanishing basins of attraction at the maximal capacity.
We have also studied the network in a multiple presentation (MP) scenario, with in which patterns presented to the network are noisy versions of a fixed set of prototypes, in the slow learning limit in which transition probabilities go to zero [23]. In the extreme case in which presented patterns are the prototypes, all synaptic weights are initially at zero, and if the synapses do not experience depression, this model is equivalent to the Willshaw model with a storage capacity of , which is about times larger than the capacity of the SP model. A more interesting scenario is when depression is present. In this case then the network has generalization properties (it can learn prototypes from noisy versions of them), as well as palimpsest properties (if patterns drawn from a new set of prototypes are presented it will eventually replace a previous set with the new one). We have quantified the trade-off between generalization and storage capacity (see Figure 3d). For instance, if the noisy patterns have of their selective neurons in common with the prototypes to be learned, the storage capacity is decreased from to .
A key step in estimating storage capacity is deriving an accurate approximation for the distribution of the inputs neurons receive. These inputs are the sum of a large number of binary variables, so the distribution is a binomial if one can neglect the correlations between these variables, induced by the learning process. Amit and Fusi [17] showed that these correlations can be neglected when . Thus, we expect the results with the binomial approximation to be exact in the large limit. We have shown that a Gaussian approximation of the binomial distribution gives inaccurate results in the sparse coding limit, because the capacity depends on the tail of the distribution, which is not well described by a Gaussian. For larger coding levels (), the binomial approximation breaks down because it does not take into account correlations between inputs. Following [19] and [20], we use a Gaussian approximation that includes the covariance of the inputs, and show that this approximation captures well the simulation results in this coding level range.
We computed storage capacities for two different learning scenarios. Both are unsupervised, involve a Hebbian-type plasticity rule, and allow for online learning (providing patterns are presented multiple times for the MP model). It is of interest to compare the performance of these two particular scenarios with known upper bounds on storage capacity. For networks of infinite size with binary synapses such a bound has been derived using the Gardner approach [25]. In the sparse coding limit, this bound is with random patterns (in which fluctuations in the number of selective neurons per pattern fluctuates), and if patterns have a fixed number of selective neurons [26]. We found a capacity of for the SP model and for the MP model, obtained both for patterns with fixed and variable number of selective neurons. The result for the MP model seems to violate the Gardner bound. However, as noticed by Nadal [21], one should be cautious in comparing these results: in our calculations we have required that a given pattern is stored perfectly with probability one, while the Gardner calculation requires that all patterns are stored perfectly with probability one. As mentioned in the section ‘Willshaw model’, the capacity of the Willshaw and MP models drops to in the case of fixed-size patterns, if one insists that all patterns should be stored perfectly, which is now consistent with the Gardner bound. This means that the MP model is able to reach a capacity which is roughly half the Gardner bound, a rather impressive feat given the simplicity of the rule. Note that supervised learning rules can get closer to these theoretical bounds [27].
We have also studied finite-size networks, in which we defined the capacity as the number of patterns for which the probability of exact retrieval is at least 50%. We found that networks of reasonable sizes have capacities that are far from the large limit. For networks of sizes storage capacities are reduced by a factor or more (see Figure 4). These huge finite size effects can be understood by the fact that the leading order corrections in the large limit are in - and so can never be neglected unless is an astronomical number (see Methods). A large part of the decrease in capacity when considering finite-size networks is due to fluctuations in the number of selective neurons from pattern to pattern. In the last section, we have used inhibition to minimize the effect of these fluctuations. For instance, for a network of neurons learning in one shot, inhibition allows to increase capacity from to . For finite size networks, memory patterns that are not perfectly retrieved can still lead to fixed points where the activity is significantly correlated with the memory patterns. We have investigated with simulations how allowing errors in the retrieved patterns modifies storage capacity. For , the capacity increases from to , i.e. by approximately 30%.
Our study focused on networks of binary neurons, connected through binary synapses, and storing very sparse patterns. These three assumptions allowed us to compute analytically the storage capacity of the network in two learning scenarios. An important question is how far real neural networks are from such idealized assumptions. First, the issue of whether real synapses are binary, discrete but with a larger number of states, or essentially continuous, is still unresolved, with evidence in favor of each of these scenarios [28], [29], [30], [31], [32]. We expect that having synapses with a finite number of states will not modify strongly the picture outlined here [17], [33], [20]. Second, it remains to be investigated how these results will generalize to networks of more realistic neurons. In strongly connected networks of spiking neurons operating in the balanced mode [34], [35], [36], [37], the presence of ongoing activity presents strong constraints on the viability of sparsely coded selective attractor states. This is because ‘non-selective’ neurons are no longer silent, but are rather active at low background rates, and the noise due to this background activity can easily wipe out the selective signal [35], [38]. In fact, simple scaling arguments in balanced networks suggest the optimal coding level would become [3], [39]. The learning rules we have considered in this paper lead to a vanishing information stored per synapse with this scaling. Finding an unsupervised learning rule that achieves a finite information capacity in the large limit in networks with discrete synapses for such coding levels remains an open question. However, the results presented here show that for networks of realistic sizes, the information capacity at such coding levels is in fact not very far from the optimal one that is reached at lower coding levels (see vertical lines in Figure 5–7). Finally, the coding levels of cortical networks during delay period activity remain poorly characterized. Experiments in IT cortex [40], [41], [42] are consistent with coding levels of order 1%. Our results indicate that in networks of reasonable sizes, these coding levels are not far from the optimal values.
The SP and MP models investigated in this paper can be thought of as minimal models for learning in hippocampus and neocortex. The SP model bears some resemblance to the function of hippocampus, which is supposed to keep a memory of recent episodes that are learned in one shot, thanks to highly plastic synapses. The MP model relates to the function of neocortex, where a longer-term memory can be stored, thanks to repeated presentations of a set of prototypes that occur repeatedly in the environment, and perhaps during sleep under the supervision of the hippocampus. The idea that hippocampal and cortical networks learn on different time scales has been exploited in several modeling studies [43], [44], [45], in which the memories are first stored in the hippocampus and then gradually transferred to cortical networks. It would be interesting to extend the type of analysis presented here to coupled hippocampo-cortical networks with varying degrees of plasticity.
We are interested at retrieving pattern that has been presented during the learning phase. We set the network in this state and ask whether the network remains in this state while the dynamics (2) is running. At the first iteration, each neuron is receiving a field(35)
Where M+1 is the number of selective neurons in pattern , with . Where we use the standard ‘Landau’ notations: means that goes to a finite limit in the large limit, while means that goes to zero in the large limit. and . We recall that and . Thus is a binary random variable which is with probability, either if is a selective neuron (sites such that ), or if is a non-selective neuron (sites such that ). Neglecting correlations between and (it is legitimate in the sparse coding limit we are interested in, see [17]), the 's are independent and the distribution of the field on selective neurons can be written as(36)where we used Stirling formula for , with defined in (13). For non-selective neurons(37)
Now write(38)
In the limit we are considering in this section, and if , the sums corresponding to the probabilities are dominated by their first term (corrections are made explicit in the following section). Keeping only higher order terms in in Eqs. (36) and (37), we have:(39)and(40)
yielding Eq. (15) with . Note that with the coding levels we are considering here (), is of order . When the number of selective neurons per pattern is fixed at , we choose for the activation threshold and these equations become:(41)where
For random numbers of selective neurons we need to compute the average over : . Since is distributed according to a binomial of average and variance , for sufficiently large , this can be approximated as where is normally distributed:(42)with(43)and(44)
When goes to infinity, we bring the limit into the integral in Eq. (42) and obtain(45)where is the Heaviside function. Thus in the limit of infinite size networks, the probability of no error is a step function. The first Heaviside function implies that the only requirement to avoid errors on selective neurons is to have a scaled activation threshold below . The second Heaviside function implies that, depending on , has to be chosen far enough from . The above equation allows to derive the inequalities (19) and (20).
We now turn to a derivation of finite-size corrections for the capacity. Here we show two different calculations. In the first calculation, we derive Eq. (32), taking into account the leading-order correction term in Eq. (43). This allows us to compute the leading-order correction to the number of patterns that can be stored for a given set of parameters. However, it does not predict accurately the storage capacity of the large-size but finite networks that we simulated. In the second calculation presented, we focus on computing the probability of no error in a given pattern , including a next-to-leading-order correction.
Eq. (32) is derived for a fixed set of parameters, assuming that the set of active neurons have a fixed size, and that the activation threshold has been chosen large enough such that the probability to have non-selective neurons activated is small. From the Stirling expansion, adding the first finite-size correction term in Eq. (41), we get(46)with . For large , the number of stored patterns can be increased until . Setting , an expansion of in allows to write(47)
The patterns are correctly stored as long as . This condition is satisfied for . For the SP model, we can deduce which value of yields this value of (see Eq. (26)). This allows to derive Eq. (32),(48)
We now turn to a calculation of the probability of no error on a given pattern , taking into account the next-to-leading order correction of order one, in addition to the term of order in Eq. (41). This is necessary to predict accurately the capacity of realistic size networks (for instance for , ). is computed for a memory pattern with selective neurons. The estimation of used in the figures is obtained by averaging over different values of , with drawn from a binomial distribution of mean .
We first provide a more detailed expansion of the sums in Eq. (38). Setting , with the Taylor expansions:(49)(50)where and . Using (37) we can rewrite:(51)
In the cases we consider, we will always have so that we can consider only the term of order in . The sum is now geometric, and we obtain(52)
The same kind of expansion can be applied for the selective neurons. Again if we are in a situation where (53)
When is close to and thus , we are then left with:(54)(55)
When is too close to , which is the case for the optimal parameters in the large limit, we need to use (55). It only contributes a term of order in and does not modify our results. In Figures 6-7, we use (53), which gives from (38) and (36), (37) and (53),(52):(56)(57)
The probability of no error is(58)which leads to Eqs. (31)
For a fixed number of selective neurons in pattern , approximating the distribution of the fields on background neurons and selective neurons with a gaussian distribution gives:(59)where(60)and(61)where(62)
The probability that those fields are on the wrong side of the threshold are:(63)and(64)
Following the same calculations presented, and keeping only terms that are relevant in the limit , the probability that there is no error is given by:(65)where the rate function is(66)
Calculations with the binomial versus the gaussian approximation differ only in the form of . Finite size terms can be taken into account in the same way it is done in the previous Methods section for the binomial approximation.
In all above calculations we assumed that fields are sums of independent random variables (35). For small correlations are negligible [17], [19]. It is possible to compute the covariances between the terms of the sum (see Eq. (3.9) in [19]), and take them into account in the gaussian approximation. This can be done using(67)(68)in Eqs. (59),(61), where(69)
|
10.1371/journal.pntd.0000365 | Molecular and Behavioral Differentiation among Brazilian Populations of Lutzomyia longipalpis (Diptera: Psychodidae: Phlebotominae) | Lutzomyia longipalpis is the primary vector of American visceral leishmaniasis. There is strong evidence that L. longipalpis is a species complex, but until recently the existence of sibling species among Brazilian populations was considered a controversial issue. In addition, there is still no consensus regarding the number of species occurring in this complex.
Using period, a gene that controls circadian rhythms and affects interpulse interval periodicity of the male courtship songs in Drosophila melanogaster and close relatives, we analyzed the molecular polymorphism in a number of L. longipalpis samples from different regions in Brazil and compared the results with our previously published data using the same marker. We also studied the male copulation songs and pheromones from some of these populations. The results obtained so far suggest the existence of two main groups of populations in Brazil, one group representing a single species with males producing Burst-type copulation songs and cembrene-1 pheromones; and a second group that is more heterogeneous and probably represents a number of incipient species producing different combinations of Pulse-type songs and pheromones.
Our results reveal a high level of complexity in the divergence and gene-flow among Brazilian populations of the L. longipalpis species complex. This raises important questions concerning the epidemiological consequences of this incipient speciation process.
| Lutzomyia longipalpis is the main vector of visceral leishmaniasis in the Americas. There is strong evidence that L. longipalpis is a species complex, but there is still no consensus regarding the number of species occurring in Brazil. We combined molecular and behavioral analyses of a number of L. longipalpis populations in order to help clarify this question. This approach has allowed us to identify two main groups of populations in Brazil. One group probably represents a single species distributed mainly throughout the coastal regions of North and Northeast Brazil and whose males produce the same type of copulation song and pheromone. The second group is more heterogeneous, probably represented by a number of incipient species with different levels of genetic divergence among the siblings that produce different combinations of copulation songs and pheromones. The high level of complexity observed raises important questions concerning the epidemiological consequences of this incipient speciation process.
| Cryptic speciation is an interesting and important issue to evolutionary biologists as organisms that are distinct in several ways can look the same even to specialist taxonomists, leading to false conclusions about their biology. Moreover, it also has practical implications for conservation management and in the identification of economically or medically important species [1]. In addition, the study of cryptic speciation in blood-sucking insects can be epidemiologically relevant as sibling species might differ in their importance as disease vectors. One classical example is the Anopheles gambiae complex whose siblings differ in their host preference and other biological characteristics that together define very different vectorial capacities (reviewed in [2]).
Lutzomyia longipalpis (Lutz & Neiva 1912) is the primary vector of American visceral leishmaniasis in the Neotropics. This sand fly has an extensive and discontinuous distribution from southern Mexico to northern Argentina, and is found in a range of different habitats [3]–[5]. Geographic isolation between numerous populations of L. longipalpis favors the process of genetic divergence and cryptic speciation [6]–[12].
L. longipalpis males show a polymorphism in the number of abdominal spots, with either a single pair on the fourth tergite (1S) or two pairs, on the third and fourth abdominal segments (2S) [13]. Experimental evidence for a L. longipalpis complex was first obtained by crosses between Brazilian populations with different abdominal spot morphology [14],[15]. It was shown that the spot phenotype cannot be used as a reliable marker to identify different species because some crosses between populations with different male phenotypes did not show reproductive isolation, while some crosses between populations with the same male phenotypes had reduced insemination rates. However, in at least one specific locality (Sobral, Ceara State), males of the two forms found in sympatry represented different species that showed strong reproductive isolation [14],[15] suggesting that this morphological variation might still be a useful tool to distinguish between sympatric siblings in some situations.
After this early work many studies have been carried out on populations from different Latin American countries to determine the taxonomic status of this important Leishmania vector, indicating that L. longipalpis is a species complex. However there is no consensus on the delineation between members of the complex as different genetic markers suggest different conclusions (reviewed in [10]).
The fact that some genetic markers show clear evidence for a complex while others do not, strongly suggests recent or incipient speciation events that are perhaps also associated with introgression among the siblings. Therefore the data for the L. longipalpis populations in Brazil and in Latin America as a whole suggest a level of complexity that resembles what is found in the Anopheles gambiae complex and An. gambiae s.s. [2],[10]. In addition, it is still not clear how many species or incipient species of the L. longipalpis complex exist within Brazil. This is not only interesting from an evolutionary point of view but may also be epidemiologically relevant since most cases of visceral leishmaniasis in Latin America occur in the northeast region of Brazil. The characterization of the number, distribution and genetic divergence of the different species of the L. longipalpis complex might allow the future study of possible coevolutionary interactions between the different siblings and parasite genotypes that might influence Leishmania transmission, virulence and clinical outcome [16], as has been recently suggested for malaria [17].
Molecular markers associated with sexual behavior and reproductive isolation, such as the period (per) gene [18], are useful tools for speciation studies. In Drosophila, per is one of the genes that control circadian rhythms of emergence from pupae and locomotor activity in adult flies [19]. This clock gene also controls species-specific differences in the courtship song rhythms that are involved in reproductive isolation between Drosophila melanogaster and its sibling species [18], [20]–[22] and for that reason it was considered an example of “speciation gene” [23]. In addition per is usually highly variable and was widely used for population genetics studies in Drosophila [24].
The per gene was also shown to be a good molecular marker for studying the differentiation among populations of the L. longipalpis complex. Bauzer et al. [25] used the per gene to study three allopatric populations from Brazil (Natal, Lapinha and Jacobina) and showed that they were genetically highly differentiated. Later, using the same marker, Bauzer et al. [26] provided the first molecular evidence in Brazil (Sobral) for the existence of two genetically highly differentiated sympatric populations, confirming therefore the early results of Ward et al. [14],[15]. These findings were later corroborated by analysis of microsatellites [27],[28] and genetic differentiation in the cacophony (cac) gene [29].
Acoustic signals are an important component of courtship behavior in insects [30] and they show rapid evolution in some Drosophila sibling species (e.g. [31]). As in Drosophila, acoustic communication is important in L. longipalpis sexual behavior [32],[33]. However, unlike D. melanogaster [34], but similar to some other Drosophila species [35], males of L. longipalpis vibrate their wings during copulation. Two main types of copulatory courtship songs, named Pulse-type and Burst-type, were found in six Brazilian populations [32]. The Burst-type song (B), composed of trains of extremely polycyclic pulses (“bursts”) modulated in frequency and amplitude, showed no significant variation among the three populations producing this pattern (Natal, Sobral 2S and Marajó) [32]. On the other hand the Pulse-type song is more variable and three different patterns (P1, P2 and P3) were identified in the populations of Jacobina, Lapinha and Sobral 1S, respectively [32]. The P1 songs are composed of trains of pulses with usually two or three cycles per pulse. P2 songs differ from P1 mainly by the presence of interspersed polycyclic pulses between nearly monocyclic pulses. The P3 songs are characterized by an almost perfect alternation of high and low amplitude pulses [32]. These data suggest the existence of four different siblings among these six populations [10],[16].
Males of the L. longipalpis complex also produce different types of terpenoid volatile compounds with either 16 or 20 carbons that function as sex pheromones [36]–[41]. Interestingly, the Pulse-type populations mentioned above produce different types of C16 pheromones while Burst-type populations produce the same type of C20 [32]. The two groups also show reproductive isolation in crossing experiments [14],[15],[42] and a clear divergence with microsatellites, cac and per gene sequences [25]–[29],[32].
Integrative approaches are particularly useful to study cryptic speciation as not all characters are necessarily going to show the same level of differentiation and studying multiple, different types of characters will more fully describe any differentiation [1]. In the present study, we carried out a combined analysis of per gene molecular polymorphisms, copulation songs and pheromones from a number of Brazilian L. longipalpis populations comparing new data we obtained to previously published results. One of our underlying hypotheses was that more species of the complex must exist in Brazil since previous molecular and behavioral analysis of the six populations described above revealed the existence of four siblings. A second hypothesis was that other sympatric siblings would probably be found in localities where L. longipalpis males with one and two spots coexist as observed in Sobral. Finally, because copulation songs and pheromones are likely to evolve under sexual selection, our third underlying hypothesis was that the observed correlation between genetic divergence in molecular markers and these two traits would not necessarily hold once a large number of populations was analyzed.
We analyzed Lutzomyia longipalpis samples from eight populations (see below) collected in six Brazilian localities: Barra de Guaratiba (23°04′S, 43°35′W), Rio de Janeiro State; Estrela de Alagoas (09°23′S, 36°45′W), Alagoas State; Jaíba (15°20′S, 43°41′W), Minas Gerais State; Mesquita (22°46′S, 43°25′W), Rio de Janeiro State; Pancas (19°13′S, 40°51′W), Espírito Santo State; and Teresina (05°05′S, 42°48′W), Piauí State. A map with the approximate position of these localities is shown in Figure 1. Species identification was carried out according to Young and Duncan [3]. In most cases the specimens analyzed were either wild-caught males or the F1 males of different wild-caught females. However, in the case of Mesquita, due to the difficulties explained below, the sequences were obtained from the F1 of a single wild-caught female. In the samples of Estrela de Alagoas and Jaíba, where males present either one (1S) or two spots (2S), the samples of the two morphotypes were considered potential sympatric siblings and therefore were analyzed separately. In Teresina nearly 95% of the flies were 1S and only those were used in the analysis. The remaining analyzed samples were also all 1S.
Genomic DNA extraction was carried out as in Jowett [43]. The PCR was done as described in Bauzer et al. [25]. The PCR fragments were purified using the Wizard PCR Prep kit (Promega) and cloned using the pMOS Blunt Ended cloning kit (GE Healthcare) or the pGEM T Easy Vector System I (Promega). The plasmid DNA was prepared with the Flexiprep kit (GE Healthcare) or using alkaline lyses method in micro-plates of 96 wells [44] and filtered in Multiscreen-filter plates. Sequencing of cloned fragments was carried out using the ABI Prism Big Dye Terminator Cycle Sequencing Ready Reaction V3.0 or V3.1 kits (Applied Biosystems) in ABI Prism 377 and ABI 3730 DNA Sequencers at Fundação Oswaldo Cruz, Rio de Janeiro, Brazil.
From three to eight sequences were obtained for each individual corresponding to the eight populations. In most cases the sequences were aligned generating a single consensus sequence per individual. For individuals with 6 or more sequences available, two consensus sequences representing both alleles were generated (denominated haplotypes A and B). Analysis of the data including and excluding these extra sequences showed no significant differences and therefore they were also used. The sequences have been submitted to GenBank (accession numbers EU713077 to EU713233).
The DNA sequences were edited and aligned using the Wisconsin Package Version 9.1; Genetic Computer Group (GCG) (Madison, Wisconsin, USA). The polymorphism analyses and three neutrality tests were performed using DnaSP 3.5 [45] and differentiation between populations was estimated using ProSeq program [46]. The genealogy of the sequences was carried out using MEGA 3.1 [47]. The analysis of molecular variance (AMOVA) was carried out using Arlequin 3.1 [48]. For these analyses the 113 sequences obtained by Bauzer et al. [25],[26] from the populations of Jacobina, Lapinha, Natal and Sobral (1S and 2S), and the five sequences from Salvaterra (Marajó Island, Pará State) from Souza et al. [32] were considered as well.
The haplotype networks were constructed using either the whole fragment of 266-bp or a 58-bp non-recombinant block selected after recombination analysis [49] available in DnaSP 3.5 and exclusion of a few recombinant sequences. The networks were estimated using statistical parsimony [50] implemented in the TCS1.21 software [51].
We recorded and analyzed male copulation songs from seven samples: Estrela de Alagoas (one and two spot males); Jaíba (one and two spot males), Pancas, Teresina and Mesquita. We attempted to obtain new samples of live males from Barra de Guaratiba, to record their songs but this population was found to be very small (see below) and ephemeral and consequently we were not successful. L. longipalpis is very rare in Mesquita, which is dominated by L. intermedia (Nataly A. Souza, unpublished results) and the collection area is surrounded by a “Favela” (shanty town) that is currently inaccessible to new collections. Therefore only the progeny of a single female was analyzed and no material was available for pheromone analysis (see below). In addition, we analyzed the songs of four different Mesquita copulations but three of the four songs were produced by the same male copulating with three different females. The three songs produced by this male were included in the analysis because the variation in the song parameters among these three copulations was higher than the variation observed between the songs of the two different males (data not shown).
Songs were recorded as in Souza et al. [32]. A male and a female were placed in a small mating chamber inside the INSECTAVOX [52] for about five minutes. If there was no copulation after 5 minutes the couple was replaced. Usually only one copulation is observed in every 15 to 20 trials. The recordings were carried out at 26°C±1°C using a Sony Hi8 CCD-TRV65 videocamera and a Sony SLV-77HFBR VCR. Some individuals from Jaíba, both 1S and 2S, and from Estrela de Alagoas 1S were recorded using a DVD recorder, Panasonic DMR-ES10. Most recordings were carried out using F1 virgin individuals (between 3 to 8 days old) reared in the lab from wild-caught females. A number of the males that had their songs recorded were also used in the molecular analysis.
The songs were digitalized on a CED 1401 A/D converter and analyzed using Spike 2 software (version 4.08), both from Cambridge Electronic Design (UK). The song parameters analyzed were: interpulse (IPI) or interburst interval (IBI), number of pulses or bursts per train (NT), train length (TL), cycles per pulse (CPP), and carrier frequency of train (Freq). In the Jaíba 1S and Teresina populations most parameters were similar (see below) but the songs of the latter tended to have pulses that alternated in amplitude, similar to what was previously seen in Sobral 1S [32]. Therefore, for these two populations we performed an analysis of the pulse amplitude alternation (AmpAlt) looking at the proportion of pulses that have amplitudes either lower or higher than the flanking pulses. The statistical analysis was carried out using STATISTICA (version 5, StatSoft, Inc., USA).
Individual males were placed in flame-cleaned Pasteur pipette ampoules and sufficient pesticide-grade hexane (∼10 ul) (BDH, Poole, Dorset, U.K) was added to cover them and extract the sex pheromone. The ampoules were heat-sealed and stored at −20°C until analysis [38].
Hexane extracts containing the sex pheromones were analyzed by Coupled Gas Chromatography - Mass Spectrometry on a HP-5MS capillary column, 30 m×0.25 mm i.d., 0.25 µm film thickness (Agilent, Stockport, Cheshire) in a Hewlett Packard 5890 II+ Gas Chromatograph coupled to a Hewlett Packard 5972A bench top mass spectrometer (electron impact, 70 eV, 180°C). Injection of the sample was via a heated injector (170°C). The carrier gas was helium at 1 ml/min. The GC was temperature-programmed with an initial 2 min at 40°C, then a rise of 10°C/min to a final isothermal period at 250°C (4 min).
We analyzed the pheromones of samples from Barra de Guaratiba, Teresina and Pancas. We were not able to obtain a new sample from Mesquita, due to the reasons stated above. Information on the pheromones of the other populations included in this work is already published [33], [36]–[38],[53].
Our previous work [32] has shown that the copulatory courtship songs produced by L. longipalpis males have two main components: a primary song produced by all males during copulation that shows remarkable differences among some populations, and a secondary song that is composed of low amplitude polycyclic pulses with highly variable intervals. The secondary song is not produced by every male and when it occurs it is usually seen between two primary song trains [32]. In the present study we also observed that the low amplitude secondary song seems to be similar in all populations and it was not analyzed in detail. However, analysis of the primary song indicates striking differences among the studied samples (song traces are shown in Supplemental Figure S1).
Males from Pancas and two spot males (2S) from Estrela de Alagoas and Jaíba produce the same type of Burst song previously observed in Natal, Marajó and Sobral 2S males [32] (see Figure 1 and Figure S1). In the sample from Teresina we observed the same Pulse-type song (P3) previously recorded in the population of Sobral 1S [32] (Figure 1). In addition, two new types of Pulse songs were observed in Jaíba 1S and Estrela 1S males that were called P4 and P5, respectively (Figure 1 and Figure S1) (see below for a description of these new Pulse songs). Note that 1S and 2S males from Jaíba and Estrela produce very distinct songs indicating that in both localities the spot phenotype might be associated with different sympatric species as observed in Sobral, and that is consistent with the molecular data (see below). Finally, males from Mesquita produce a new song type that seems to present a mixed (M) pattern between Burst and Pulse songs (Figure 1 and Figure S1).
Table 1 shows the mean (±SEM) for the different song parameters. Because of the very different patterns observed in Burst and Pulse-type songs their statistical analyses were carried out separately comparing the data to our previously published results for other populations [32].
Analysis of variance (ANOVA) comparing the Burst-type songs of the populations of Pancas, Estrela 2S, Jaíba 2S (this study), Natal, Marajó and Sobral 2S [32] detected a significant difference only for IBI (F[5,38] = 3.89; p<0.01). We performed a Post-Hoc LSD (“Least Significant Difference”) test for this parameter and found that the difference occurs mainly because of the high IBI found in the population of Jaíba 2S (data not shown).
An ANOVA was also carried out comparing the populations that produce Pulse-type songs: Teresina, Jaíba 1S and Estrela de Alagoas 1S (this study), Jacobina, Lapinha and Sobral 1S [32]. The results showed significant differences in all parameters analyzed (IPI: F[5,45] = 43.76; NP: F[5,45] = 49.18; TL: F[5,45] = 12.97; Freq: F[5,45] = 25.18; CPP: F[5,45] = 23.09; AmpAlt: F[2,17] = 56.42; p<0.0001 in all cases).
As mentioned above males from Teresina sing like Sobral 1S males. They produce a P3 type song which shows an alternation of high and low amplitude pulses along the train with a ∼65 ms mean inter-pulse interval and about 2 s train length (Table 1). Post-Hoc LSD tests show no significant differences between Teresina and Sobral 1S males in any of the parameters analyzed (data not shown).
The song produced by Jaíba 1S males, named P4 (Figure 1 and Figure S1), has similar mean parameter values to the songs of Teresina and Sobral 1S (Table 1; [32]). However, the pattern is different and the P4 song does not show the alternating high and low amplitude pulses observed in the P3 song (Figure 1 and Figure S1). In P4 songs, the pulse amplitude grows and decreases gradually through out the train. Post-Hoc LSD analysis between Jaíba 1S and the populations of Sobral 1S and Teresina showed a significant difference only in the proportion of alternated amplitude pulses (AmpAlt) (data not shown).
Estrela de Alagoas 1S males produce the type 5 Pulse song (P5). This pattern presents similar pulses along the train with a few polycyclic pulses in the end (Figure 1 and Figure S1). This song has a very short IPI (∼37 ms) and a long train (∼3.6 s) compared with the other populations of L. longipalpis (Table 1) [32]. Post-Hoc LSD analysis shows significant differences in all parameters in most comparisons (data not shown).
Mesquita males produce a new song pattern that we called Mix (M). The M song type is characterized by the presence of polycyclic pulses in the first third of the train followed by highly polycyclic pulses resembling bursts at the end (Figure 1 and Figure S1). Table 1 shows the means (±SEM) of analyzed parameters for the whole song and for the more Pulse-like and Burst-like segments. These values show a clear separation of the two parts that constitute this new song pattern. ANOVA and LSD analysis were carried out comparing the two segments, the first segment to the populations with Pulse song and the second segment to the populations with Burst song. Significant differences were obtained in most comparisons (Table S1).
Finally, Table 2 shows the results of the Gas Chromatography and Mass Spectrometry analysis. In Pancas, males produce only diterpenes (20 carbon terpenes with a molecular weight of 272 amu) which can be characterized as cembrenes. The main compound found in males from this locality (Rt = 22.95) has a mass spectrum similar to the main cembrene-1 present in Sobral 2S males and some other populations [38]. The main compound found in the male pheromones of Barra de Guaratiba and Teresina is (S)-9-methylgermacrene-B (9MGB), a 16 carbon terpene with a molecular weight of 218 amu. However, males from these two localities, particularly Barra de Guaratiba, also produce several diterpenes.
Figure 1 shows a map with the approximate geographic position of the different localities with a summary of the information available for the different copulation songs and main pheromone types from 14 different populations based on this study and previously published data [32], [33], [36]–[38],[53]. This information will be compared to the results of the period gene molecular analysis that follows.
We amplified the same 266-bp fragment of the L. longipalpis per gene used by Bauzer et al. [25],[26] which includes a 54-bp intron. This fragment encodes part of the region between the PAS dimerization domain and the Thr-Gly region, including most of the cytoplasmic localization domain (CLD). We analyzed 157 sequences of the L. longipalpis per gene obtained from six localities (Barra de Guaratiba, Estrela de Alagoas, Jaíba, Mesquita, Pancas and Teresina). The sequences from Estrela de Alagoas and Jaíba were separated according to the male spot phenotypes (1S and 2S). The sequences from Marajó were included in this analysis as they were not analyzed in detail in Souza et al. [32]. For the 266 sites analyzed, 56 (21.05%) were variable. Most of the base substitutions were silent or occurred in the intron (see alignment in Figure S2). We found five non-silent substitutions occurring in four different sequences from Estrela de Alagoas 1S, all in the second exon. We also observed deletions which included either one or two sites in the intron region of two sequences from Estrela de Alagoas 2S and 18 sequences from Estrela de Alagoas 1S. The remaining sequences did not show any deletions or amino acid changes (Figure S2).
The number of segregating sites (S), nucleotide diversity (π) and the neutral parameter (θ) were calculated for each population (Table 3). Interestingly, the Barra de Guaratiba population did not present polymorphism in any of the 24 sequences analyzed. The samples from Estrela de Alagoas 1S and Jaíba 1S were the most polymorphic of the L. longipalpis samples analyzed. We carried out three tests of selective neutrality, Tajima's D [54], Fu's FS [55] and Ramos-Onsins and Rozas' R2 [56]. In all cases the values were not significant after Bonferroni's correction (Table 3), indicating no obvious departures from neutrality in L. longipalpis per gene.
For the analysis of molecular differentiation all pair-wise comparisons were performed including the five populations obtained by Bauzer et al. [25]–[26] plus the samples of L. longipalpis analyzed in this study, excluding the monomorphic Barra de Guaratiba population and the Mesquita sample that was derived from the F1 of a single female. Table 4 shows the pair-wise fixation index (FST) [57], and Tables S2 and S3 show the number of shared/fixed polymorphic sites and the number of exclusive polymorphic sites, respectively. The results of the FST analysis clearly indicate that the 1S and 2S males from Estrela de Alagoas (AL) and Jaíba (MG) belong to different sympatric species. Therefore we started analyzing in more detail these two localities before carrying out a general analysis of all populations.
A highly significant genetic differentiation was found between Estrela 1S and 2S, a result that is similar to that observed between Sobral 1S and 2S [26]. Although no fixed differences in per were observed between Estrela 1S and 2S, perhaps due to introgression (see below), the number of exclusive and shared polymorphisms is similar (20 and 19, respectively) (Tables S2 and S3). In addition, the results presented in Table 4 suggest that Estrela 2S is essentially the same gene pool as Jaíba 2S, Natal, Marajó, Sobral 2S and Pancas, with all FST values low and/or non-significant. This is consistent with the song analysis presented above. On the other hand all comparisons involving samples from Estrela 1S showed high levels of differentiation. The lowest FST value for Estrela 1S was observed in the comparison with Sobral 1S.
Figure 2 shows a haplotype network of all sequences from Estrela de Alagoas (1S and 2S) (see also Figure S3 for a minimum evolution tree of the same sequences). The haplotypes are clearly separated into two major groups. The first one consists mainly of Estrela 1S sequences, plus three haplotypes of Estrela 2S. The second group consists of Estrela 2S sequences, plus three haplotypes of Estrela 1S. Note in Figure 2 the few available sequences from Estrela males that have had also their songs analyzed. Haplotypes from males producing Burst song (marked with a “B”) were all Estrela 2S while those from males producing Pulse P5 songs (marked with a “P5”) were Estrela 1S. However, among the other sequences two interesting features are also observed in the network. Estrela 1S sequence Est1S13A clusters with Estrela 2S sequences while the other allele from the same fly (Est1S13B) clusters with other Estrela 1S sequences. The same happens to Estrela 1S sequences Est1S14A and Est1S14B; and Estrela 2S sequences Est2S24A and Est2S24B. This could indicate retention of ancestral polymorphisms, introgression or even possible hybrids between the two siblings. In addition, although the sequences Est2S30A and Est2S30B are the two alleles of one Estrela 2S fly both cluster with Estrela 1S sequences. This could indicate that the spot phenotype is not 100% reliable for identifying the two putative species in this locality.
Our data show that Jaíba 1S and 2S morphotypes are also highly differentiated, as observed in Estrela de Alagoas and Sobral (Table 4, [26]). One fixed difference, 29 exclusive polymorphisms and only four shared polymorphic sites were observed between Jaíba 1S and 2S (Tables S2 and S3). The differentiation observed between the two Jaíba putative siblings is consistent with the song analysis (see above). Jaíba 2S shows very low and non-significant FST values when compared to Natal, Sobral 2S, Estrela 2S, Marajó and Pancas suggesting that all these populations represent essentially the same species (Table 4). Comparisons involving Jaíba 1S show a low and non-significant level of differentiation with Sobral 1S and Lapinha (Table 4). A haplotype network including all sequences from Jaíba (1S and 2S) is shown in Figure 3 (see also Figure S4 for a minimum evolution tree of the same sequences). The results are similar to Figure 2 and the network clearly separates Jaíba 1S and 2S sequences. Note also that the haplotypes from Jaíba 1S and 2S males that have had their songs analyzed (marked with “P4” and “B”, respectively) cluster consistently with the other sequences of their respective populations.
Table 4 shows that the lowest genetic differentiation involving Teresina was observed against Sobral 1S. On the other hand, the lowest and non-significant FST value observed for Pancas was against Sobral 2S. Pancas also showed low and non-significant values against Natal and Jaíba 2S, low but significant differentiation against Estrela 2S and moderate, but non-significant differentiation against Marajó. The latter is more closely related to other Burst song populations and showed higher levels of genetic differentiation when compared to Pulse song samples but all comparisons resulted in non-significant values, probably due to the small number of sequences available for this locality (Tables 3 and 4).
Figure 4 shows two haplotype networks based on a 58-bp non-recombinant segment within the 266-bp fragment. An attempt to construct a network based on the whole fragment resulted in too many ambiguities due to many recombination events (see also Bauzer et al. [25]) and therefore this smaller non-recombinant segment was used. Twenty-three haplotypes were identified with 18 segregating sites. Using a 95% connection limit, two networks were constructed. In network 1 (Figure 4 and Table S4) we found the two more frequent haplotypes (H2 and H3) connected to H1. The H2 is the predominant haplotype in Burst song populations of Natal, Sobral 2S, Estrela 2S, Pancas, Jaíba 2S and Marajó. On the other hand H3 is the predominant haplotype in the Pulse song populations of Lapinha, Sobral 1S, Jaíba 1S, and Teresina. Mesquita and Barra de Guaratiba show only the haplotype H3. The main haplotype found in Jacobina (H4) is connected to H3 by a single mutation. In network 2 we observed one major haplotype with 16 sequences of the Estrela 1S, and other two minor haplotypes with sequences of Estrela 1S and Estrela 2S (Figure 4 and Table S4). The two networks are connected if the significance level is decreased to 94%. In this case, the Estrela haplotype H21 is connected to H09 by two hypothetical haplotypes.
We examined the partitioning of per sequence variation within and between groups by performing an AMOVA. For this analysis, groups were defined based on copulation songs and pheromones (Table 5), the samples from Barra de Guaratiba and Mesquita were not included. Most of the total variation (over 50%) is distributed within populations in both analyses. For the pheromone groups based on the number of carbons (C16 and C20) or based on the different pheromone types (cembrene-1, cembrene-2, 9-methylgermacrene-B and himachalene), the variation was almost equally distributed between populations within the groups (18.73 and 22.90%, respectively) and among groups (23.25 and 18.37%, respectively). This result probably reflects the fact that Jaíba 1S and Jaíba 2S are both C20 but genetically quite different, and that Estrela 1S is cembrene-1 (and therefore C20) but genetically differentiated from the rest of these groups.
Considering the two main groups of copulation song patterns (Burst and Pulse), 13.79% of the total molecular variation is distributed among populations within groups. This variation is much smaller when we consider the different Pulse-type songs (P1, P2, P3, P4 and P5) as separate groups (1.68%) and reflects the high similarity among populations with Burst songs and between the two P3 populations (Sobral 1S and Teresina).
The low differentiation among Burst song populations contrasts to the differentiation among Pulse song populations (Table 4). Hence, the mean pairwise FST value among the Burst populations is only 0.058±0.017 while among samples of Pulse song populations it is 0.260±0.037. Finally, the differentiation among Pulse populations is nearly half of that observed for comparisons between the two song groups (mean pairwise FST = 0.449±0.011).
There are a number of difficulties associated with the study of recently diverged species and populations in an incipient speciation process [58]–[60] such as the members of the L. longipalpis complex within Brazil [10]. Although the role of sexual selection as a major cause of speciation still needs further support [61], it is likely that the rapid divergence of mating signals is particularly important in the evolution of reproductive isolation in cryptic species complexes [62]. Therefore, in order to enhance our knowledge on the taxonomic status and geographic distribution of the different L. longipalpis siblings in Brazil, we combined a comparison of period gene sequences with an analysis of male copulation songs and sex pheromones, traits that probably have an important role in the reproductive isolation among these closely related species.
Our results reveal a high level of complexity in the divergence and gene-flow among Brazilian populations of the L. longipalpis species complex. The available data suggest that the sibling species producing Burst-type copulation songs and cembrene-1 is distributed mainly throughout the coastal regions of North and Northeast Brazil reaching the Southeast in Pancas (Figure 1). In contrast, the populations producing different types of Pulse songs and pheromones are far more heterogeneous and probably represent five incipient species with different levels of divergence among the siblings.
The existence of pairs of sympatric species in three different localities (Sobral, Jaíba and Estrela de Alagoas) raises the question whether reinforcement of reproductive isolation [63] is occurring in this species complex as preliminary evidence suggests [42]. In each case a Burst song, cembrene-1 population is sympatric to a different Pulse song sibling. In Sobral, the two siblings also differ in their pheromone types, and their genetic differentiation and reproductive isolation have been confirmed by microsatellites [28], the cacophony gene [29] and crossing experiments [14],[15],[42]. In Jaíba, the two siblings also differ in the type of diterpene isomers they carry [53]. Finally, Estrela 1S and 2S males differ only in their copulation songs, but they share the same type of pheromone [38]. Interestingly though, Estrela 1S is the most genetically divergent among the Pulse-type populations. Acoustic signals and pheromones are both likely to have a role in the reproductive isolation of the L. longipalpis siblings as observed in other insects [64]. The pheromones are probably involved in pre-mating isolation [42] while copulation songs might be the main signal involved in the insemination failure observed in copulations between the siblings [14],[15],[42]. The fact that in Estrela, males of the two siblings have the same pheromone suggests that in this locality copulation songs might have a more important role as isolation mechanism or, that other signals, such as, visual or putative cuticular pheromones are also involved [65]. Our results with Estrela also show how important is to carry out an integrative approach using molecular and behavioral data as the analysis of pheromone alone in this locality would suggest a single species. It is possible that future work will reveal siblings that differ in pheromones but not in their copulation songs.
Besides the Burst and Pulse songs, analysis of the acoustic signals produced by Mesquita males revealed a new pattern that we called Mix because it presents characteristics that resemble superficially the two other types. The trains of the Mix song begin like a polycyclic Pulse song and end with a more Burst-like pattern. Interestingly, the switch from one pattern to the other in the Mesquita song is also associated with characteristic changes in male behavior. Males that produce Burst song not only vibrate their wings but also swing their bodies continuously about 30° degrees to each side during copulation. Males producing Pulse songs do not do that. Mesquita males produce songs in both ways. They start the copulation like Pulse song males and then begin swinging their bodies the same way Burst song males do (Vigoder and Peixoto, unpublished observations). Unfortunately the Mesquita sample we analyzed was very small and new samples from the same collection site are impossible to obtain at the moment. We are currently attempting to obtain further samples from other localities in Rio de Janeiro State, a region where L. longipalpis is usually quite rare. Analysis of other populations producing this new song pattern might offer interesting clues about the evolution of the copulation songs and the speciation in the L. longipalpis complex.
Barra de Guaratiba is also an interesting population because although males from this locality produce mainly the 9-methylgermacrene pheromone, large amounts of cembrenes are also found in their extracts. We observed that Barra de Guaratiba is a monomorphic population for the 266 bp fragment of the per gene analyzed. This lack of molecular variation is consistent with a very small population size. However, a recent bottleneck or a selective sweep event in or near the per locus with fixation of a single haplotype are also possible alternative explanations. Analysis of a different gene is needed to settle the issue.
Our data show evidence for the persistence of ancestral polymorphisms and/or introgression, suggesting that the separation between the Brazilian siblings is probably recent and perhaps incomplete as suggested by crossing experiments [14],[15]. Differential introgression across the genome might be occurring in L. longipalpis causing a mosaic of genetic divergence as observed in other sand flies [66] and in the Anopheles gambiae complex and An. gambiae s.s. incipient species [67]–[71]. That is consistent with the fact that different conclusions concerning the taxonomic status of L. longipalpis in Brazil were reached by studies using different genetic markers (reviewed in [10]).
Introgression between closely related or incipient vector species can have very important epidemiological consequences allowing the spread of insecticide resistance genes and adaptive traits as well as facilitating the bridge between sylvatic and peri-urban cycles [66]–[68],[72],[73]. In addition, man made environmental changes might promote contact between different incipient species with incomplete reproductive isolation that will in turn exchange reservoirs of genetic variability that might promote adaptation to modified habitats. Whether this could be one of the explanations for the spread of visceral leishmaniasis observed in the last decades in Brazil [5], is still an open question. To address the issue of introgression between the L. longipalpis siblings in Brazil in more detail, we are currently carrying out a multilocus analysis of the sympatric siblings from Sobral that we hope will help understanding the mechanisms shaping the differentiation in this complex. Extending our studies to other populations will also give us a better view of the geographical distribution of the L. longipalpis sibling species in Brazil and their potential implication to Leishmania transmission.
|
10.1371/journal.pntd.0001523 | A Novel G Protein-Coupled Receptor of Schistosoma mansoni (SmGPR-3) Is Activated by Dopamine and Is Widely Expressed in the Nervous System | Schistosomes have a well developed nervous system that coordinates virtually every activity of the parasite and therefore is considered to be a promising target for chemotherapeutic intervention. Neurotransmitter receptors, in particular those involved in neuromuscular control, are proven drug targets in other helminths but very few of these receptors have been identified in schistosomes and little is known about their roles in the biology of the worm. Here we describe a novel Schistosoma mansoni G protein-coupled receptor (named SmGPR-3) that was cloned, expressed heterologously and shown to be activated by dopamine, a well established neurotransmitter of the schistosome nervous system. SmGPR-3 belongs to a new clade of “orphan” amine-like receptors that exist in schistosomes but not the mammalian host. Further analysis of the recombinant protein showed that SmGPR-3 can also be activated by other catecholamines, including the dopamine metabolite, epinine, and it has an unusual antagonist profile when compared to mammalian receptors. Confocal immunofluorescence experiments using a specific peptide antibody showed that SmGPR-3 is abundantly expressed in the nervous system of schistosomes, particularly in the main nerve cords and the peripheral innervation of the body wall muscles. In addition, we show that dopamine, epinine and other dopaminergic agents have strong effects on the motility of larval schistosomes in culture. Together, the results suggest that SmGPR-3 is an important neuronal receptor and is probably involved in the control of motor activity in schistosomes. We have conducted a first analysis of the structure of SmGPR-3 by means of homology modeling and virtual ligand-docking simulations. This investigation has identified potentially important differences between SmGPR-3 and host dopamine receptors that could be exploited to develop new, parasite-selective anti-schistosomal drugs.
| Bloodflukes of the genus Schistosoma are the causative agents of human schistosomiasis, a debilitating disease that afflicts over 200 million people worldwide. There is no vaccine for schistosomiasis and treatment relies heavily on a single drug, praziquantel. Recent reports of praziquantel resistance raise concerns about future control of the disease and show the importance of developing new anti-schistosomal drugs. The focus of this research is on the nervous system of the model fluke, Schistosoma mansoni. Many pesticides and antiparasitic drugs act by interacting with neuronal proteins and therefore the nervous system is a particularly attractive target for chemotherapeutic intervention. Here we describe a novel receptor of S. mansoni that is activated by dopamine, an important neurotransmitter of the schistosome nervous system. The study provides a first in-depth analysis of this receptor and suggests that it plays an important role in the control of muscle function and movement. We also show that the schistosome receptor is substantially different from dopamine receptors of the mammalian host, both in terms of structure and functional properties. We propose that this novel protein could be used to develop new, schistosome-specific drugs aimed at disrupting parasite motility within the host.
| The bloodfluke Schistosoma mansoni is one of three species of schistosomes that cause significant disease in humans. Approximately 200 million people are infected and another 600 million are at risk of infection. Over 90% of all human schistosomiasis is due to S. mansoni. This species exists in Africa, the Middle East, South America and the Caribbean, in regions where the intermediate snail host, Biomphalaria glabrata, is also present. There is no vaccine for schistosomiasis and the arsenal of drugs available for treatment is limited. Praziquantel is the drug of choice but concerns over praziquantel resistance [1]–[3] have renewed interest in the search for alternative drug therapies. The nervous system of helminth parasites is considered to be an excellent target for chemotherapeutic intervention. Most of the anthelmintics currently in use, including the mainstay of nematode control, ivermectin, act by interacting with neurotransmitter receptors and cause disruption of neuronal signalling [4]. Recent drug screens conducted on cultured S. mansoni suggest that biogenic amine (BA) neurotransmitters may be particularly suitable for development of anti-schistosomal drugs [5], [6]. Substances that normally disrupt BA neurotransmission, such as dopaminergic and serotonergic drugs were shown to halt larval development [5] and to produce aberrant motor phenotypes in culture [6]. The BA systems of schistosomes have not been widely investigated at the molecular level and not much is known about the receptors or other proteins involved. More information is needed to elucidate the mode of action of these neurotransmitters and to identify potential targets for drug discovery.
BAs constitute a group of structurally related amino acid derivatives that function broadly as neurotransmitters and modulators in a variety of organisms. Included in this group are catecholamines (dopamine, noradrenaline, adrenaline), serotonin (5-hydroxytryptamine: 5-HT), histamine and the invertebrate-specific amines, tyramine and octopamine. In flatworms, including S. mansoni, BAs play important roles in the control of muscle contraction and movement, activities that are crucial for survival of the parasite within the host [7]–[9]. The best characterized of these amines is serotonin, which is myoexcitatory in all the flatworm species studied to date. Serotonin is synthesized by the parasite [10], it is widely distributed in the nervous system and there is evidence for the existence of a serotonin transport system in S. mansoni [11], [12]. At least two putative serotonin receptors are encoded in the S. mansoni genome [13], though neither has yet been characterized at the protein level. Besides serotonin, flatworms have both dopamine and histamine within their nervous system [14]–[20]. Dopamine, in particular, has important neuromuscular activities, which can be either excitatory or inhibitory depending on the flatworm species. In S. mansoni, dopamine causes relaxation of the body wall muscles [21], possibly by activating a receptor that is associated with neuromuscular structures [22]. In addition to motor effects, BAs have been implicated in the regulation of metabolic activity in several flatworms [8] and recent evidence has shown that serotonin and dopamine are both involved in the transformation of S. mansoni miracidia to sporocyst stage [5], suggesting a probable role in parasite development.
BAs exert their effects by interacting with cell-surface receptors, the majority of which belong to the superfamily of G protein-coupled receptors (GPCR) and are structurally related to rhodopsin. GPCRs have a distinctive topology consisting of seven transmembrane (TM) domains separated by loops, the longest of which is the third intracellular loop (il3). Rhodopsin-like (or Class A) GPCRs are further identified by having a relatively short extracelullar N-terminus, which is typically glycosylated, and an intracellular C-terminal tail of variable length [23]. In mammals, BA receptors are classified according to their amine specificity, sequence homology, signalling mechanisms and pharmacological profiles. Each BA interacts with multiple receptors. Dopamine, in particular, interacts with five different receptors (D1–D5), which are classified according to two major structural types, D1- and D2-like [24]. The current annotation of the S. mansoni genome has a total of 16 predicted BA receptors, all Class A GPCRs [13]. A few of these sequences, for example the D2-like dopamine receptor of S. mansoni (SmD2) [22] share sufficient homology with mammalian prototypes to be classified accordingly. The majority, however, are novel sequences that share about the same level of homology with all different types of BA receptors and can only be defined as BA-like. Among these sequences is a new clade of BA-like GPCRs (named SmGPR) that were previously described in our laboratory [15] and thus far have been detected only in schistosomes. These receptors could prove to be particularly good candidates for selective drug targeting and deserve further investigation.
Two SmGPRs of S. mansoni were shown earlier to be functional histamine receptors [15], [25]–[27]. Here we describe a third structurally related receptor (SmGPR-3), which has different agonist specificity. The results presented here show that SmGPR-3 is activated by dopamine and other catecholamines but does not resemble any one particular type of host dopamine receptor, either in terms of overall sequence homology, pharmacological profile or the predicted organization of the binding pocket. Further analysis of the receptor's tissue distribution revealed exceptionally abundant expression throughout the nervous system and suggests an important role for this novel dopamine receptor in neuronal and neuromuscular signalling.
A Puerto Rican strain of S. mansoni -infected Biomphalaria glabrata snails were kindly provided by Dr. Fred Lewis, Biomedical Research Institute, Rockville, Maryland, USA. S. mansoni cercaria were collected from infected snails 35–45 days post-infection. Schistosomula were produced from cercaria by mechanical transformation, as described [27], [28] and were cultured at 37°C and 5% CO2 in OPTI-MEM I medium (Invitrogen) supplemented with 10% FBS, streptomycin 100 µg/ml, penicillin 100 U/ml and fungizone 0.25 µg/ml. Adult parasites were obtained by infecting 28-day old female CD-1 mice with freshly collected cercaria (150 cercaria/animal) by skin penetration. Adult S. mansoni worms were recovered 7 weeks post-infection by perfusion of the liver [28], washed extensively and either flash-frozen in liquid nitrogen for subsequent RNA extraction or fixed in 4% paraformaldehyde (PFA) for immunolocalization experiments. Animal procedures were reviewed and approved by the Facility Animal Care Committee of McGill University (Protocol No. 3346) and were conducted in accordance with the guidelines of the Canadian Council on Animal Care.
The full-length SmGPR-3 cDNA was cloned from adult S. mansoni based on a predicted coding sequence (Smp_043290) obtained from the S. mansoni genome database (S. mansoni GeneDB; http://www.genedb.org/Homepage/Smansoni). Total RNA was purified from adult S. mansoni worms (Qiagen RNeasy kit) and was oligo-dT reverse-transcribed with MMLV reverse transcriptase (Invitrogen), according to standard procedures. SmGPR-3 was cloned with primers that targeted the beginning and end of the predicted coding sequence. The primer sequences were as follows: 5′-ATGAATTTCATAAGAAACAAAACCAATTATTC-3′ (sense) and 5′-CTATCTACATCCTTTCAAAAGTACAATATG-3′ (antisense). A proofreading Platinum Pfx DNA polymerase (Invitrogen) was used to amplify the cDNA in a standard PCR reaction (35 cycles of 94°C/15 s, 53.1°C/30 s and 68°C/90 s). The resulting amplicon (1,494 bp) was ligated to pGEM-T Easy vector (Promega) and verified by DNA sequencing of two independent clones.
The SmGPR-3 coding sequence was sub-cloned between the NcoI/XbaI restriction sites of the yeast expression vector Cp4258 (kindly provided by Dr J. Broach, Princeton University, NJ, USA). The functional expression assay was adapted from the protocol of Wang and colleagues [29] as described [15], [30]. The receptor was expressed in Saccharomyces cerevisiae strain YEX108 (MATα PFUS1-HIS3 PGPA1-Gαq(41)-GPA1-Gaq(5) can1 far1Δ 1442 his3 leu2 lys2 sst2Δ2 ste14::trp1::LYS2 ste18Δ6-3841 ste3Δ1156 tbt1-1 trp1 ura3; kindly provided by J. Broach, Princeton University, NJ, USA). This strain expresses the HIS3 reporter gene under the control of the FUS1 promoter and contains an integrated copy of a chimeric Gα gene in which the first 31 and last five codons of native yeast Gα (GPA1) were replaced with those of human Gαq [29]. Other strains carrying chimeras of GPA1 and human Gαi2, Gα12, Gαo or Gαs were tested in preliminary experiments but were found to yield lower or no receptor activity compared with strain YEX108. YPD culture medium (1% yeast extract, 2% peptone and 2% dextrose) was used to culture YEX108 strain, according to standard conditions. The lithium acetate method was performed to transform yeast with either empty Cp4258 vector (mock) or Cp4258/SmGPR-3 expression plasmid, using 200 µl mid-log phase cells, 200 µg carrier single stranded DNA (Invitrogen) and 1 µg plasmid. Positive transformants were selected on synthetic complete (SC) 2% glucose solid medium lacking leucine (SC/leu−). For agonist assays, single colonies of transformants were cultured overnight in SC/leu− liquid medium at 250 rpm/30°C. The next day, cells were washed four times in SC 2% glucose liquid medium that lacked both leucine and histidine (SC/leu−/his−). Cells were finally resuspended in SC/leu−/his− medium supplemented with 50 mM 3-(N-morpholino) propanesulfonic acid (MOPS), pH 6.8 and 1.5 mM 3-Amino-1, 2, 4-Triazole (3-AT). The latter was used to reduce basal growth due to endogenous background signalling as it inhibits the gene product of HIS3 [29]. Aliquots containing approximately 3,000 cells were added to individual wells of a 96-well plate containing test agonist or vehicle plus additional medium for a total reaction volume of 100 µl. The plates were incubated at 30°C for 22–26 h, after which 10 µl of Alamar blue (Invitrogen) was added to each well. The plates were returned to the 30°C incubator until the Alamar blue color started to shift to pink (approximately 1–4 h) and fluorescence (560 nm excitation/590 nm emission) was measured at 30°C every hour up to three hours, using a plate fluorometer (FlexStation II, Molecular Devices, US). Antagonist assays were done in a similar way, except that each well contained 10−4 M agonist (dopamine) and the antagonist at the specified concentration. Data analyses and dose-response curve fits were performed using Prism v5.0 (GraphPad software Inc.).
Polyclonal antibodies were produced in rabbits against two SmGPR-3 synthetic peptides. The first peptide (CYISYSKEYRIYSSV) is located in the predicted 2nd extracellular loop (ECL2) of SmGPR-3 (positions 186–200) and the second peptide sequence (CERKTERTIKTQRQF) is in the third intracellular loop (il3) (positions 395–409). The two peptide sequences were tested against the S. mansoni GeneDB database and the general database at The National Center for Biotechnology Information (NCBI) to insure specificity. The peptides were both conjugated to ovalbumin to increase immunogenicity. The conjugated peptides and custom antibodies were purchased from 21st Century Biochemicals, Malboro, MA, USA. Antibodies were raised in two rabbits, each of which was inoculated five times and the serum was collected prior to injection (preimmune) and up to 72 days following injection (see http://21stcenturybio.com/ for further details of the antibody protocol). ELISA was done to test the specificity of the antibodies to each of the two peptides. SmGPR-3 –specific antibodies were subsequently affinity-purified, using the MicroLink Peptide Coupling kit (Pierce), as described previously [22].
Immunoprecipitation (IP) was performed with the Seize Primary Immunoprecipitation kit (Pierce, USA), as described previously [22]. Briefly, IP affinity columns were prepared by covalent coupling of purified anti-SmGPR-3 IgG to AminoLink Plus gel in the presence of sodium cyanoborohydride. A solubilized membrane fraction was prepared from adult S. mansoni, using a commercial kit (ProteoExtract Native Membrane Protein Extraction Kit, Calbiochem) and aliquots of solubilized membrane proteins (14 µg protein) were mixed with 50 µl IgG-linked gel and incubated overnight at 4°C with gentle rotation. After incubation, the gel was washed extensively and the bound proteins were eluted under acidic conditions, using the elution buffer supplied by the kit. The immunoprecipitated proteins were resolved on 4–12% Tris-Glycine precast gel (Invitrogen) and transferred to polyvinylidene fluoride (PVDF) membranes (Millipore). Western blot analysis was performed according to standard protocols, using purified anti-SmGPR-3 antibody (dilution 1/10,000) and a goat anti-rabbit Horseradish Peroxidase (HRP)-conjugated antibody (Calbiochem) (1/20, 000). To test for specificity, the western blots were repeated with primary antibody that was preadsorbed with 0.5 mg/ml of pooled peptide antigens (0.25 mg of each peptide) or preimmune serum.
The procedure is based on the protocol of D. Halton and colleagues [31], as described previously [22], [27]. S. mansoni cercaria and adult worms were fixed in 4% paraformaldehyde (PFA) for 4 h at 4°C, washed three times in blocking buffer (1× PBS, pH 7.4; 1% bovine serum albumin; 0.1% sodium azide and 0.5% triton X-100) and were incubated in 10 mM sodium citrate for 1 h at 70°C. Animals were subsequently washed twice with PBS, incubated with affinity-purified anti-SmGPR-3 antibody (diluted 1∶25 in blocking buffer) for three days at 4°C, washed overnight with the same blocking buffer and finally incubated with goat anti-rabbit IgG conjugated to FITC (Sigma, Canada) for another three days at 4°C (dilution 1∶300). If used as a counterstain, 200 µg/ml TRITC-labelled phalloidin was added during the last two days of incubation. The samples were mounted using anti-quench mounting medium (Sigma, Canada) and examined with a BIO-RAD RADIANCE 2100 confocal laser scanning microscope (CLSM) equipped with a Nikon E800 fluorescence microscope for confocal image acquisition and the LASERSHARP 2000 software package. As negative controls, we used preimmune serum, omitted the primary antibody and used primary antibody that was preadsorbed with 0.5 mg/ml of pooled peptide antigens (0.25 mg of each peptide).
3-day old in vitro transformed schistosomula were placed in individual wells of a 24-well plate (30–40 animals/well) in 300 µl of OPTI-MEM+10% dialyzed serum. Following an adaptation period of 15 min at room temperature, test substances were added at a final concentration of 100 µM or as indicated. Animals were monitored 5 min after drug addition by placing the 24-well plate on a compound microscope (Nikon, SMZ1500) equipped with a digital video camera (QICAM Fast 1394, mono 12 bit, Qimaging) and SimplePCI version 5.2 (Compix Inc.) for image acquisition. Images were obtained at a rate of ≈3 frames/s for a period of 1 minute and the data were analyzed with ImageJ software (version 1.41, NIH, USA). Cultured schistosomula display complex motor behaviours that are dominated by repeated changes in length, both shortening and elongation. To quantify this type of movement, we used the “Fit-Ellipse” command of ImageJ to draw best-fit ellipses of individual animals in each recorded frame. An estimate of body length was obtained by measuring the principal (“major”) axis of each ellipse, using calibrated units (µm), and the frequency of length changes during the observation period was calculated. Any change representing >10% of body length, whether an increase or decrease, was included in the calculation; changes ≤10% were disregarded. Between 12–15 animals were monitored per well and the experiment was repeated a minimum of 3 times. To monitor for possible drug induced toxicity, viability tests were performed routinely using the methylene blue dye exclusion assay described by Gold [32].
Homology searches were performed by BLAST analyses (tBLASTn or BLASTp) of the S. mansoni Genome Database (S. mansoni GeneDB; www.genedb.org/genedb/smansoni/) [13], the S. japonicum Transcriptome and Proteome Database (SjTPdb) [33], the most current genome annotations of the planarians, Schmidtea mediterranea (SmedGD version 1.3.14) [34] and Macrostomum lignano (www.macgenome.org/index.html) and the general database available at the National Center for Biotechnology Information (NCBI). Sequences showing significant homology with SmGPR-3 were aligned with ClustalW and inspected manually for the presence of conserved Class A (rhodopsin-like) GPCR motifs [23]. Phylogenetic trees were generated with MEGA4 [35], using two different methods, neighbour-joining and Unweighted Pair Group Method with Arithmetic mean (UPGMA) with similar results. The trees were tested by bootstrap analysis with 1,000 replicates. Predictions of transmembrane (TM) regions were made using the TMpred server (http://www.ch.embnet.org) and by comparison with the crystal structure of the human β2 adrenergic GPCR [36]. To facilitate identification, S. mansoni sequences are described using both their S. mansoni GeneDB designation [13] and the corresponding GenBank accession numbers. S. mediterranea sequences are identified by their SmedGD designation [34]. All other sequences are identified by their GenBank accession numbers. GPCR residues located within TM regions are described according to the system of Ballesteros and Weinstein [37]. Each amino acid within a TM region is identified by the TM number (1–7) followed by the position in the TM helix relative to an invariant reference residue, which is arbitrarily assigned the number 50. Residue D3.32, for example, is located in TM3, 18 residues upstream of the invariant reference residue. Amino acids of relevance to this study are as follows (corresponding residue of SmGPR-3): R2.64 (Arg96), D3.32 (Asp117), S5.42 (Ser198), S5.43 (Ser199), T7.39 (Thr462) and Y7.43 (Tyr466).
A theoretical model of SmGPR-3 was built with Accelrys Discovery Studio (DS). Prior to generating the model, SmGPR-3 was aligned with the sequences of GPCR crystal structures available in the general protein database (PDB) (Accession numbers: 2rh1, 3eml, 1u19, 2vt4, 2z73) and the β-2 adrenergic receptor (2hr1) was selected as the best template based on similarity scores. The sequence alignment was inspected to ensure that the positions of conserved residues in the structural template were properly aligned with those of SmGPR-3, including the reference residues designated at position 50 of each helix [37] and conserved motifs, such as the DRY peptide at the end of TM3 and the NPxxY motif of TM7. Subsequent preparation of the template and construction of the model were performed with DS using default parameters. The first stretch of 28 amino acid residues corresponding to the extracellular N-terminal end was not constructed due to lack of structural information. In addition, we note that the β-2 adrenergic template is a chimeric receptor in which the region corresponding to the third intracellular loop (il3) was replaced with the sequence of T4 lysozyme [36]. Therefore the il3 of SmGPR-3 (positions 224–417) could not be aligned and was omitted from the model. Energy minimization was performed using the CHARMm forcefield in DS. The conserved disulfide linkage that occurs between the beginning of TM3 and extracellular loop 2 (Cys110 and Cys186 in SmGPR-3) was constrained during optimization of the model. The energy-minimized model was subsequently verified by means of a Ramachandran plot analysis and the PROFILES-3D evaluation method available in DS. The quality score obtained with PROFILES-3D was within acceptable range and the Ramachandran plot analysis showed 94.8% of the residues occurring in favourable regions of the plot, suggesting the model was reliable. Superimposition of the model with the β-2 adrenergic receptor template showed a backbone (Cα) root-mean-square-deviation (RMSD) of 0.83 Å and the overall protein RMSD was 1.38 Å. For ligand-docking, the structure of the ligand (dopamine or epinine) was generated with the molecular builder panel available in the software and was energy-minimized, as recommended. Next, we used DS to search for potential binding cavities. Six potential sites were identified, of which only one was located in the correct region based on the position of the binding pocket in the structural template. The ligand was subsequently docked onto this site of the SmGPR-3 model with CDOCKER. Each ligand was docked in multiple conformational states and orientations, which resulted in 240 different ligand poses for dopamine and 290 for epinine. These were examined and evaluated using the CHARMm scoring method of CDOCKER to identify potential binding residues for each ligand.
Protein content was measured with a Lowry assay (BioRad). Indirect ELISA was performed in 96-well plates coated with individual or pooled SmGPR-3 peptides (50–500 ng/well) and incubated with a serial dilution of rabbit anti-SmGPR-3 antiserum or preimmune serum (1∶30,000–1∶100), followed by incubation with a horseradish peroxidase (HRP)-labeled secondary antibody (goat anti-rabbit IgG, 1∶2,000). Quantitative PCR (qPCR) was performed as described previously [15], [52] using the Platinum SYBR Green qPCR SuperMix-UDG kit (Invitrogen) and a Rotor-Gene RG3000 (Corbett Research) real-time PCR cycler. Statistical comparisons were done with Student t-tests or a one-way ANOVA, followed by a Tukey pairwise comparison. P≤0.05 was considered statistically significant.
Predicted S. mansoni BA-like receptors [13] were aligned with BA receptors from other species, including other flatworms for which genomic data are available (S. japonicum, Dugesia japonica and S. mediterranea) and both vertebrate and invertebrate representatives of dopaminergic, serotonergic, adrenergic, histaminergic, tyramine/octopamine and structurally related cholinergic muscarinic (mACh) receptors. A phylogenetic tree of the alignment (Fig. 1) shows a subset of schistosome GPCRs (SmGPR) that are derived from a common node and constitute a separate clade within the tree. Included in this clade are seven S. mansoni sequences and two homologues from S. japonicum but no sequences from any of the other species examined, including the free-living planarians. We have previously described two members of this new clade, SmGPR-1 (formerly SmGPCR; AAF21638; Smp_043260;) and SmGPR-2 (GQ397114; Smp_043340) [15], [25]–[27]. In the present study, we cloned a third SmGPR (SmGPR-3) cDNA from adult S. mansoni by RT-PCR. The cDNA was verified by DNA sequencing (Accession # GQ259333) and was found to be identical to the corresponding genomic prediction available at the S. mansoni GeneDB (Smp_043290). SmGPR-3 has 497 amino acids and a predicted MW of 58.4 kDa.
NCBI BLASTp analyses confirmed the identity of SmGPR-3 as a member of the BA receptor family. According to pairwise alignment analyses, the most closely related sequences are those of the SmGPR clade including (% homology): SmGPR-1 (Smp_043260, 53.4%), Smp_043300 (47.4%), SmGPR-2 (Smp_043340, 46.9%), Smp_043270 (45.5%), Smp145520 (40.4%) and the S. japonicum receptor FN328430 (46.1%). SmGPR-3 is also related to other schistosome BA receptors (non-SmGPRs), as well as BA receptors from other organisms but the level of homology is generally lower (<40%). Further analysis of the SmGPR-3 protein sequence detected all the hallmark features of Class A (rhodopsin-like) GPCRs (Fig. 2). Aside from having the expected 7-TM topology, SmGPR-3 carries the signature DRY motif at the intracellular boundary of TM3, the NPxxY motif of TM7 and all the conserved reference residues at position #50 of each TM helix [37]. We also identified several residues that have been implicated in BA binding and receptor activation, notably the aromatic cluster FxxCWxPFF of TM6 and a highly conserved aspartate at position 3.32 of TM3 (D3.32/Asp117), which is considered to be one of the core binding sites in BA GPCRs [23], [36], [38], [39]. The presence of D3.32 marks an important difference between SmGPR-3 and other members of this clade. The schistosome SmGPRs are unusual in that they carry an asparagine substitution at this position [15]. SmGPR-3 is the only receptor in this group where the aspartate D3.32 is conserved.
Functional expression assays were performed in yeast. Preliminary experiments failed to detect receptor expression in mammalian cells (data not shown) and therefore we used yeast as a heterologous expression system throughout the study. The full-length SmGPR-3 cDNA was ligated to the Cp4258 vector and the recombinant plasmid was transformed into Saccharomyces cerevisae to test for receptor activity. We used a genetically modified S. cerevisae strain that is designed for GPCR activity assays [29]. The yeast is auxotrophic for histidine and expresses a HIS3 reporter gene under the control of the FUS1 promoter. Activation of a recombinant GPCR in this system increases expression of the HIS3 reporter via the yeast's endogenous pheromone response, which in turn allows the cells to grow in selective histidine-deficient medium. Thus receptor activity can be quantified based on measurements of yeast growth in the selective medium, using a fluorometric Alamar Blue assay. Cells transformed with either empty vector (mock control) or SmGPR-3 were initially tested with all different biogenic amines, each at 2×10−4 M (Fig. 3A). The results obtained from five to six individual clones showed that SmGPR-3 was selectively activated by dopamine and its naturally occurring metabolite epinine (deoxyepinephrine). Other catecholamines, including noradrenaline, adrenaline and the adrenaline metabolite, metanephrine (not shown) also stimulated SmGPR-3 activity but not to the same extent as dopamine or epinine. The receptor exhibited partial constitutive activity in the absence of agonist but there was further activation in the presence of catecholamines, whereas other biogenic amines had no significant effect relative to the no drug control. Experiments were repeated with different concentrations of dopamine or epinine and their responses were shown to be dose-dependent. The half maximal effective concentration (EC50) for dopamine and epinine activation in the yeast expression system are 3.1×10−5 M and 2.85×10−5 M, respectively (Fig. 3B).
Next we examined the effects of classical (mammalian) dopaminergic and other BA antagonists on the activity of SmGPR-3. Drugs were tested initially at a single concentration of 100 µM (10 µM in the case of flupenthixol) in the presence of 100 µM DA (Fig. 4A). The drug effects revealed an unusual pharmacological profile, which did not resemble any of the dopaminergic or adrenergic receptors of mammals. The most surprising observation was that spiperone, a mammalian D2 antagonist enhanced the activity of SmGPR-3 nearly 2-fold, thus behaving more as an agonist than a receptor blocker. Propranolol, a β-adrenoceptor antagonist had no effect on this receptor, while the remaining drugs showed various degrees of inhibition. Those drugs that produced significant inhibition (>50%) in the initial screen were subsequently tested at different concentrations to obtain dose response curves (Fig. 4B–F). The half-maximal inhibitory concentrations (IC50) for these drugs were as follows: Haloperidol, 1.4×10−6 M; Flupenthixol, 3.9×10−6 M; Promethazine, 2.8×10−5 M; Mianserin, 4.5×10−5 M; Clozapine >10−4 M. Based on this analysis, the most effective antagonists of SmGPR-3 were haloperidol and flupenthixol, two classical DA antagonists, followed by promethazine, an antihistaminic drug not known to interact with dopaminergic receptors. Mianserin, (mixed adrenergic/5HT antagonist) and cyproheptadine, (mixed histamine/5-HT antagonist) both produced significant inhibition of ≈70% at the highest concentration tested. Finally the remaining drugs caused modest or no significant inhibition at 100 µM, including clozapine, a classical dopamine antagonist. Because the assay is based on cell growth, we questioned whether the strong inhibition induced by promethazine, flupenthixol and haloperidol were due to drug-induced toxicity leading to cell death. To test this possibility, we repeated the assay in medium supplemented with histidine (100 µM), which enables cell growth irrespective of receptor activation. The results showed normal or nearly normal growth in the drug-treated cells in the presence of histidine, indicating that the inhibitory effect of the drug was receptor-mediated and not the product of generalized toxicity (Fig. 4A).
To investigate the tissue localization of SmGPR-3 we obtained a specific antibody that targets two unique peptides of the receptor. The antibody was affinity-purified and verified first by ELISA. To test if the antibody could recognize the native receptor, we immunoprecipitated (IP) SmGPR-3 from solubilized S. mansoni membranes, using covalently-coupled anti-SmGPR-3 antibody beads and then probed the IP eluate by western blotting with affinity-purified anti-SmGPR-3 antibody. The results (Supplemental Fig. S1) detected a single major band of about 60 kDa, which is consistent with the expected size of SmGPR-3. The negative controls were similarly immunoprecipitated with the same antibody beads but were probed either with peptide-preadsorbed antibody or preimmune serum. The results show much diminished or no immunoreactivity in the negative controls, suggesting the antibody recognizes SmGPR-3 specifically. For the in situ immunolocalization studies, larval and adult stages of S. mansoni were probed with affinity-purified anti-SmGPR-3, followed by a FITC-labelled secondary antibody. Some animals were also treated with TRITC-conjugated phalloidin to label cytoskeletal elements and muscle [31]. The results show strong SmGPR-3 green fluorescence in the nervous system of both cercaria and adult S. mansoni. In cercaria we see immunoreactivity primarily in the CNS along the main longitudinal nerve cords and transverse commissures (Fig. 5). The pattern of labelling is similar to that of dopamine itself, which localizes to the same regions of the CNS in cercaria of both S. mansoni and S. japonicum [14]. Adult worms have high levels of SmGPR-3 immunoreactivity in the cerebral ganglia and major longitudinal nerve cords. This was observed in adult males (Fig. 6A) as well as female worms (not shown). The peripheral nervous system (PNS) is also rich in SmGPR-3. Immunoreactivity can be seen in the innervation of the caecum (Fig. 6B) and the peripheral plexuses and fibers innervating the parasite musculature (Fig. 6C), both circular and longitudinal muscles (Fig. 6D, E). There is no apparent co-localization of SmGPR-3 labelling (green) and the somatic muscles (red) that were counterstained with TRITC-conjugated phalloidin, suggesting the receptor is probably associated with the innervation of the musculature rather than the muscle itself. Other regions of significant labelling include the peripheral plexus of the ventral sucker (Fig. 6F) and, in male worms, the tubercles and the innervation of the testes (Fig. 6G–I). No significant fluorescence was observed in any of the negative controls tested, including worms probed with pre-immune serum, secondary antibody only and antibody that was pre-adsorbed with the peptide antigens.
The prevalence of SmGPR-3 in the peripheral innervation of the somatic muscles suggests this receptor could be involved in the control of motor activity. To explore this possibility we tested the effects of several SmGPR-3 agonists and antagonists on the motor activity of intact schistosomula in culture. The goal of these studies was to determine whether substances that interact with the recombinant receptor in vitro also influence worm movement, which could suggest involvement of SmGPR-3 in motor control. We used a larval stage (schistosomula) instead of adult worms because the larvae are better suited for quantitative measurements of movement. A preliminary analysis by confocal immunofluorescence detected expression of SmGPR-3 in the nervous system of the schistosomula (data not shown). We also determined that SmGPR-3 could be amplified from schistosomula oligo-dT reverse-transcribed cDNA using both end-point and real-time quantitative PCR (Supplemental Fig. S2), thus confirming that the receptor is expressed in this larval stage. Measuring schistosome movement is challenging because the worms do not travel (i.e. swim) in culture. To quantitate motor activity, we monitored individual schistosomula in the presence and absence of test substances, using a microscope equipped with a video camera and imaging software. Approximately 180 consecutive frames were recorded for each animal over 1 min of observation and the approximate length of the animal in each frame was measured in calibrated units (µm). An individual recording of a control (untreated) animal is shown in Fig. 7A (top panel). When cultured in vitro, schistosomula exhibit repeated cycles of elongation and contraction, which cause the animal to increase or decrease its body length by as much as 20% in either direction. Treatment with catecholamines significantly altered this behaviour. The addition of dopamine or epinine, each at a concentration of 100 µM caused marked inhibition of motor activity. Dopamine decreased both the amplitude and frequency of contractions and epinine caused virtual paralysis in all animals tested. To quantify the level of motor activity, we repeated experiments at different concentrations of each amine and then calculated the frequency of length changes, both decrease and increase for each individual animal during the 1 min of observation. Mean values were obtained from the average of 12–15 animals/experiment and each experiment was repeated a minimum of 3 times. The data (Fig. 7B) confirmed the inhibition caused by dopamine and epinine and showed that the effect was dose-dependent. In both cases, significant inhibition was observed at concentrations equal to or >10 µM.
The study was subsequently expanded to test other drugs that normally interact with BA receptors and were shown to have activity towards recombinant SmGPR-3 in the yeast assay. Substances were tested at 50 µM (flupenthixol, promethazine, epinine, dopamine) or 500 µM (adrenaline, metanephrine, haloperidol). With the exception of haloperidol, which had no significant motor effects, all the other substances produced significant changes (P<0.05) in motor activity compared to the untreated larvae (Fig. 7C). However the results do not show a correlation between the phenotypic effects of these drugs and those seen in vitro with the recombinant receptor. Adrenaline and its methylated derivative, metanephrine have weak agonist activity towards SmGPR-3 in the yeast assay, yet they were found to have opposite effects on larval motility. Adrenaline produced the same decrease in motility as dopamine and epinine, whereas metanephrine caused marked hyperactivation, increasing the frequency of body wall movements ≈3-fold. Among the antagonists of SmGPR-3, haloperidol had no effect but flupenthixol and promethazine both caused a significant decrease in motor activity. Drug-treated animals were routinely assayed for viability, using the methylene blue dye exclusion assay described by Gold [32]. There were no measurable effects on viability at the drug concentrations tested, though we cannot rule out some degree of drug-induced toxicity that would not be detected by this assay.
Studies of mammalian BA receptors have localized the agonist binding pocket to a crevice formed by residues near the extracellular side of the TM helices, particularly TM3, 5, 6 and 7 [23], [36], [38]–[43]. Several ligand-interacting residues have been identified within this pocket. Of particular importance in catecholamine (dopamine, adrenergic) receptors is the aforementioned aspartate of TM3 (D3.32), which anchors the protonated amino group of the ligand, and three closely positioned serines in TM 5 (S5.42, S5.43, S5.46), which interact with the hydroxyl groups on the catechol ring. Three of these residues are also conserved in SmGPR-3 (D3.32, S5.42, S5.43) and therefore we questioned whether they might be involved in the interaction with dopamine. A homology model of SmGPR-3 was obtained from a structural alignment with the β2-adrenergic receptor (2rh1) and used for docking simulations. The hypothetical structure shows the typical topology of a Class A GPCR with 7 TM helices and one additional short helix (helix 8) at the C-terminal end that runs parallel to the membrane (Fig. 8A). Starting with this structure, we searched for potential binding cavities, using the site finder command in Discovery Studio. The analysis identified a possible site, which was located in the expected region of the receptor and included the canonical D3.32 of TM3 (residue Asp117). The dopamine structure was docked virtually onto this site and 240 docked ligand poses were examined and scored to search for the most favourable binding interactions. We identified approximately 20 residues located within 5 Å of the docked ligand, which are predicted to form the binding pocket of the receptor (Table S1). Among these residues there are five amino acids showing direct interactions with dopamine (Arg96/R2.64, Asp117/D3.32, Ser198/S5.42, Thr462/T7.39 and Tyr466/Y7.43) (Fig. 8A, B). The protonated amino group of dopamine is anchored to Asp117/D3.32 of TM3, as expected, and also forms additional H-bond interactions with Thr462/T7.39 and Tyr466/Y7.43of TM7. As for the catechol ring of dopamine, we observed more that one set of interactions depending on the orientation of the ligand. The most favourable orientation (highest docking scores) had the catechol ring interacting with Arg96/R2.64 of TM2 (Fig. 8C, yellow). Dopamine could also be docked in a different orientation where the catechol ring was pointed towards TM5 and interacted with one of the conserved serines of TM5, Ser198/S5.42 (Fig. 8C, magenta) but the docking scores were lower and we did not see interactions with the other conserved serine of TM5 (Ser199/S5.43). About 85% of the docked poses examined had interactions with Arg96/R2.64 of TM2 instead of Ser198/S5.42 (Fig. 8D). Since SmGPR-3 is as responsive to epinine as it is to dopamine, we repeated the docking simulation using epinine as the ligand. An overlay of all the potential docked poses shows that epinine binds to the same pocket as dopamine and interacts with the same core residues (Fig. 8E). Like dopamine, the most favourable epinine interactions involved residues of TM3 (Asp117/D3.32), TM2 (Arg96/R2.64) and TM7 (Thr462/T7.39 and Tyr466/Y7.43). We also observed less favourable interactions between the catechol ring of epinine and TM5, similar to dopamine, except in this case interactions occurred with both conserved serines, Ser198/S5.42 and Ser199/S5.43.
The S. mansoni genome encodes 16 GPCR-like sequences that share significant homology with aminergic receptors from other species [13]. A comparative bioinformatics analysis of these putative S. mansoni BA receptors with homologues of vertebrate and invertebrate origins enabled us to identify a new clade, the SmGPRs, which contains approximately half of all S. mansoni aminergic receptors and at least another two homologues of S. japonicum [15]. SmGPRs are easily recognized as Class A GPCRs due to their heptahelical organization, short N-terminus and the presence of signature motifs such as the DRY peptide (near TM3), FxxCWxxFF (TM6) and NPxxY (TM7) among others. Nevertheless, these receptors are considered novel; they are clearly more related to each other than they are to other BA receptors in the database, including other schistosome BA receptors. The dendogram analysis suggests that the SmGPRs diverged from a common ancestor early in evolution, probably through a series of gene duplications that gave rise first to the SmGPR-1/SmGPR-3 branch and subsequently the remaining sequences. We note that the SmGPR sequences grouped within this clade are all located in relatively close proximity to each other in the S. mansoni genome (Smp_scaff000103) [13], which is consistent with a mechanism of gene duplication. A distinctive feature of the SmGPRs is the absence of residue D3.32, which is replaced with an asparagine in all of these sequences except for SmGPR-3. As mentioned earlier, D3.32 is highly conserved in BA receptors [22], [36], [38], [39] and therefore the asparagine substitution marks a significant departure from current models of receptor structure. Thus far SmGPRs have been found only in S. mansoni and S. japonicum. We checked all the available planarian genomes as well as the general database at NCBI and found no other sequences that aligned within this clade. As more genomes become available, it will be of interest to determine if SmGPRs are present in other parasitic flatworms or if they are unique to schistosomes.
The functional expression analysis of SmGPR-3 was conducted in yeast cells. Yeast provides a robust expression system for recombinant, hard to express proteins. SmGPR-3 could not be expressed in mammalian cells (not shown) but it produced a functional receptor when expressed in the budding yeast S. cerevisiae. SmGPR-3 is the third member of this clade to be cloned and characterized. The first two (SmGPR-1 and -2) were found to be selectively activated by histamine when expressed in vitro [15], [25]. Given the structural similarity among these receptors, we questioned whether SmGPR-3 might also be responsive to histamine. However the activity assays revealed that SmGPR-3 was preferentially activated by catecholamines. The strongest response was obtained with dopamine and its methylated derivative, epinine, but the receptor also recognized noradrenaline and adrenaline, suggesting broad specificity for catechol derivatives. When tested against classical dopaminergic and other BA antagonists, SmGPR-3 exhibited a mixed pharmacological profile that did not conform to any known mammalian receptor. The most potent inhibitors of SmGPR-3 included two classical DA antagonists (haloperidol, flupenthixol), an anti-histaminic drug (promethazine) and a relatively nonselective inhibitor (cyproheptadine) that normally targets serotonin and histaminergic receptors. In contrast, anti-dopaminergics such as clozapine and spiperone either had little inhibitory effect or exhibited agonist activity. Interestingly, we noted that two of the most potent inhibitors of SmGPR-3, promethazine and flupenthixol, are also the most potent inhibitors of SmGPR-2 [15], suggesting that SmGPRs may have common pharmacological properties, even if the amine ligands are different.
To explore the role of SmGPR-3 in the bloodfluke, we produced a specific polyclonal antibody and then examined the tissue distribution of the receptor in different developmental stages of the parasite by confocal immunofluorescence. The results show that this receptor is predominantly neuronal and is expressed in the nervous system of both larvae (cercaria) and adult worms. The distribution of SmGPR-3 immunoreactivity in cercaria closely parallels that of catecholamine-containing neurons, which were previously described in S. mansoni and S. japonicum by biochemical methods [14]. In cercaria these neurons localize in part to the major longitudinal nerve cords and the posterior transverse commissure, sites that are also enriched in SmGPR-3. SmGPR-3 is probably activated by dopamine released from these centrally located neurons and mediates inter-neuronal dopamine signalling within the cercarial CNS. In the adults, SmGPR-3 is more widespread and localizes both to the CNS and elements of the PNS, suggesting a greater diversity of activities. As in the larvae, we detected strong immunoreactivity throughout the major longitudinal nerve cords and cerebral ganglia. In addition, the adult worms express high levels of SmGPR-3 outside the CNS, particularly in the smaller peripheral nerve fibers and plexuses that supply the body wall muscles and the suckers among other tissues. We do not know how this compares with the distribution of dopamine neurons in the adults, which has yet to determined. However there is evidence from other flatworms that dopamine neurons are present both in the CNS and the peripheral innervation of the somatic muscles, where dopamine is believed to play an important role in motor control [16], [18]. Our results thus suggest that SmGPR-3 has important activities both within and outside the CNS in adult worms. The prevalence of SmGPR-3 in the nerve fibers of the circular muscles strongly implicates this receptor in the control of muscle function and suggests this is one of the mechanisms by which dopamine controls movement of the worm.
The first evidence of dopaminergic activity in schistosomes dates back to the 1970s, when researchers reported that addition of dopamine caused a pronounced lengthening of the body in intact adult worms [44], [45]. Subsequent studies demonstrated that the lengthening effect was due to relaxation of the body wall muscles [21]. It was further suggested that dopamine had both direct and indirect effects on the musculature, possibly by acting through more than one receptor [21], though the molecular mechanisms were unknown. Our continuing investigation of this system is beginning to shed new light on these early observations. SmGPR-3 is the second dopamine receptor to be identified in S. mansoni. The tissue distribution of the other receptor, named SmD2 is quite different from that described here. SmD2 is absent in the CNS and it is enriched in the body wall muscles [22]. Although SmGPR-3 is also present in the body wall, it is associated with neuronal elements rather than the muscle per se; we did not see any co-labelling of SmGPR-3 with phalloidin, the counterstain that was used to probe the muscle layers. Thus, we conclude that there are at least two routes of dopaminergic motor control in S. mansoni involving both direct and indirect mechanisms, as was hypothesized in the earlier studies. One pathway is mediated by SmD2, which is predicted to act directly on the musculature, whereas the second is a more indirect neuronal pathway mediated by SmGPR-3. The localization of SmGPR-3 suggests that the receptor probably controls neuronal input to the body wall muscles, most likely by modulating release and/or signalling activity of another transmitter. This type of indirect activity is consistent with the notion that BAs can function both as neuromodulators and classical neurotransmitters. For example, in C. elegans dopamine controls motor activity indirectly through centrally located neurons that in turn modulate muscle function [46].
While it is well established that dopamine has inhibitory neuromuscular effects in schistosomes [21], the outcome of these effects on movement is unclear. Some early studies reported an increase in the length of the body but no effect on motility [44] whereas others reported both an increase in length and a decrease in motor activity when adult worms were treated with dopamine [45]. These studies were based on qualitative assessments of worm movement and did not include larval stages of the parasite. We have revisited this issue using a more quantitative imaging assay and in vitro transformed schistosomula, which are best suited for measurements of motility in culture. The results support the notion that dopamine has inhibitory motor effects in schistosomes; dopamine-treated larvae were significantly less motile than the controls. The inhibition was dose-dependent, starting at about 10 µM, and resulted in nearly complete paralysis at higher concentrations. Similar effects were observed in larvae treated with adrenaline and the dopamine metabolite, epinine, suggesting this is a common response to catecholamines. The inhibitory effect of dopamine described here is much stronger than that reported earlier for adult worms [44], [45] but, surprisingly, we did not see an overall increase in body length, which was a predominant response in the adults. This discrepancy could be due to differences in experimental protocol or it could be a function of the different developmental stages used in the studies, larvae versus adults. As mentioned earlier, the distribution of SmGPR-3 is more widespread in the adult worms than the larvae. SmGPR-3 localizes to the peripheral innervation of the somatic muscles in the adults whereas in the larvae it is more restricted to the CNS. It is possible the different dopamine-induced behaviours, paralysis versus elongation, are due to developmental changes in the expression patterns of these receptors.
Besides catecholamines, we repeated motility assays with a variety of substances that were shown to interact with the recombinant SmGPR-3 in the yeast system. Given that exogenous dopamine paralyzes the larvae, we had expected antagonists of this receptor to cause hyperactivity but that was not observed. One of the antagonists tested, haloperidol, caused a small increase in motility but this was not statistically significant, whereas the other two (flupenthixol and promethazine) had the opposite effect and caused marked inhibition of movement at the concentrations tested. A probable explanation for these results is that the drugs are interacting with more than one receptor in the intact parasite, for example the aforementioned SmD2 receptor, a SmGPR homologue, or others that have yet to be identified. The paralysis caused by promethazine and flupenthixol could be due to interactions with SmGPR-2, which is also strongly inhibited by these drugs [15]. Another consideration is the possibility of general (receptor-independent) toxicity effects. The drugs used in this study did not affect parasite viability but we cannot rule out other, more subtle effects that might have hindered motility and were not detected in this study. One of the more unexpected results of this survey was the effect of metanephrine on the larvae. Metanephrine is structurally related to adrenaline but it had the opposite effect on larval movement. Whereas adrenaline caused paralysis, the metabolite caused very strong hyperactivity, more than doubling the frequency of movement. Metanephrine was found to have weak agonist activity towards SmGPR-3 in vitro (not shown); it is possible that an excess of exogenous metanephrine is able to compete with endogenous dopamine (or adrenaline) for this receptor, thus decreasing endogenous signalling. Although the mode of action remains unclear, the results nonetheless show that catecholamine metabolites and other dopaminergic agents have strong effects on larval motility. The challenge for future studies is to determine which of the S. mansoni dopamine receptors, smGPR-3, SmD2 or others is targeted by these drugs and to identify more selective inhibitors that could be used to disrupt parasite movement.
The strong effects of epinine and metanephrine on larval motility raise interesting questions about the possible role of these substances in the parasite. In mammals, epinine is a naturally occurring but relatively minor by-product of DA metabolism. It is produced from DA by the activity of phenylethanolamine N-methyltransferase (PNMT) and can be further metabolized to adrenaline, though this reaction rarely occurs in mammals [47]. Metanephrine is one of the major metabolites of adrenaline in humans; it is synthesized through the action of catechol-O-methyl transferase (COMT) and it is normally excreted in the urine. Although they are considered to be inactive metabolites in higher organisms, epinine and metanephrine are biologically active in some protozoa and invertebrates. For instance, epinine replaces noradrenaline as the major substrate for adrenaline biosynthesis in the unicellular protozoan Tetrahymena pyriformi [48]. In the cnidarian, sea pansy Renilla koellikeri, epinine, metanephrine and another related metabolite, normetanephrine, are all present at high levels and are believed to be neuroactive [49]. It is unknown if these substances also occur in flatworms. We note, however, that at least some of the enzymes required for endogenous biosynthesis of these substances are present in schistosomes. The S. mansoni genome encodes a putative O-methyltransferase (CAZ32787, Smp_052470) that shares significant homology with COMTs from other species. If these metabolites are produced by the parasite, their interaction with SmGPR-3 and strong effects on larval motility could prove to be biologically important.
In the absence of a crystal structure, which is lacking for all but a few GPCRs, researchers often resort to homology models to explore the structural organization of new receptors. We have produced a homology model of SmGPR-3, using the human β2-adrenergic receptor as a structural template, and then performed virtual docking of dopamine onto the hypothetical structure. It has been published that dopamine/adrenergic receptors bind their BA ligands through the highly conserved D3.32 in TM3 and through one or more of three serines (S5.42, S5.43 and S5.46) present in TM5 [23], [36], [39]–[43]. SmGPR-3 has the conserved D3.32 of TM3 and two of the serines (S5.42 and S5.43) in TM5 but S5.46 is missing. We questioned whether a different residue might substitute for the missing serine or, given the novelty of the sequence, dopamine could interact with different residues entirely. Our docking simulation verified the interaction with D3.32 (Asp117) but suggests that the serines of TM5 are unlikely to play a major role in dopamine binding. Instead, we found potential binding sites in TM2 and TM7. One of the more interesting findings of this analysis is the predicted interaction with R2.64 (Arg96) of TM2. This arginine is not conserved in the mammalian dopamine receptors and the cognate residue (alanine or hydrophobic) is not known to be directly involved in dopamine binding. Another surprising difference is the apparent absence of ligand interactions with aromatic residues of TM6, which are conserved in SmGPR-3 (F6.51, F6.52) and would normally be expected to contribute to the binding of the catechol moiety. These differences must be viewed with caution, given the low sequence homology of SmGPR-3 compared to the template and the inevitable artefacts associated with homology models. Nevertheless, the results identify potentially important differences between the parasite and host receptors that can now be tested by mutagenesis and functional analyses. The predicted involvement of R2.64 is particularly noteworthy because of its location near the extracellular junction of TM2. The TM2 junction is a major component of the antagonist binding pocket in the human D3 receptor [40] and therefore this region could be important for the function and pharmacology of SmGPR-3. The fact that R2.64 is not conserved in the host (but is present in all the SmGPRs) further identifies this region as a potential target for the development of selective receptor blockers.
This study adds to a growing body of molecular evidence that points to dopamine as a major neurotransmitter of the schistosome nervous system. Besides SmGPR-3 and the previously described SmD2 receptor [22], researchers have characterized a dopamine biosynthetic enzyme in S. mansoni [50] and, more recently, discovered a high-affinity dopamine transporter [51], which is likely involved in the recycling/inactivation of the amine. The broad distribution of SmGPR-3 reported here suggests a great diversity of dopamine activities in this parasite, more than was previously believed. Although we have focussed our attention on neuromuscular and motor effects, it should be noted that SmGPR-3 was also found in the innervation of the caecum, the tubercles and, interestingly, the male reproductive system. Very little is known about the neuronal control of reproduction in schistosomes. We have previously identified a glutamate receptor in the female reproductive tract of S. mansoni [52] but, to our knowledge, this is the first evidence of a neurotransmitter receptor expressed in the male testes and it suggests a novel role for dopamine in this system. The presence of SmGPR-3 in the male tubercles is also noteworthy. Several neurotransmitters and neuronal proteins have been identified in the tubercles of schistosomes, where they are believed to be associated with sensory nerve endings [53]–[55]. SmGPR-3 expressed in these sites could play an important role in chemosensory signalling, either as part of an endogenous pathway or in response to exogenous (host-derived) catecholamines. More research is needed to clarify the nature of these effects and to elucidate the mode of action of dopamine in the parasites.
|
10.1371/journal.pcbi.1000713 | Estimating the Stoichiometry of HIV Neutralization | HIV-1 virions infect target cells by first establishing contact between envelope glycoprotein trimers on the virion's surface and CD4 receptors on a target cell, recruiting co-receptors, fusing with the cell membrane and finally releasing the genetic material into the target cell. Specific experimental setups allow the study of the number of trimer-receptor-interactions needed for infection, i.e., the stoichiometry of entry and also the number of antibodies needed to prevent one trimer from engaging successfully in the entry process, i.e., the stoichiometry of (trimer) neutralization. Mathematical models are required to infer the stoichiometric parameters from these experimental data. Recently, we developed mathematical models for the estimations of the stoichiometry of entry [1]. In this article, we show how our models can be extended to investigate the stoichiometry of trimer neutralization. We study how various biological parameters affect the estimate of the stoichiometry of neutralization. We find that the distribution of trimer numbers—which is also an important determinant of the stoichiometry of entry—influences the estimated value of the stoichiometry of neutralization. In contrast, other parameters, which characterize the experimental system, diminish the information we can extract from the data about the stoichiometry of neutralization, and thus reduce our confidence in the estimate. We illustrate the use of our models by re-analyzing previously published data on the neutralization sensitivity [2], which contains measurements of neutralization sensitivity of viruses with different envelope proteins to antibodies with various specificities. Our mathematical framework represents the formal basis for the estimation of the stoichiometry of neutralization. Together with the stoichiometry of entry, the stoichiometry of trimer neutralization will allow one to calculate how many antibodies are required to neutralize a virion or even an entire population of virions.
| A large part of the research on the Human Immunodeficiency Virus focuses on how virus particles attach and enter their target cells, and how entry can be inhibited by antibodies or antiretroviral drugs. Because virus particles are too small to be observed in action the inference of the details of HIV entry has to be indirect—involving the genetic manipulation of virions, and often mathematical modeling. It is known that virus particles establish contact to their target cells with spikes on their surface, and antibodies binding to these spikes can inhibit virus entry. It is not known, however, how many antibodies are needed to neutralize a spike. In this article, we develop a mathematical framework to estimate this number, called the stoichiometry of neutralization, from data obtained in experiments with genetically engineered virions. An estimate of the stoichiometry of neutralization for different antibodies is important, as it will allow us to calculate the amount of antibodies required to abrogate virus replication.
| Virions of human immunodeficiency virus (HIV) are coated by a lipid bilayer. Trimers of the dimeric envelope proteins (Envs) gp120 and gp41 are inserted into this membrane [3]–[5]. These trimers, often also referred to with the more general term spikes, can bind to CD4 receptors [6],[7]. After successful engagement of CD4, the envelope trimer undergoes conformational changes that allow a coreceptor, most commonly chemokine receptors CCR5 and CXCR4, to bind [8]. Binding to the coreceptor initiates a series of rearrangements in the viral envelope protein gp41, which upon insertion of the fusion peptide into the cellular membrane brings together viral and cellular membrane and triggers the fusion process. Possible targets for neutralizing antibodies have been identified over the past decades and are restricted to accessible conserved regions on the Env trimers [9],[10].
Estimating the number of monoclonal antibodies or inhibitory molecules needed to block a single trimer together with estimates of other parameters that characterize the molecular interaction of the virus with its target cells and antibodies, may eventually allow us to predict the antibody concentrations required to inhibit viral replication in vitro and within the infected individual. This quantitative understanding of neutralizing antibody activity will aid the development of antibody vaccines and entry inhibitors.
In this paper, we develop a mathematical framework to estimate how many antibodies are needed to neutralize a single trimer. This number is referred to as stoichiometry of trimer neutralization or short stoichiometry of neutralization. The mathematical models, which we introduce here, are based on the models we developed for the analysis of the number of trimers required for cell entry [1]. As for the stoichiometry of entry, we investigate models differing with respect to the biological assumptions about the exact molecular mechanisms involved in the generation of pseudotyped virions. To illustrate how to use our model to estimate the stoichiometry of neutralization, we reanalyze previously published data by [2].
Here, we briefly introduce the experimental setup for the determination of the stoichiometry of neutralization, in particular those aspects relevant for the development of the mathematical models in the next section. The experimental setup is described in more detail in [1],[2],[11].
Envelope-pseudotyped HIV virions are generated by transfection of virus producer cells (293T) with a set of plasmids. One plasmid provides all the genetic information to assemble infectious virions with the exception of the viral envelope. The genetic information for viral envelope proteins is provided on separate plasmids. A third plasmid encodes for the firefly luciferase reporter gene under the control of HIV LTR, which allows rapid detection of infected target cells. The resulting virions contain viral envelope proteins and are infectious but are only capable of completing one infection cycle, as the genetic information packaged into the virions lacks essential genes.
To study the stoichiometry of neutralization pseudotyped virions with mixed envelope proteins are generated. Hereby, plasmids encoding for wild-type envelope proteins are transfected along with plasmids encoding for neutralization-resistant envelope proteins. As a result, the plasmid pool in the producer cell consists of a mixture of wild-type and mutant envelope proteins. Proteins from this pool trimerize and, as a consequence, a fraction of the envelope trimers are wild-type/mutant hetero-trimers. We denote the fraction of mutant envelope protein encoding plasmids by . The mutant envelopes harbor only one (or few) amino acid changes compared to the wild-type that render them resistant to a specific neutralizing antibody. Otherwise the mutant envelope proteins are fully functional and can form functional hetero-trimers [12].
The infectivity of these pseudotyped virions with mixed envelope protein trimers is then measured. Before these virions infect target cells, they are saturated with monoclonal antibodies that bind to all wild-type envelope proteins. As a consequence, only mutant envelope proteins can take part in attachment to CD4-receptors. In this assay, infectivity is measured via the luciferase reporter gene, which is expressed upon infection of susceptible target cells. The luciferase activity (measured as emitted relative light units) in the infected cell population is proportional to the number of virions that successfully entered and integrated into a cell. The infectivity is normalized to a virus stock that contains 100% wild-type Envs. Similar to the study of the stoichiometry of entry [1], the relative infectivity, RI, is determined for different fractions of mutated envelope encoding plasmids.
The mathematical models to infer the stoichiometry of (trimer) neutralization, , incorporate the combinatorial aspects of the assembly of pseudotyped virions with mixed envelope proteins and the infection of cells in the infectivity assay.
One of the most important input parameter in all of these models is the distribution of the number of trimers on the surface of virions. We include this distribution in a generic form with the parameters , , where denotes the fraction of virions with trimers. Note that this distribution only describes the numerical and not the spatial distribution of trimers on the virion's surface.
For the basic model we assume that the envelope proteins to be assembled into trimers are sampled out of an envelope pool. The fraction of mutated envelope proteins in this pool is equal to the fraction of mutant Env encoding plasmids in the transfection medium, . Trimers are formed perfectly randomly from the envelope proteins in the pool, i.e. the number of mutated Env proteins is binomial distributed. Virions can infect a cell if they have at least functional trimers.
In the four model extensions we relax different assumptions of the basic model.
Which virions end up infecting a cell? To answer this question we first have to zoom in on the trimeric level. A trimer is called functional if it is able to take part in mediating cell entry. As virions are saturated with antibodies before the infection experiments, this ability is dependent on the stoichiometry parameter . In the absence of antibodies, both mutant and wild-type Envs are assumed to be perfectly functional and give rise to infectious particles. In the investigated setup however, antibodies bind to wild-type Envs and all wild-type Envs are assumed to be bound by one antibody. If a trimer has or more wild-type envelope proteins, this trimer is neutralized. Hence, in this setup only trimers with more than mutated envelope proteins are functional trimers. Figure 1 gives an overview of functional and non-functional trimers depending on the stoichiometry of neutralization . Here lies the important difference between the scenario studied in our work on HIV-entry [1] and the assays to estimate the neutralization parameter [2]. For estimating the entry parameter a mutation was used which renders the complete trimer binding-incapable, i.e. only trimers without any mutated Env protein are functional ones. In the neutralization assay, both wild-type and mutant Envs are infectious and only wild-type Envs can be rendered non-infectious by binding neutralizing antibody.
Not all virions that can potentially infect a cell end up in successfully infecting a cell. We call a virion infectious if it has the potential to infect a cell. Therefore it has to fulfill special conditions concerning the number of functional trimers which depend on the model and which are defined for every model separately. We assume that every infectious virion has the same probability to infect a cell independent of the number of functional trimers. Since we study the infectivities of a mixed virion stock in comparison to a wild-type stock this quantity cancels out in the calculations.
In this section, we first analyze the effects of the input parameters on the predicted relative infectivity. We show predictions for the basic model in detail. Reversing these predictions, we estimate the stoichiometry of neutralization below by fitting our mathematical models to data obtained by experiments of Yang et al. [2].
The relative infectivity RI increases with the fraction of neutralization resistant envelope proteins . This is quite intuitive because, the more neutralization resistant envelope proteins exist in the transfectable cell, the more trimers with a high number of mutant envelopes are assembled. These trimers are more likely to be functional in the presence of antibody.
The higher the stoichiometry of neutralization , the more the relative infectivity curve is shifted to the left (see figure 3 (A)). For example, let us assume only virions with exactly 10 trimers of which 8 trimers are needed for entry (). If three antibodies are needed to neutralize one trimer a trimer with one or more mutant envelope proteins is functional. For a fraction of neutralization resistant envelope proteins , the probability for a functional trimer is 78.4% and this leads to a relative infectivity (see figure 3 (A)). In contrast, the relative infectivity for and are almost 0 due to the small probabilities for functional trimers. Only 35.2% of all trimers are functional if 2 antibodies are needed for neutralization (), and 6.4% are functional for .
The stoichiometry of entry describes the minimal number of trimer-target cell receptor interactions needed to mediate cell entry. If only few trimers are necessary for attachment and entry, the probability that a virion has enough functional trimers is much higher than for a large number of trimer–target interactions . Therefore the predicted relative infectivity decreases with the stoichiometry of entry (cf. the blue and black solid curves in figure 3 (B)).
A problem for the estimation of the stoichiometric parameters is that there are various combinations of and which give rise to very similar predicted relative infectivities. As an example, let us assume a virion population with exactly 10 trimers per virion. For this situation the stoichiometry parameter pairs ; and predict similar relative infectivity values (see the blue curves in figure 3 (B)). This suggests that it is not advisable to estimate both stoichiometry parameters and simultaneously (see also below).
As shown in figure 3 (C) mean and variance of the distribution of trimer numbers play also important roles for the predictions of the relative infectivity. The higher the mean trimer number, the faster the relative infectivity increases. The variance of the trimer number distribution changes the smoothness of the predicted relative infectivity curve. Since parameter estimations are based on the predicted relative infectivites, it is necessary to include as much information as possible about the distribution of trimer numbers. [18] investigated trimers on HIV-1 virions and found trimer numbers of trimers on a virion's surface. Since we do not have more detailed information about trimer number distribution, our estimates are based on a discretized -distribution with mean 14 and standard deviation 7 (see figure 5 in [1]).
Essential for the estimation of the stoichiometric parameters is the presence of hetero-trimers. As one can see in figure 1 wild-type homo-trimers are neutralized for every stoichiometry of neutralization and mutant homo-trimers are not neutralized for any . Therefore, most of the information about the stoichiometry of neutralization is contained in the infectivity of virions with hetero-trimers.
In the imperfect transfection model, we introduce the variance coefficient to allow the envelope pool to vary from the fraction of antibody-resistant Env-mutant encoding plasmids . A high corresponds to a scenario in which almost all cells are transfected exclusively with either wild-type or mutant Env encoding plasmids. As the presence of hetero-trimers is crucial for the determination of the stoichiometry of neutralization, the distinguishability between different estimates of the stoichiometry of neutralization decreases with increasing . This effect is depicted in figure 4 (A) and (B). From this figure it becomes clear that for variance coefficients close to one different values of the stoichiometry of neutralization lead to almost identical predictions of the relative infectivity and make a reliable estimate for the stoichiometry of neutralization impossible.
In the segregation model, the segregation coefficient allows a deviance from the binomial sampling from envelope proteins out of the envelope pool. When is very close to 1, almost only homo-trimers are formed. Therefore the distinguishability between the predictions of the relative infectivity for the different stoichiometries of neutralization decreases with increasing (see figure 4 (C) and (D)). As for high values of , the estimation of the stoichiometry of neutralization is extremely difficult when the segregation parameter is close to 1.
To estimate the stoichiometry of neutralization and the other model parameters, we re-analyze data obtained in [2]. Yang et al. investigated 4 different HIV1-strains (ADA, YU2, HXBc2 and KB9) which in sum had 11 different envelope glycoprotein mutants that rendered them insensitive to one or several of the 9 different neutralizing antibodies (b12, 2F5, 2G12, 1121 F105, F91,15e,17b and 48d). Infectivity and neutralization was studied on two different target cell types (Cf2Th-CD4/CCR5 and Cf2Th-CD4/CXCR4). In total, 15 different virus antibody combinations were available for our reanalysis [2]. To demonstrate how our models can be used to derive the stoichiometry of neutralization, we treat this data set in two different ways. First, we include all data points into our estimation. This assumes that all antibodies have the same stoichiometry. However, it could be possible, that the stoichiometry of trimer neutralization varies between different antibodies, i.e. for one type of antibodies only one antibody is sufficient for trimer neutralization whereas for another sort of antibodies two or three abs could be needed to neutralize one trimer. For this analysis, a statistically sufficient number of experiments for the same combination of viral strain, envelopes mutation, antibody and target cell would be required. Since the data set [2] is not sufficient to analyze single antibody-virus combinations, we divide these combinations into 5 groups according to the antibody binding sites. Antibodies F105, b12, 15e and F91 interfere with the CD4 binding site [20]–[23] and are classified as CD4BS-group. Antibodies 17b and 48d bind to a highly conserved region induced upon CD4 engagement which is important for gp120-chemokine receptor interaction [24],[25] and belong to the CD4i-group. The other three monoclonal antibodies have different binding sites and could not be grouped together. These are 2F5 that binds a linear gp41 epitope proximal to the viral membrane [26], 2G12 that recognizes a carbohydrate-dependent epitope on the gp120 surface [27] and antibody 1121 that recognizes the gp120 V3 loop (ImmunoDiagnostics, Inc.).
We first assume that model parameters for entry derived in [1] are valid for the neutralization assays data from Yang et al. [2]. Under this assumption, we analyze the data either pooling over all antibody-virus combinations, or stratifying with respect to antibody binding site (grouped data). Then we also estimate the parameter of neutralization along with the parameters of entry for the different models, and compare these estimates with the estimates for the neutralization parameter alone.
In this paper, we developed a framework for the estimation of the stoichiometry of HIV neutralization and, as an example how to apply these models, we reanalyze neutralization data [2]. As in our framework for the estimation of the stoichiometry of entry [1], we find that the distribution of trimer numbers is essential for the estimation of the stoichiometry of neutralization. A second major finding is, that the stoichiometry of neutralization may not be estimable if the variation in the number of plasmids that transfect the virus-producer cells in the generation of pseudotyped virions, or the segregation of envelope proteins within the transfected cells are too large. This is due to the fact that, in this case, virions do not express many hetero-trimers, which contain most information on the stoichiometry of neutralization. To ascertain that the experimental procedure is indeed a viable approach for the estimation of the stoichiometry of entry, the variation in transfection and segregation coefficient should be determined.
As defined in our study on HIV entry [1], the measurements for the amount of virions that productively infect a cell is the relative infectivity, RI. In contrast, Yang et al. [2] define the percent neutralization sensitivity, %NS, for studying the stoichiometry of neutralization. The relation between these variables is simply . Yang's model expresses the stoichiometry of entry and the stoichiometry of neutralization as continuous parameters. Since only an integer valued number of trimers can actually bind to the CD4-receptor and 1,2 or 3 monoclonal antibodies can bind to one trimer, and have to be discrete variables, as we modeled them in all models. The most important difference between our models and those of Yang and Klasse [2],[28] is that we include the distribution of trimer numbers. We show, that this distribution, i.e. the frequencies of virions with trimers, is an important input factor which affects the predictions for the relative infectivities and therefore the estimates of the stoichiometric parameters strongly.
The neutralization data of Yang et al. [2] can be analyzed in two ways. Either we use stoichiometric parameters of entry that were independently estimated to estimate the stoichiometry of neutralization from the neutralization data. Or, we attempt to estimate both, the stoichiometric parameters of entry and neutralization from the neutralization data. For the basic, proximity, and soft threshold models, we can infer the stoichiometry of neutralization. The stoichiometry of neutralization cannot be inferred from the imperfect transfection model and the segregation model if we use previous estimates for the parameters of these models. This is due to the lack of hetero-trimers according to the previously estimated parameters. Pooling over all antibodies, the fit to the data for all these models suggest that, on average, one antibody is sufficient for trimer neutralization. We also obtain a stoichiometry of neutralization of one if we stratify the data by antibody binding sites. Using the second approach in which we try to estimate the stoichiometric parameters of neutralization and entry simultaneously from the neutralization data, we find that the estimates of the stoichiometry of neutralization largely agree with those obtained with the first approach. However, the estimates for the entry parameters deviate from the parameters obtained by analyzing entry data only [1]. This is due to the extremely small differences between the predictions of the relative infectivities for some parameter combinations. Hence, we suggest to determine the entry parameters independently from experiments similar to those presented in [2].
As we suggested previously [1], a reliable estimate of the stoichiometric parameters of entry requires elucidating certain aspects of the experimental assays further. We suggest the following line of experiments (in the order of importance):
These experiments allow a deeper insight into the biological processes during transfection and infection of target cells. Some models could be falsified or combined to one model, which explains the infectivity assays best. The entry parameters for this final model should be studied first and then, the stoichiometry of trimer neutralization can be studied.
In the current analysis, experimental data from 4 different virus strains neutralized by monoclonal antibodies with different specificities and mode of action are included. Due to the relatively low number of data points available, stoichiometries of individual antibodies could not be assessed. While the majority of interactions appear to follow a stoichiometry, we can currently not rule out that stoichiometric differences between monoclonal antibodies exist.
The stoichiometry of trimer neutralization, which is the focus of this paper, should not be confused with the single- and multi-hit model parameters proposed by McLain and Dimmock [29]. Their models aimed to determine the number of antibodies required for the neutralization of an entire virion. Since it has been established that virions differ in the number of trimers on their surface [18],[30] we cannot expect to describe virion neutralization with a single parameter. More recently, stoichiometric parameters have been proposed for the quantitative description of HIV entry and neutralization [2],[11],[14],[31]. In particular, HIV neutralization is currently described by the number of trimers on the virion's surface, the stoichiometry of entry and trimer neutralization. In combination with the stoichiometry of entry, the stoichiometry of trimer neutralization can be used to estimate the mean number of antibodies that neutralize a single virion. The number of trimers per virion varies and so does the number of trimers, which have to be neutralized. As an example, let us assume that the basic model is valid and that the stoichiometry of entry is 8, i.e. 8 functional trimers are needed to infect a target cell. Imagine a virion with 10 trimers. If two of them are neutralized, this virion is still infectious. Neutralizing one more trimer renders the virion non-infectious. In total, at least 3 trimers have to be neutralized for neutralizing the virion. Assuming that one antibody is able to neutralize one trimer, i.e. , at least 3 antibodies are needed to neutralize the virion with 10 trimers. However, 3 antibodies may not be sufficient. Imagine, for example, that the 3 antibodies bind to the 3 envelope proteins of the same trimer, then only one trimer is neutralized. While we need at least 3 antibodies to neutralize this virion, 7 antibodies will be sufficient. Still assuming and , a virion with 30 trimers would be neutralized if at least 23 trimers loose their functionality. Therefore at least 23 but not more than 67 antibodies are needed. We can estimate the mean number of monoclonal antibodies required to neutralize virions with 10 trimers as and virions with 30 trimers as (these estimates are based on simulations with virtual virions). We plan to study how the stoichiometries of entry and neutralization relate to the neutralization of a population of viruses in the future.
The models we present here address the question of how many monoclonal antibodies are needed to neutralize a single trimer in vitro. In vivo however, there will always be a mixture of different monoclonal antibodies attacking the virions. To predict the effect of a polyclonal antibody response on virus replication, it will be necessary, in addition to estimating the stoichiometries for each antibody clone, to investigate how they synergize or antagonize each other. To illustrate what exactly we mean by synergy and antagonism assume, for example, we have two antibody clones, A and B with stoichiometries of neutralization of , and , respectively. If one antibody A is bound to a trimer already, how many antibodies B are then required for neutralization? Further, does it matter where the B antibodies bind, i.e. whether they bind to the same envelope protein as antibody A or to a different one? To assess if the antibodies synergize or antagonize in this sense, one can perform experiments using pseudotyped viruses with mixed envelope proteins (very similar to those that have been conducted to estimate the stoichiometry of neutralization) in combination with our mathematical models. For this particular question, one should mix envelope proteins resistant to neutralization by antibody A with envelope proteins resistant to neutralization by antibody B. The relative infectivity of these pseudotyped viruses has to be measured under saturation of antibody A and B. If antibody A and B synergize, the relative infectivity in this experiment will be lower than if the antibodies act independently. This is because the only trimer that is not neutralized is one consisting of two envelope proteins with A-resistance and one with B-resistance. If antibodies A and B act independently these trimers are not neutralized because they bind fewer than necessary numbers of A and B. How the understanding of the stoichiometry of neutralization by a mixture of antibodies scales up to the level of the entire virion depends strongly on whether the antibody binding sites overlap. If they do, the number of antibodies required for neutralization will be lower than in the case of non-overlapping antibody binding sites because there is less opportunity for antibodies to bind uselessly.
We presented a modeling framework which enables us to investigate the number of antibodies that are needed to neutralize a single trimer and if this quantity varies between different antibodies. As the stoichiometry of trimer neutralization is the basis for the calculation of the stoichiometry of virion and population neutralization, it is an important parameter for the quantitative understanding of the protection antibodies may confer.
|
10.1371/journal.pcbi.1003311 | Stochastic Computations in Cortical Microcircuit Models | Experimental data from neuroscience suggest that a substantial amount of knowledge is stored in the brain in the form of probability distributions over network states and trajectories of network states. We provide a theoretical foundation for this hypothesis by showing that even very detailed models for cortical microcircuits, with data-based diverse nonlinear neurons and synapses, have a stationary distribution of network states and trajectories of network states to which they converge exponentially fast from any initial state. We demonstrate that this convergence holds in spite of the non-reversibility of the stochastic dynamics of cortical microcircuits. We further show that, in the presence of background network oscillations, separate stationary distributions emerge for different phases of the oscillation, in accordance with experimentally reported phase-specific codes. We complement these theoretical results by computer simulations that investigate resulting computation times for typical probabilistic inference tasks on these internally stored distributions, such as marginalization or marginal maximum-a-posteriori estimation. Furthermore, we show that the inherent stochastic dynamics of generic cortical microcircuits enables them to quickly generate approximate solutions to difficult constraint satisfaction problems, where stored knowledge and current inputs jointly constrain possible solutions. This provides a powerful new computing paradigm for networks of spiking neurons, that also throws new light on how networks of neurons in the brain could carry out complex computational tasks such as prediction, imagination, memory recall and problem solving.
| The brain has not only the capability to process sensory input, but it can also produce predictions, imaginations, and solve problems that combine learned knowledge with information about a new scenario. But although these more complex information processing capabilities lie at the heart of human intelligence, we still do not know how they are organized and implemented in the brain. Numerous studies in cognitive science and neuroscience conclude that many of these processes involve probabilistic inference. This suggests that neuronal circuits in the brain process information in the form of probability distributions, but we are missing insight into how complex distributions could be represented and stored in large and diverse networks of neurons in the brain. We prove in this article that realistic cortical microcircuit models can store complex probabilistic knowledge by embodying probability distributions in their inherent stochastic dynamics – yielding a knowledge representation in which typical probabilistic inference problems such as marginalization become straightforward readout tasks. We show that in cortical microcircuit models such computations can be performed satisfactorily within a few . Furthermore, we demonstrate how internally stored distributions can be programmed in a simple manner to endow a neural circuit with powerful problem solving capabilities.
| The question whether brain computations are inherently deterministic or inherently stochastic is obviously of fundamental importance. Numerous experimental data highlight inherently stochastic aspects of neurons, synapses and networks of neurons on virtually all spatial and temporal scales that have been examined [1]–[5]. A clearly visible stochastic feature of brain activity is the trial-to-trial variability of neuronal responses, which also appears on virtually every spatial and temporal scale that has been examined [2]. This variability has often been interpreted as side-effect of an implementation of inherently deterministic computing paradigms with noisy elements, and it has been attempted to show that the observed noise can be eliminated through spatial or temporal averaging. However, more recent experimental methods, which make it possible to record simultaneously from many neurons (or from many voxels in fMRI), have shown that the underlying probability distributions of network states during spontaneous activity are highly structured and multimodal, with distinct modes that resemble those encountered during active processing. This has been shown through recordings with voltage-sensitive dyes starting with [6], [7], multi-electrode arrays [8], and fMRI [9], [10]. It was also shown that the intrinsic trial-to-trial variability of brain systems is intimately related to the observed trial-to-trial variability in behavior (see e.g. [11]). Furthermore, in [12] it was shown that during navigation in a complex environment where simultaneously two spatial frames of reference were relevant, the firing of neurons in area CA1 represented both frames in alternation, so that coactive neurons tended to relate to a common frame of reference. In addition it has been shown that in a situation where sensory stimuli are ambiguous, large brain networks switch stochastically between alternative interpretations or percepts, see [13]–[15]. Furthermore, an increase in the volatility of network states has been shown to accompany episodes of behavioral uncertainty [16]. All these experimental data point to inherently stochastic aspects in the organization of brain computations, and more specifically to an important computational role of spontaneously varying network states of smaller and larger networks of neurons in the brain. However, one should realize that the approach to stochastic computation that we examine in this article does not postulate that all brain activity is stochastic or unreliable, since reliable neural responses can be represented by probabilities close to 1.
The goal of this article is to provide a theoretical foundation for understanding stochastic computations in networks of neurons in the brain, in particular also for the generation of structured spontaneous activity. To this end, we prove here that even biologically realistic models for networks of neurons in the brain have – for a suitable definition of network state – a unique stationary distribution of network states. Previous work had focused in this context on neuronal models with linear sub-threshold dynamics [17], [18] and constant external input (e.g. constant input firing rates). However, we show here that this holds even for quite realistic models that reflect, for example, data on nonlinear dendritic integration (dendritic spikes), synapses with data-based short term dynamics (i.e., individual mixtures of depression and facilitation), and different types of neurons on specific laminae. We also show that these results are not restricted to the case of constant external input, but rather can be extended to periodically changing input, and to input generated by arbitrary ergodic stochastic processes.
Our theoretical results imply that virtually any data-based model , for networks of neurons featuring realistic neuronal noise sources (e.g. stochastic synaptic vesicle release) implements a Markov process through its stochastic dynamics. This can be interpreted – in spite of its non-reversibility – as a form of sampling from a unique stationary distribution . One interpretation of , which is in principle consistent with our findings, is that it represents the posterior distribution of a Bayesian inference operation [19]–[22], in which the current input (evidence) is combined with prior knowledge encoded in network parameters such as synaptic weights or intrinsic excitabilities of neurons (see [23]–[26] for an introduction to the “Bayesian brain”). This interpretation of neural dynamics as sampling from a posterior distribution is intriguing, as it implies that various results of probabilistic inference could then be easily obtained by a simple readout mechanism: For example, posterior marginal probabilities can be estimated (approximately) by observing the number of spikes of specific neurons within some time window (see related data from parietal cortex [27]). Furthermore, an approximate maximal a posteriori (MAP) inference can be carried out by observing which network states occur more often, and/or are more persistent.
A crucial issue which arises is whether reliable readouts from in realistic cortical microcircuit models can be obtained quickly enough to support, e.g., fast decision making in downstream areas. This critically depends on the speed of convergence of the distribution of network states (or distribution of trajectories of network states) from typical initial network states to the stationary distribution. Since the initial network state of a cortical microcircuit depends on past activity, it may often be already quite “close” to the stationary distribution when a new input arrives (since past inputs are likely related to the new input). But it is also reasonable to assume that the initial state of the network is frequently unrelated to the stationary distribution , for example after drastic input changes. In this case the time required for readouts depends on the expected convergence speed to from – more or less – arbitrary initial states. We show that one can prove exponential upper bounds for this convergence speed. But even that does not guarantee fast convergence for a concrete system, because of constant factors in the theoretical upper bound. Therefore we complement this theoretical analysis of the convergence speed by extensive computer simulations for cortical microcircuit models.
The notion of a cortical microcircuit arose from the observation that “it seems likely that there is a basically uniform microcircuit pattern throughout the neocortex upon which certain specializations unique to this or that cortical area are superimposed” [28]. This notion is not precisely defined, but rather a term of convenience: It refers to network models that are sufficiently large to contain examples of the main types of experimentally observed neurons on specific laminae, and the main types of experimentally observed synaptic connections between different types of neurons on different laminae, ideally in statistically representative numbers [29]. Computer simulations of cortical microcircuit models are practically constrained both by a lack of sufficiently many consistent data from a single preparation and a single cortical area, and by the available computer time. In the computer simulations for this article we have focused on a relatively simple standard model for a cortical microcircuit in the somatosensory cortex [30] that has already been examined in some variations in previous studies from various perspectives [31]–[34].
We show that for this standard model of a cortical microcircuit marginal probabilities for single random variables (neurons) can be estimated through sampling even for fairly large instances with 5000 neurons within a few of simulated biological time, hence well within the range of experimentally observed computation times of biological organisms. The same holds for probabilities of network states for small sub-networks. Furthermore, we show that at least for sizes up to 5000 neurons these “computation times” are virtually independent of the size of the microcircuit model.
We also address the question to which extent our theoretical framework can be applied in the context of periodic input, for example in the presence of background theta oscillations [35]. In contrast to the stationary input case, we show that the presence of periodic input leads to the emergence of unique phase-specific stationary distributions, i.e., a separate unique stationary distribution for each phase of the periodic input. We discuss basic implications of this result and relate our findings to experimental data on theta-paced path sequences [35], [36] and bi-stable activity [37] in hippocampus.
Finally, our theoretically founded framework for stochastic computations in networks of spiking neurons also throws new light on the question how complex constraint satisfaction problems could be solved by cortical microcircuits [38], [39]. We demonstrate this in a toy example for the popular puzzle game Sudoku. We show that the constraints of this problem can be easily encoded by synaptic connections between excitatory and inhibitory neurons in such a way that the stationary distribution assigns particularly high probability to those network states which encode correct (or good approximate) solutions to the problem. The resulting network dynamics can also be understood as parallel stochastic search with anytime computing properties: Early network states provide very fast heuristic solutions, while later network states are distributed according to the stationary distribution , therefore visiting with highest probability those solutions which violate only a few or zero constraints.
In order to make the results of this article accessible to non-theoreticians we present in the subsequent Results section our main findings in a less technical formulation that emphasizes relationships to experimental data. Rigorous mathematical definitions and proofs can be found in the Methods section, which has been structured in the same way as the Results section in order to facilitate simultaneous access on different levels of detail.
A simple notion of network state at time simply indicates which neurons in the network fired within some short time window before . For example, in [20] a window size of 2ms was selected. However, the full network state could not be analyzed there experimentally, only its projection onto 16 electrodes in area V1 from which recordings were made. An important methodological innovation of [20] was to analyze under various conditions the probability distribution of the recorded fragments of network states, i.e., of the resulting bit vectors of length 16 (with a “1” at position if a spike was recorded during the preceding 2ms at electrode ). In particular, it was shown that during development the distribution over these network states during spontaneous activity in darkness approximates the distribution recorded during natural vision. Apart from its functional interpretation, this result also raises the even more fundamental question how a network of neurons in the brain can represent and generate a complex distribution of network states. This question is addressed here in the context of data-based models for cortical microcircuits. We consider notions of network states similar to [20] (see the simple state in Figure 1C) and provide a rigorous proof that under some mild assumptions any such model represents and generates for different external inputs associated different internal distributions of network states . More precisely, we will show that for any specific input there exists a unique stationary distribution of network states to which the network converges exponentially fast from any initial state.
This result can be derived within the theory of Markov processes on general state spaces, an extension of the more familiar theory of Markov chains on finite state spaces to continuous time and infinitely many network states. Another important difference to typical Markov chains (e.g. the dynamics of Gibbs sampling in Boltzmann machines) is that the Markov processes describing the stochastic dynamics of cortical microcircuit models are non-reversible. This is a well-known difference between simple neural network models and networks of spiking neurons in the brain, where a spike of a neuron causes postsynaptic potentials in other neurons - but not vice versa. In addition, experimental results show that brain networks tend to have a non-reversible dynamics also on longer time scales (e.g., stereotypical trajectories of network states [40]–[43]).
In order to prove results on the existence of stationary distributions of network states , one first needs to consider a more complex notion of network state at time , which records the history of all spikes in the network since time (see Figure 1C). The window length has to be chosen sufficiently large so that the influence of spikes before time on the dynamics of the network after time can be neglected. This more complex notion of network state then fulfills the Markov property, such that the future network evolution depends on the past only through the current Markov state. The existence of a window length with the Markov property is a basic assumption of the subsequent theoretical results. For standard models of networks of spiking neurons a value of around 100ms provides already a good approximation of the Markov property, since this is a typical time during which a post-synaptic potential has a non-negligible effect at the soma of a post-synaptic neuron. For more complex models of networks of spiking neurons a larger value of in the range of seconds is more adequate, in order to accommodate for dendritic spikes or the activation of receptors that may last 100ms or longer, and the short term dynamics of synapses with time constants of several hundred milliseconds. Fortunately, once the existence of a stationary distribution is proved for such more complex notion of network state, it also holds for any simpler notion of network state (even if these simpler network states do not fulfill the Markov property), that results when one ignores details of the more complex network states. For example, one can ignore all spikes before time , the exact firing times within the window from to , and whether a neuron fired one or several spikes. In this way one arrives back at the simple notion of network state from [20].
Theorem 1 also applies to networks which generate stereotypical trajectories of network activity [41]. For such networks it may be of interest to consider not only the distribution of network states in a short window (e.g. simple states with , or ), but also the distribution of longer trajectories produced by the network. Indeed, since Theorem 1 holds for Markov states with any fixed window length , it also holds for values of that are in the range of experimentally observed trajectories of network states [41], [49], [50]. Hence, a generic neural circuit automatically has a unique stationary distribution over trajectories of (simple) network states for any fixed trajectory length . Note that this implies that a neural circuit has simultaneously stationary distributions of trajectories of (simple) network states of various lengths for arbitrarily large , and a stationary distribution of simple network states. This fact is not surprising if one takes into consideration that if a circuit has a stationary distribution over simple network states this does not imply that subsequent simple network states represent independent drawings from this stationary distribution. Hence the circuit may very well produce stereotypical trajectories of simple network states. This feature becomes even more prominent if the underlying dynamics (the Markov process) of the neural circuit is non-reversible on several time scales.
We address two basic types of knowledge extraction from a stationary distribution of a network : the computation of marginal probabilities and maximal a posteriori (MAP) assignments. Both computations constitute basic inference problems commonly appearing in real-world applications [51], which are in general difficult to solve as they involve large sums, integrals, or maximization steps over a state space which grows exponentially in the number of random variables. However, already [21], [25] noted that the estimation of marginal probabilities would become straightforward if distributions were represented in the brain in a sample-based manner (such that each network state at time represents one sample from the distribution). Theorem 1 provides a theoretical foundation for how such a representation could emerge in realistic data-based microcircuit models on the implementation level: Once the network has converged to its stationary distribution, the network state at any time represents a sample from (although subsequent samples are generally not independent). Simultaneously, the subnetwork state of any subset of neurons represents a sample from the marginal distribution . This is particularly relevant if one interprets in a given cortical microcircuit as the posterior distribution of an implicit generative model, as suggested for example by [20] or [21], [22].
In order to place the estimation of marginals into a biologically relevant context, assume that a particular component of the network state has a behavioral relevance. This variable , represented by some neuron , could represent for example the perception of a particular visual object (if neuron is located in inferior temporal cortex [52]), or the intention to make a saccade into a specific part of the visual field (if neuron is located in area LIP [53]). Then the computation of the marginal(1)would be of behavioral significance. Note that this computation integrates information from the internally stored knowledge with evidence about a current situation . In general this computation is demanding as it involves a sum with exponentially many terms in the network size .
But according to Theorem 1, the correct marginal distribution is automatically embodied by the activity of neuron . Hence the marginal probability can be estimated by simply observing what fraction of time the neuron spends in the state , while ignoring the activity of the remaining network [21]. In principle, a downstream neuron could gather this information by integrating the spike output of over time.
Marginal probabilities of subpopulations, for example , can be estimated in a similar manner by keeping track of how much time the subnetwork spends in the state (1,0,1), while ignoring the activity of the remaining neurons. A downstream network could gather this information, for example, by integrating over the output of a readout neuron which is tuned to detect the desired target pattern (1,0,1).
Notably, the estimation of marginals sketched above is guaranteed by ergodic theory to converge to the correct probability as observation time increases (due to Theorem 1 which ensures that the network is an ergodic Markov process, see Methods). In particular, this holds true even for networks with prominent sequential dynamics featuring, for example, stereotypical trajectories. However, note that the observation time required to obtain an accurate estimate may be longer when trajectories are present since subsequent samples gathered from such a network will likely exhibit stronger dependencies than in networks lacking sequential activity patterns. In a practical readout implementation where recent events might be weighed preferentially this could result in more noisy estimates.
Approximate maximal a posteriori (MAP) assignments to small subsets of variables can also be obtained in a quite straightforward manner. For given external inputs , the marginal MAP assignment to the subset of variables (with some ) is defined as the set of values that maximize(2)
A sample-based approximation of this operation can be implemented by keeping track of which network states in the subnetwork occur most often. This could, for example, be realized by a readout network in a two stage process: first the marginal probabilities of all subnetwork states are estimated (by 8 readout neurons dedicated to that purpose), followed by the selection of the neuron with maximal probability. The selection of the maximum could be achieved in a neural network, for example, through competitive inhibition. Such competitive inhibition would ideally lead to a winner-take-all function such that the neuron with the strongest stimulation (representing the variable assignment with the largest probability) dominates and suppresses all other readout neurons.
Whereas many types of computations (for example probabilistic inference via the junction tree algorithm [51]) require a certain computation time, probabilistic inference via sampling from an embodied distribution belongs to the class of anytime computing methods, where rough estimates of the result of a computation become almost immediately available, and are automatically improved when there is more time for a decision. A main component of the convergence time to a reliable result arises from the time which the distribution of network states needs to become independent of its initial state . It is well known that both, network states of neurons in the cortex [54] and quick decisions of an organism, are influenced for a short time by this initial state (and this temporary dependence on the initial state may in fact have some behavioral advantage, since may contain information about preceding network inputs, expectations, etc.). But it has remained unknown, what range of convergence speeds for inference from is produced by common models for cortical microcircuits .
We address this question by analyzing the convergence speed of stochastic computations in the cortical microcircuit model of [30]. A typical network response of an instance of the cortical microcircuit model comprising 560 neurons as in [30] is shown in Figure 2A. We first checked how fast marginal probabilities for single neurons converge to stationary values from different initial network Markov states. We applied the same analysis as in Figure 1H to the simple state () of a single representative neuron from layer 5. Figure 2B shows quite fast convergence of the “on”-state probability of the neuron to its stationary value from two different initial states. Note that this straightforward method of checking convergence is rather inefficient, as it requires the repetition of a large number of trials for each initial state. In addition it is not suitable for analyzing convergence to marginals for subpopulations of neurons (see Figure 2G).
Various more efficient convergence diagnostics have been proposed in the context of discrete-time Markov Chain Monte Carlo theory [55]–[58]. In the following, we have adopted the Gelman and Rubin diagnostic, one of the standard methods in applications of MCMC sampling [55]. The Gelman Rubin convergence diagnostic is based on the comparison of many runs of a Markov chain when started from different randomly drawn initial states. In particular, one compares the typical variance of state distributions during the time interval within a single run (within-variance) to the variance during the interval between different runs (between-variance). When the ratio of between- and within-variance approaches 1 this is indicative of convergence. A comparison of panels B and C of Figure 2 shows that in the case of marginals for single neurons this interpretation fits very well to the empirically observed convergence speed for two different initial conditions. Various values between 1.02 [58] and 1.2 [57], [59], [60] have been proposed in the literature as thresholds below which the ratio signals that convergence has taken place. The shaded region in Figure 2C–G corresponds to values below a threshold of 1.1. An obvious advantage of the Gelman-Rubin diagnostic, compared with a straightforward empirical evaluation of convergence properties as in Figure 2B, is its substantially larger computational efficiency and the larger number of initial states that it takes into account. For the case of multivariate marginals (see Figure 2G), a straightforward empirical evaluation of convergence is not even feasible, since relative frequencies of states would have to be analyzed.
Using the Gelman-Rubin diagnostic, we estimated convergence speed for marginals of single neurons (see Figure 2C, mean/worst in Figure 2E), and for the product of the simple states of two neurons (i.e., pairwise spike coincidences) in Figure 2D. We found that in all cases the Gelman-Rubin value drops close to 1 within just a few . More precisely, for a typical threshold of convergence times are slightly below in Figure 2C–E. A very conservative threshold of yields convergence times close to .
The above simulations were performed in a circuit of 560 neurons, but eventually one is interested in the properties of much larger circuits. Hence, a crucial question is how the convergence properties scale with the network size. To this end, we compared convergence in the cortical microcircuit model of [30] for four different sizes (500, 1000, 2000 and 5000). To ensure that overall activity characteristics are maintained across different sizes, we adopted the approach of [30] and scaled recurrent postsynaptic potential (PSP) amplitudes inversely proportional to network size. A comparison of mean (solid line) and worst (dashed line) marginal convergence for networks of different sizes is shown in Figure 2F. Notably we find that the network size has virtually no effect on convergence speed. This suggests that, at least within the scope of the laminar microcircuit model of [30], even very large cortical networks may support fast extraction of knowledge (in particular marginals) from their stationary distributions .
In order to estimate the required computation time associated with the estimation of marginal probabilities and MAP solutions on small subpopulations , one needs to know how fast the marginal probabilities of vector-valued states of subnetworks of become independent from the initial state of the network. To estimate convergence speed in small subnetworks, we applied a multivariate version of the Gelman-Rubin method to vector-valued simple states of subnetworks (Figure 2G, dotted lines, evaluated for varying circuit sizes from 500 to 5000 neurons). We find that multivariate convergence of state frequencies for a population of neurons is only slightly slower than for uni-variate marginals. To complement this analysis, we also investigated convergence properties of a “random readout” neuron which integrates inputs from many neurons in a subnetwork. It is interesting to note that the convergence speed of such a readout neuron, which receives randomized connections from a randomly chosen subset of 500 neurons, is comparable to that of single marginals (Figure 2F, solid lines), and in fact slightly faster.
An interesting research question is which dynamic or structural properties of a cortical microcircuit model have a strong impact on its convergence speed to the stationary distribution . Unfortunately, a comprehensive treatment of this question is beyond the scope of this paper, since virtually any aspect of circuit dynamics could be investigated in this context. Even if one focuses on a single aspect, the impact of one circuit feature is likely to depend on the presence of other features (and probably also on the properties of the input). Nonetheless, to lay a foundation for further investigation, first empirical results are given in Figure 3.
As a reference point, Figure 3A shows a typical activity pattern and convergence speed of single marginals in the small cortical microcircuit model from Figure 1. To test whether the overall activity of a network has an obvious impact on convergence speed, we constructed a small network of 20 neurons (10 excitatory, 10 inhibitory) and tuned connection weights to achieve sparse overall activity (Figure 3B). A comparison of panels A and B suggests that overall network activity has no significant impact on convergence speed. To test whether the presence of stereotypical trajectories of network states (similar to [41]) has a noticeable influence on convergence, we constructed a small network exhibiting strong sequential activity patterns (see Figure 3C). We find that convergence speed is hardly affected, except for the first (see Figure 3C). Within the scope of this first empirical investigation, we were only able to produce a significant slow-down of the convergence speed by building a network that alternated between two attractors (Figure 3D).
In Theorem 1 we had already addressed one important case where the network receives dynamic external inputs: the case when external input is generated by some Markov process. But many networks of neurons in the brain are also subject to more or less pronounced periodic inputs (“brain rhythms” [61]–[63]), and it is known that these interact with knowledge represented in distributions of network states in specific ways. For instance, it had been shown in [35] that the phase of the firing of place cells in the hippocampus of rats relative to an underlying theta-rhythm is related to the expected time when the corresponding location will be reached. Inhibitory neurons in hippocampus have also been reported to fire preferentially at specific phases of the theta cycle (see e.g. Figure S5 in [12]). Moreover it was shown that different items that are held in working memory are preferentially encoded by neurons that fire at different phases of an underlying gamma-oscillation in the monkey prefrontal cortex [64] (see [65] for further evidence that such oscillations are behaviorally relevant). Phase coding was also reported in superior temporal sulcus during category representation [66]. The following result provides a theoretical foundation for such phase-specific encoding of knowledge within a framework of stochastic computation in networks of spiking neurons.
Whenever an inhibitory neuron fires, it reduces for a short while the probability of firing for its postsynaptic targets. In fact, new experimental data [71] show that inhibitory neurons impose quite powerful constraints on pyramidal cells. But also how pyramidal cells are embedded into their network environment imposes constraints on local network activity. From this perspective, the resulting firing patterns of a cortical microcircuit can be viewed as stochastically generated solutions of an immensely complex constraint satisfaction problem, that is defined both by external inputs to the circuit and by the way each excitatory and inhibitory neuron is embedded into its circuit environment. Constraint satisfaction problems are from the computational perspective a particularly interesting class of problems, because many tasks that a brain has to solve, from the generation of a percept from unreliable and ambiguous sources to higher level tasks such as memory recall, prediction, planning, problem solving, and imagination, can be formulated as constraint satisfaction problems [72]. However, numerous constraint satisfaction problems are known to be NP-hard, thereby limiting the applicability of exact solution strategies. Instead, approximate or heuristic algorithms are commonly used in practice (for example evolutionary algorithms [73]). Here we propose that networks of spiking neurons with noise have an inherent capability to solve constraint satisfaction problems in an approximate (heuristic) manner through their stochastic dynamics. The key principle is that those network states , which satisfy the largest number of local constraints, have the highest probability under the distribution . These constraints are imposed by the way each neuron of is embedded into the circuit, and the current external input which can selectively activate or deactivate specific constraints.
We have selected a specific constraint satisfaction problem for demonstrating the capability of networks of spiking neurons to generate rapidly approximate solutions to constraint satisfaction problems through their inherent stochastic dynamics: solving Sudoku puzzles (see Figure 5A). Sudoku is a well-suited example because it is complex enough to be representative for many problem solving tasks, and lends itself well to visual interpretation and presentation (but note that we do not aim to model here how humans solve Sudoku puzzles). The rules of the Sudoku game can be easily embedded into common models for cortical microcircuits as recurrent networks of Winner-Take-All (WTA) microcircuit motifs [29]. Each WTA motif is an ensemble of pyramidal cells (on layers 2/3 or 5/6) that are subject to lateral inhibition (see Figure 5B). Each pyramidal cell can in fact be part of several interlocking WTA motifs (Figure 5B, right).
This architecture makes it easy to impose the interlocking constraints of Sudoku (and of many other constraint satisfaction problems). Each pyramidal cell (or each local group of pyramidal cells) votes for placing a particular digit into an empty field of the grid, that is not dictated by the external input . But this pyramidal cell is subject to the constraints that only one digit can be placed into this field, and that each digit occurs only once in each column, in each row, and in each 3×3 sub-grid. Hence each pyramidal cell is simultaneously part of four inhibitory subnetworks (WTA motifs).
A specific puzzle can be entered by providing strong input to those neurons which represent the given numbers in a Sudoku (Figure 5A, left). This initiates a quite intuitive dynamics: “Clamped” neurons start firing strongly, and as a consequence, neurons which code for conflicting digits in the same Sudoku field, the same row, column or 3×3 sub-grid, become strongly inhibited through di-synaptic inhibition. In many Sudoku fields this will lead to the inhibition of a large number of otherwise freely competing neurons, thereby greatly reducing the space of configurations generated by the network. In some cases, inhibition will immediately quieten all neurons except those associated with a single remaining digit (the only choice consistent with the givens). In the absence of competition, these uninhibited neurons will start firing along with the givens, thereby further constraining neighboring neurons. This form of inhibitory interaction therefore implicitly implements a standard strategy for solving easy Sudokus: checking for fields in which only one possibility remains. In harder Sudokus, however, this simple strategy alone would be typically insufficient, for example when several possibilities remain in all fields. In such cases, where inhibition leaves more than one possible digit open, a tentative digit will be automatically picked randomly by those neurons which happen to fire first among its competitors. This ensures that, instead of getting stuck, the network automatically explores potential configurations in situations where multiple possibilities remain. Altogether, through this combination of constraint enforcement and random exploration, those network states which violate few constraints (good approximate solutions) are visited with much higher probability than states with conflicting configurations. Hence, most time is spent in good approximate solutions. Furthermore, from all Sudoku configurations the solving configuration is visited in this process especially often.
Figure 5C shows a typical network run during the last seconds (out of a total simulation time of approximately ) before the correct solution was found to the Sudoku puzzle from Figure 5A. For this simulation we modeled lateral inhibition in each WTA motif by reciprocally connecting each neuron in the subnetwork to a single inhibitory neuron. For each of the 9 digits in a Sudoku field, we created an associated local group of four pyramidal cells. This can be seen in Figure 5C, where spike responses of pyramidal cells associated with three different Sudoku fields are shown (the three colored fields in Figure 5A and B). Each field has 9 possible digits, and each digit has four associated neurons. Hence, for each of the three Sudoku fields (WTA motifs), neurons are shown. Spikes are colored black for those neurons which code for a wrong digit, and green for the four neurons which code for the correct digit in a Sudoku field (the correct digits in Figure 5C are 6, 8 and 4). The overall performance of the network (fraction of correctly solved fields) during the last 1.5 seconds before the solution is found is shown in Figure 5C above.
In our simulations we found that the solve time (the time until the correct solution is found for the first time) generally depends on the hardness of the Sudoku, in particular on the number of givens. For the “hard” Sudoku with 26 givens from Figure 5A, solve times are approximately exponentially distributed at an average of 29 seconds (Figure 5D). The average performance during the first five seconds of a run (obtained from 1000 independent runs) is shown in Figure 5E. The plot shows quick convergence to a (stationary) average performance of approximately 0.9. This demonstrates that the network spends on average most time in approximate solutions with high performance. Among these high-performance solutions, the correct solution occurs especially often (on average 2% of the time).
We have shown that for common noise models in cortical microcircuits, even circuits with very detailed and diverse non-linear neurons and synapses converge exponentially fast to a stationary distribution of network states . This holds both for external inputs that consist of Poisson spike trains of a fixed rate, and for the case where is periodic, or generated by some Markov process with a stationary distribution. The same mathematical framework also guarantees exponentially fast convergence to a stationary distribution of trajectories of network states (of any fixed time length), thereby providing a theoretical foundation for understanding stochastic computations with experimentally observed stereotypical trajectories of network states. These results extend and generalize previous work in [17] and [18] in two ways. First, previous convergence proofs had been given only for networks of simplified neurons in which the (sub-threshold) neuronal integration of pre-synaptic spikes was assumed a linear process, thereby excluding the potential effects of dendritic non-linearities or synaptic short-term dynamics. Second, previous work had focused only on the case where input is provided by neurons with fixed firing rates (a special case of Theorem 1). In addition we show that these convergence proofs can be derived from a fundamental property of stochastic spiking networks, that we have formulated as the Contraction Lemma (Lemma 1 in Methods).
The stationary distribution provides an attractive target for investigating the stochastic computing capabilities of data-based models for local circuits or larger networks of neurons in the brain. In contrast to the much simpler case of Boltzmann machines with non-spiking linear neurons and symmetric synaptic connections, it is unlikely that one can attain for cortical microcircuit models a simple analytical description of . But our computer simulations have shown that this is not necessarily an obstacle for encoding salient constraints for problem solving in , and for merging knowledge that is encoded in with online information from external inputs in quite fast stochastic computations. In fact, the resulting paradigm for computations in cortical microcircuits supports anytime computing, where one has no fixed computation time. Instead, first estimates of computational results can be produced almost immediately, and can be rapidly communicated to other circuits. In this way, no processor (circuit) has to idle until other processors have completed their subcomputations, thereby avoiding the arguably most critical general bottleneck of massively parallel computing systems. Instead, each microcircuit can contribute continuously to an iterative refinement of a global computation.
Our computer simulations for a standard cortical microcircuit model suggest that convergence to is fast enough to support knowledge extraction from this distribution within a few , i.e. within the typical computation time of higher-level brain computations. These first estimates need to be corroborated by further theoretical work and computer simulations. In particular, the relationship between the structure and dynamics of cortical microcircuits and their convergence speed merits further investigation. Furthermore, in the case where is a multi-modal distribution there exists an obvious tradeoff between the convergence speed to and the typical duration of staying in an “attractor” (i.e., a region of the state space which has high probability under ). Staying longer in an attractor obviously facilitates the readout of the result of a computation by downstream networks. A number of experimental data suggest that neuromodulators can move neural circuits (at least in the prefrontal cortex) to different points on this tradeoff curve. For example it is argued in [74], [75] that the activation of receptors through dopamine deepens all basins of attraction, making it harder for the network state to leave an attractor. Additional molecular mechanisms that shift the tradeoff between fast sampling (exploration) and the temporal stability of found solutions are reviewed in [76]. Another interesting perspective on convergence speed is that slow convergence may be beneficial for certain computations in specific brain areas (especially early sensory areas). Slow convergence enlarges the time span during which the network can integrate information from non-stationary external inputs [77]–[79]. In addition the initial state of a network may contain information about preceding events that are computationally useful. Those considerations suggest that there exist systematic differences between the convergence speed to in different neural systems , and that it can be modulated in at least some systems dependent on the type of computational task that needs to be solved.
Another important issue is the tradeoff between sampling time and sampling accuracy. In high-level cognitive tasks, for example, it has been argued that “approximate and quick” sample-based decisions are often better than “accurate but slow” decisions [80], [81]. Of particular interest in this context is the analysis of [81] who studied the time-accuracy tradeoff during decision making, under the assumption that the mind performs inference akin to MCMC sampling. Due to the nature of MCMC sampling, early samples before convergence (during the burn-in period) are biased towards the initial state of the system. In the absence of time pressure, the optimal strategy is therefore to wait and collect samples for a long period of time (in theory indefinitely). In the presence of even moderate time costs, however, the optimal sampling time can be shown to be finite, a result which can provide a rational explanation of the anchoring effect in cognitive science [81] (under time pressure people's decisions are influenced by their “initial state”). Notably, the analysis of [81] was based on the assumption that the MCMC algorithm exhibits geometric convergence, the discrete-time equivalent to the exponential convergence speed proved in this paper for stochastic spiking networks. Applying a similar analysis to study optimal time-accuracy tradeoff points in cortical microcircuits therefore presents a promising avenue for future research.
It had been shown in [21] and [22] that, under certain assumptions on the neuron models and circuit structure, in principle every joint distribution over discrete-valued random variables can be represented as a stationary distribution of some network of spiking neurons. Forthcoming unpublished results suggest that such internal representations of a given distribution can even be learned from examples drawn from . This will provide a first step towards understanding how the stationary distribution of a microcircuit can be adapted through various plasticity processes to encode salient constraints, successful solution strategies (rules), and other types of knowledge. This research direction promises to become especially interesting if one takes into account that knowledge can not only be encoded in the stationary distribution of network states, but also in the simultaneously existing stationary distribution of trajectories of network states.
Attractor neural networks [82] were originally deterministic computational models, where gradient descent leads the network from some given initial state (the input for the computation) to the lowest point of the attractor (the output of the computation) in whose basis of attraction lies. The computational capability of an attractor neural network is substantially larger if its attractor landscape can be reconfigured on the fly by external input , as in [83] and in the Sudoku example of this article. This usually requires that the attractors are not programmed directly into the network parameters, but emerge from some more general computational principles (e.g. constraint satisfaction). Attractor neural networks gain additional computational capability if there is some noise in the system [84]. This enables the network to leave after a while suboptimal solutions [85]. Alternative modeling frameworks for the transient dynamics of neural systems are provided by the liquid computing model [77], and on a more abstract level by sequences of metastable states in dynamical systems [86]. Here we propose to view both transient and attractor dynamics of complex data-based circuits from the perspective of probabilistic inference, in particular as neural sampling [21] (or more abstractly: as MCMC sampling) from their inherent probability distribution over network states (or trajectories of network states), that serves as the knowledge base of these neural systems.
We had focused in our computer simulations on the investigation of the stationary distribution for models of cortical microcircuits. But the results of Theorem 1 and Theorem 2 are of course much more general, and in principle apply to models for networks of neurons in the whole brain [87]. This perspective suggests understanding spontaneous brain activity (see [9]) as sampling from this global distribution in the absence of external input, and brain computations with external inputs as sampling of brain states from conditional distribution , thereby merging the knowledge base of the brain with incoming new information . This computational framework could in principle explain how the brain can merge both types of information in such seemingly effortless manner, a capability that can only partially be reproduced in artificial devices with current technology. Large-scale computer simulations will be needed to test the viability of this hypothesis, in particular the relationship between the known global structure of the brain network and properties of its stationary distribution , and the convergence speed to . Possibly the brain uses an important trick to speed up convergence during brain-wide sampling, for example by sampling during any concrete brain computation only from a subnetwork of : those brain areas that control variables that are relevant for this computation. Functional connectivity would be explained from this perspective as opening of communication channels that support sampling from the (marginal) joint distribution of those variables that are stored within the functionally connected brain areas. Structured spontaneous brain activity [9] would then receive a functional interpretation in terms of updating these marginal joint distributions on the basis of newly acquired knowledge.
A surprisingly large number of computational tasks that the brain has to solve, from the formation of a percept from multi-modal ambiguous sensory cues, to prediction, imagination, motor planning, rule learning, problem solving, and memory recall, have the form of constraint satisfaction problems: A global solution is needed that satisfies all or most of a set of soft or hard constraints. However, this characterization per se does not help us to understand how the brain can solve these tasks, because many constraint satisfaction problems are computationally very demanding (in fact, often NP-hard [88]), even for a fast digital computer. In the Sudoku example we have shown that the inherent stochastic dynamics of cortical microcircuits provides a surprisingly simple method for generating heuristic solutions to constraint satisfaction problems. This is insofar remarkable, as this computational organization does not require that specific algorithms are programmed into the network for solving specific types of such problems (as it is for example needed for solving Sudoku puzzles according to the ACT-R approach [89]). Rather, it suffices that salient constraints are encoded into the network (e.g. through learning) in such a way that they make certain firing patterns of a subset of neurons more or less likely.
Future work will need to investigate whether and how this approach can be scaled up to larger instances of NP-complete constraint satisfaction problems. For example, it will be interesting to see whether stochastic networks of spiking neurons can also efficiently generate heuristic solutions to energy minimization problems [90] arising in visual processing.
Furthermore, additional research is needed to address suitable readout mechanisms that stabilize and evaluate promising candidate solutions (see [76] for an experimentally supported mechanism that might contribute to this function). This is an important issue since, in its current form, the network will simply continue the stochastic exploration of heuristic solutions even after it has found the optimal solution. Therefore, in the absence of additional mechanisms the network is not able to hold on to (or store) previously found (near-)optimal solutions. To solve this issue one could consider, for example, one or several networks which generate in parallel heuristic solutions to a given problem. The output of these networks could then be further processed and integrated by a readout network which attempts to extract a MAP solution, for example by adopting a solution from some only if it has higher value than the currently stored state. Hence, the sampling networks would have stationary distributions which encourage exploration and broadly assign probability to many different heuristic solutions, whereas the readout network would ideally exhibit a sharply peaked stationary distribution at the global optimum of the constraint satisfaction problem. Studying the feasibility of this approach requires further research.
A substantial number of behavioral studies in cognitive science (see e.g. [69], [91]–[94]) have arrived at the conclusion that several of the previously discussed higher level mental operations are implemented through probabilistic inference. Some of the underlying data also suggest that probabilistic inference is implemented in the brain through some form of sampling (rather than through arithmetical approaches such as belief propagation [44]). But according to [94]: “The key research questions are as follows: What approximate algorithms does the mind use, how do they relate to engineering approximations in probabilistic AI, and how are they implemented in neural circuits?” This article contributes to these fascinating questions by providing a rigorous theoretical foundation for the hypothesis that neural circuits in the brain represent complex probability distributions through sampling. In addition, we have provided evidence that this form of sampling in cortical microcircuits may be fast enough to facilitate the approximate estimation of marginals or marginal MAP assignments, which commonly appear in real-world inference tasks, within a few . A major challenge for future work will be to understand also neuronal plasticity on the implementation level from this perspective. For example, how can prior knowledge be acquired and integrated into the stationary distribution of a realistic circuit (featuring short-term plasticity, dendritic processing, etc.) in an autonomous fashion, and in a manner consistent with statistically optimal learning [25]?
In biological networks it is reasonable to assume that the network dynamics unfolds on a continuum of time scales from milliseconds to days. Our goal in this article was to focus on stochastic computations on shorter time scales, between a few milliseconds to seconds. To this end we assumed that there exists a clear separation of time scales between fast and slow dynamical network features, thus allowing us to exclude the effect of slower dynamical processes such as long-term plasticity of synaptic weights during these shorter time scales. In network models and experimental setups where slower processes significantly influence (or interfere with) the dynamics on shorter time scales, it would make sense to extend the concept of a stationary distribution to include, for example, also the synaptic parameters as random variables. A first step in this direction has been made for neurons with linear sub-threshold dynamics and discretized synapses in [18].
Deterministic network models such as leaky integrate-and-fire neurons without noise (no external background noise, no synaptic vesicle noise and no channel noise) violate the assumptions of Theorem 1 and 2. Furthermore, although realistic neurons are known to possess various noise sources, the theoretical assumptions could in principle still fail if the network is not sufficiently stochastic: this would happen, for example, if there exists some strong input (within the limits of typical input activity) which entirely overrules the noise, leading to a firing probability in some time interval during the network simulation. Such deterministic behavior would correspond to the instantaneous firing rate of a stochastic neuron becoming infinite at some point during that interval (in violation of assumption A2, see Methods: Scope of theoretical results). From an empirical perspective, a simple necessary condition for sufficient stochasticity is the presence of trial-to-trial variability for each single spike produced by a network. Consider, for example, the spike times generated by a specific neuron in a network simulation, in response to some fixed input spike train. If there exists a spike which always occurs at the exact same time during multiple repetitions of this experiment starting from identical initial states, then the assumptions of Theorem 1 and 2 are obviously violated.
For deterministic (or insufficiently stochastic) networks the question arises whether convergence to a unique stationary distribution may still occur under appropriate conditions, perhaps in some modified sense. Notably, it has been recently observed that deterministic networks may indeed lead to apparently stochastic spiking activity [95], [96]. This apparent stochasticity was linked to chaotic spiking dynamics. This suggests that chaos may act as a substitute for “real” noise in deterministic networks (similar to pseudo random-number generators emulating true randomness): Chaotic systems are sensitive to small perturbances in initial conditions, and may thus exponentially amplify otherwise insignificant noise sources such as ubiquitous thermal noise [5]. Thus, chaos could play an important role in both emulating and amplifying stochasticity on the network level.
[96] focused their analysis of stochasticity on firing rate fluctuations and spiking irregularity, and it remains unclear whether these networks would still appear stochastic if one takes into account full network states (as in this article). The Gelman-Rubin convergence analysis of population activity proposed in this paper could be applied to provide some insight into this question. A more thorough investigation of chaos in the context of our results would also call for a rigorous theoretical analysis of ergodic properties of chaotic spiking networks.
Our theoretical results demonstrate that every neural system has a stationary distribution of network states . This can be tested experimentally, for various behavioral regimes and external inputs . A first step in this direction has already been carried out in [20] (see also the discussion in [97]). The hypothesis that serves (for “neutral” external inputs ) as a prior for probabilistic inference through sampling suggests that is constantly modified through prior experience (see [98], [99] for first results) and learning (see [10] for fMRI data).
Our Theorem 2 suggests in addition that neural systems that have a prominent rhythm (such as for example the theta oscillation in the hippocampus) are able to store several stationary distributions of network states, one for each clearly separable phase of this rhythm. It has already been shown in a qualitative manner that in some behavioral situations certain states appear with substantially high probability at specific phases of the rhythm (see e.g. [36], [62], [64], [66], [100]). But a systematic experimental analysis of phase-dependent distributions of network states in the style of [20] is missing.
Our Theorem 1 predicts in addition that a generic neural circuit also has a stationary distribution over trajectories of network states. The existence of stereotypical trajectories of network states in the awake brain has been frequently reported (see e.g. [40], [41], [98], [101]). But a statistical analysis of the distribution of such trajectories, especially also during spontaneous activity, is missing. Of particular interest is the relationship between the distribution of trajectories and the stationary distribution of (simple) network states. Do some network states typically have a high probability because they occur in some high probability trajectory? And how does the distribution of trajectories change during learning?
The model for problem solving that we have presented in Figure 5 suggests that external constraints have a significant and characteristic impact on the structure of the stationary distribution , by reducing the probability of network states which are inconsistent with the current constraints . In principle, this could be analyzed experimentally. In addition, this model suggests that there may be special mechanisms that prolong the time span during which a neural system stays in a network state with high probability under , in order to support a readout of by downstream networks. These mechanisms need to be revealed through experiments.
The Sudoku example has shown that networks of spiking neurons with noise are in principle able to carry out quite complex computations. The constraints of many other demanding constraint satisfaction problems, in fact even of many NP-complete problems, can be encoded quite easily into circuit motifs composed of excitatory and inhibitory spiking neurons, and can be solved through the inherent stochastic dynamics of the network. This provides new computational paradigms and applications for various energy-efficient implementations of networks of spiking neurons in neuromorphic hardware, provided they can be equipped with sufficient amounts of noise. In particular, our results suggest that attractive computational properties of Boltzmann machines can be ported into spike-based hardware. These novel stochastic computing paradigms may also become of interest for other types of innovative computer hardware: Computer technology is approaching during the coming decade the molecular scale, where noise is abundantly available (whether one wants it or not) and it becomes inefficient to push through traditional deterministic computing paradigms.
The results of this article show that stochastic computation provides an attractive framework for the investigation of computational properties of cortical microcircuits, and of networks of microcircuits that form larger neural systems. In particular it provides a new perspective for relating the structure and dynamics of neural circuits to their computational properties. In addition, it suggests a new way of understanding the organization of brain computations, and how they are modified through learning.
If the input sequence is periodic with period , i.e. for all , then the Markov process will be time-periodic, in the sense that transition kernels are invariant to shifts which are multiples of the period :(52)
This implies the following result, which is a more precise version of Theorem 2:
Lemma 4 Under periodic input, i.e. for all with some , the time-periodic Markov process with period has a periodically stationary distribution , to which convergence occurs exponentially fast from any initial state. In particular, for every there exists a unique stationary distribution such that,(53)from any initial Markov state .
Proof: For each there exists a skeleton chain , with transition probability kernel , which is time-homogeneous, irreducible, and aperiodic and thus has a unique stationary distribution . An application of , which corresponds to a full period, decreases the total variation distance to by at least :(54)(55)(56)The first inequality follows from the fact that applying the remaining can only further decrease the total variation distance between the two distributions, according to (23). The second inequality is due to Lemma 1.
Lemma 4 then follows from recursive application of (54)–(56) for multiple periods, and choosing a singleton as initial distribution.
In the main text, we use the notation for a phase-specific stationary distribution, where denotes a specific periodic input sequence.
Previous work on the question whether states of spiking neural networks might converge to a unique stationary distribution had focused on the case where neuronal integration of incoming spikes occurs in a linear fashion, i.e., linear subthreshold dynamics followed by a single output non-linearity [17], [18]. In addition these earlier publications did not allow for the experimentally observed short term dynamics of synapses. The earlier publication [17] had studied this question as a special case of the mathematical framework of non-linear Hawkes processes, a class of mutually exciting point processes (see also [109]). The authors had arrived for the more restricted type of neurons which they considered at exponential convergence guarantees under a similar set of assumptions as in this article, in particular bounded memory and bounded instantaneous firing rates (and these results can thus be seen as a special case of Theorem 1, for the case of constant external input). [17] also derived convergence results for linearly integrating neurons with unbounded memory dynamics under a different set of assumptions, in particular Lipschitz conditions on the output non-linearity and constraints on the effective connectivity matrix of the network. Whether such alternative set of assumptions can be found also in the context of non-linear integration of incoming spikes (needed e.g. for synaptic short-time dynamics or dendritic non-linearities) remains an open question.
The recent publication [18] also focused on neurons with linear sub-threshold dynamics followed by an output non-linearity (termed there non-linear Poisson neurons) with static synapses, and extended the convergence results of [17] to networks with Hebbian learning mechanisms. In addition, an important methodological innovation by [18] was the introduction of spike history states (which are equivalent to the Markov states in this article) which allowed them to study convergence in the framework of general Markov processes (in contrast to point processes in [17]). Theorem 1 in this article contains as a special case the convergence results of [18] for their Model I (non-linear Poisson neurons in the absence of Hebbian learning). We note that although [18] focused on neurons with linear sub-threshold dynamics (and required that firing rates are strictly greater than ), their method of proof for Model I could be readily extended to cover also non-linear sub-threshold dynamics to yield the first part of our Theorem 1 (the case where inputs have constant firing rates).
We are not aware of previous work that studied convergence in spiking networks with dynamic synapses, or in the presence of stochastic or periodic inputs (see the second part of Theorem 1 concerning Markov processes as input, and Theorem 2). We further note that our method of proof builds on a new and rather intuitive intermediate result, Lemma 1 (Contraction Lemma), which may be useful in its own right for two reasons. On the one hand it provides more direct insight into the mechanisms responsible for convergence (the contraction between any two distributions). On the other hand, it holds regardless of the input trajectory , and hence has in fact an even larger scope of applicability than Theorem 1 and 2. Hence, Lemma 1 could be, for example, applied to study non-stationary evolutions of state distributions in response to arbitrary input trajectories.
A key advantage of sample-based representations of probability distributions is that probabilities and expected values are in principle straightforward to estimate: To estimate the expected value of a function under a distribution from a number of samples , simply apply the function to each sample and compute the time average . As long as the samples are distributed according to , either independently drawn, or as the result of an ergodic Markov chain/process with stationary distribution , this estimate is guaranteed to converge to the correct value as one increases the number of samples [110], i.e. . Estimates based on a finite observation window represent an approximation to this exact value.
Under the mild assumptions of Theorem 1 the dynamics of a stochastic spiking network in response to an input are exponentially ergodic and there exists a unique stationary distribution , according to which network states are distributed. Hence, the expected value of any function under the stationary distribution can be estimated by computing the sample-based time average(57)
This approach can also be used to estimate marginal probabilities, since probabilities can be expressed as expected values, for example,(58)(59)where if and otherwise. Hence, in order to estimate the probability it suffices to measure the relative time the neuron spends in its active state, i.e. . Similarly, to estimate the probability it is sufficient to keep track of the relative frequency of the pattern , by computing .
All simulations of microcircuit models for Figures 1–4 were carried out in PCSIM [111]. A time step of was chosen throughout. Further analysis of spike trains was performed in Python [112].
The small cortical microcircuit model of Figure 1B was constructed based on the cortical column template of [30]: Synaptic connections between neurons and their weights were chosen to approximately reflect connection probabilities and mean synaptic strengths of the cortical column template [30]. Due to the very small size of this network, the resulting dynamics was not immediately satisfactory (for example, the influence of inputs on Layer 5 neurons was too weak). To shift the circuit into a more responsive regime, we manually adjusted a few synaptic weights and neuronal excitabilities. In particular, we injected small constant currents into some of the neurons to modulate their intrinsic excitability. Furthermore, to increase activity and correlations between highlighted neurons 2, 7 and 8, we increased synaptic weights and by factors 5 and 10, respectively. To set the initial Markov state of the network, preparatory input was shown for before the actual start of the simulation. Two different preparatory inputs were injected to set the two initial states considered in Figure 1F–H (first: at , at , second: both and at ). To reproduce the same initial Markov state in multiple trials (for example the two trials shown in Figure 1F), the same random seed was used during the preparatory phase for these trials. The random seed was then reinitialized at to different values for each trial.
In Figure 3 we compared convergence times in four different neural circuits. The first circuit was identical to the small cortical microcircuit from Figure 1. For the remaining three circuits, the same stochastic point neurons and conductance-based dynamic synapses with delays were used as for the data-based cortical microcircuit model. Dynamic synaptic parameters were set to the corresponding mean values of parameters used in the cortical column model. Synaptic delays of were used for all networks, except for the network with sequential structure (Figure 3C) where delays were . To modulate the intrinsic excitability of neurons we injected small currents to each neuron. The strengths of injected currents and connections were tuned for each network until the desired network activity was achieved. Synaptic background inputs were injected as in the cortical microcircuit model. To set different initial states (needed for Gelman Rubin analysis), during a preparatory phase of we injected into each neuron a random current chosen from . These small random input currents were strong enough to yield sufficiently diverse initial states. Gelman-Rubin values were then calculated based on runs, where the duration of each run (after the preparatory phase) was of biological time. Convergence analysis was performed on marginals (individual simple states with ). Mean and worst marginals were computed as described in the previous section.
Below are additional details to the circuits used for Figure 3B–D: The sparsely active network of Figure 3B comprises one excitatory (E) and one inhibitory (I) population (each 10 neurons). Connections between neurons were drawn randomly according to the following set of connection probabilities: EE = 0.1, EI = 0.1, II = 0.9, IE = 0.9. The network with sequential structure of Figure 3C consists of two interconnected subnetworks where each one of them produces a stereotypical trajectory. Each subnetwork consists of a trigger neuron, a subsequent chain of neurons, and a pool of inhibitory neurons. Shown in Figure 3C are only the excitatory chain neurons from each subnetwork (neurons 1–15: first subnetwork; neurons 16–30: second subnetwork). Each excitatory neuron in the chain projects to all other neurons in the same chain with synaptic strengths decreasing with distance according to where applies to the forward direction in the chain and to the backward direction. The trigger neuron projects (forward) to the chain in the same fashion with . All neurons in the chain project to the inhibitory pool, and all neurons in the inhibitory pool project back to the trigger neuron and to the chain. Finally, the two subnetworks are combined such that the inhibitory pool of one subnetwork projects to the trigger neuron and the chain of the other subnetwork, and vice versa. This ensures that only one of the two subnetworks can be active at a time (competition between two trajectories). The bistable network of Figure 3D consists of two populations which strongly inhibit each other (each population comprising 10 neurons).
The theoretical proof for Theorem 2 can be found after the proof of Theorem 1 above. For Figure 4F, a single long simulation () of the bi-stable network in Figure 4E was carried out. Each of the two pools was defined active at time if more than two neurons from the pool had an active simple state at time (with ). A transition was defined as the succession of a period in which one pool was active and the other pool inactive by a period in which the other became active and the first pool turned inactive. Between those two periods it typically occurs that either both pools are active or both are inactive for some short time. The exact time (and phase within the current cycle) of each transition was defined as the point in the middle of this intermediate period.
|
10.1371/journal.pgen.1004467 | MDRL lncRNA Regulates the Processing of miR-484 Primary Transcript by Targeting miR-361 | Long noncoding RNAs (lncRNAs) are emerging as new players in gene regulation, but whether lncRNAs operate in the processing of miRNA primary transcript is unclear. Also, whether lncRNAs are involved in the regulation of the mitochondrial network remains to be elucidated. Here, we report that a long noncoding RNA, named mitochondrial dynamic related lncRNA (MDRL), affects the processing of miR-484 primary transcript in nucleus and regulates the mitochondrial network by targeting miR-361 and miR-484. The results showed that miR-361 that predominantly located in nucleus can directly bind to primary transcript of miR-484 (pri-miR-484) and prevent its processing by Drosha into pre-miR-484. miR-361 is able to regulate mitochondrial fission and apoptosis by regulating miR-484 levels. In exploring the underlying molecular mechanism by which miR-361 is regulated, we identified MDRL and demonstrated that it could directly bind to miR-361 and downregulate its expression levels, which promotes the processing of pri-miR-484. MDRL inhibits mitochondrial fission and apoptosis by downregulating miR-361, which in turn relieves inhibition of miR-484 processing by miR-361. Our present study reveals a novel regulating model of mitochondrial fission program which is composed of MDRL, miR-361 and miR-484. Our work not only expands the function of the lncRNA pathway in gene regulation but also establishes a new mechanism for controlling miRNA expression.
| Long non-coding RNAs (lncRNAs) have been shown to be involved in a wide range of biological functions. However, studies linking individual lncRNA to the mitochondrial fission program remain scarce. Also, it remains unknown whether lncRNAs can operate in the processing of miRNA primary transcript. Here, we provide causal evidence for the involvement of the lncRNA MDRL in the mitochondrial dynamics and the processing of miR-484 primary transcript in cardiomyocyte. We identified MDRL which can act as an endogenous ‘sponge’ that directly binds to miR-361 and downregulates its expression levels. miR-361 can directly bind to primary transcript of miR-484 and prevent its processing by Drosha into pre-miR-484. MDRL inhibits mitochondrial fission and apoptosis by miR-361 and miR-484. Our present study reveals a novel regulating model which is composed of MDRL, miR-361 and miR-484. Modulation of their levels may provide a new approach for tackling myocardial infarction.
| Long non-coding RNAs (lncRNAs) are non-protein coding transcripts longer than 200 nucleotides. A number of lncRNAs have been shown to be involved in a wide range of biological functions including RNA processing [1], gene transcription regulation [2], miRNAs' host genes [3], modulation of apoptosis and invasion [4], marker of cell fate [5], chromatin modification [6], etc. Also, misregulation of lncRNAs has been observed in human diseases such as cancer and neurological disorders. In addition, lncRNAs not only can act as antisense transcripts or as decoy for splicing factors leading to splicing malfunctioning [7], [8], but also can act as a competing endogenous RNA (ceRNA) in mouse and human myoblasts [9]. However, it is not yet clear whether lncRNA can be involved in the processing of pri-miRNA and the regulation of mitochondrial network.
MicroRNAs (miRNAs) act as negative regulators of gene expression by inhibiting the translation or promoting the degradation of target mRNAs. Growing evidence has demonstrated that miRNAs can play a significant role in the regulation of development, differentiation, proliferation and apoptosis. Recent studies have identified the functional role of miRNA in numerous facets of cardiac biology, including the control of myocyte growth, contractility, fibrosis, angiogenesis, heart failure, and myocardial infarction, providing potential therapeutic targets for heart disease. To prevent and reverse myocardial infarction, it is critical to identify those miRNAs that are able to regulate myocardial infarction and to characterize their signal transduction pathways in the apoptotic cascades.
Mature miRNAs execute their functions mainly in the cytoplasm. Some studies have observed that there exist mature miRNAs in the nucleus [10], [11]. Their work demonstrated that additional sequence elements of specific miRNAs can control their posttranscriptional behavior, including the subcellular localization. Recent work reported that the miRNA pathway targets non-coding RNAs across species. It showed that let-7 could regulate its biogenesis autonomously through a conserved complementary site in its own primary transcript, creating a positive-feedback loop [11]. However, the molecular mechanism of miRNAs and their regulation model in the nucleus remains to be fully elucidated.
Mitochondria are highly dynamic organelles that constantly undergo fusion and fission to form a network that spans the entire area of the cell. Mitochondrial fission and fusion are crucial for maintaining mitochondrial function and are necessary for the maintenance of organelle fidelity [12]. The disruption of mitochondrial fission and fusion has been linked to the development and progression of some diseases [13]–[17]. Most recent studies have revealed that abnormal mitochondrial fusion and fission participate in the regulation of apoptosis. Mitochondrial fusion is able to inhibit apoptosis, while mitochondrial fission is necessary for initiation of apoptosis [12], [18]–[23]. Thus, exploring the function of mitochondrial fission and fusion regulators will unveil their roles in various pathway and diseases.
Our present work revealed that nuclear miR-361 can directly bind to pri-miR-484 and inhibiting its processing into pre-miR-484 which is mediated by Drosha. miR-361 participates in the regulation of mitochondrial network and apoptosis through the miR-484 pathway. Moreover, our study further suggested that a long noncoding RNA, MDRL, can directly bind to miR-361 and act as an endogenous “sponge” which downregulates miR-361 expression levels and promotes the processing of pri-miR-484. In short, MDRL regulates mitochondrial network and apoptosis through the miR-361 and miR-484. Our data reveal a novel role for lncRNA and miRNA in promoting biogenesis of other miRNA primary transcript, expanding the functions of the lncRNA and miRNA in gene regulation and mitochondrial network.
Many studies have observed that there exist mature miRNAs in the nucleus [12], [18]–[23] and recent work also reports that let-7 miRNA in the nucleus can regulate its own primary transcript through a conserved complementary site, thus creating a positive-feedback loop [11]. Our previous work has showed that transcription factor could regulate miR-484 expression [24]. To further explore other underlying mechanism responsible for miR-484 regulation under anoxia/reoxygenation (A/R) condition, we tested whether miRNA in the nucleus participates in the regulation of miR-484 expression. To understand which nuclear miRNA is involved in the apoptosis pathway of A/R, we performed a microarray to detect nuclear miRNAs in response to A/R treatment (Figure 1A, Figure 1B and Table S1) and among these miRNAs induced by A/R, only knockdown of endogenous miR-361 (Figure S1A) induced an increase in the miR-484 expression levels (Figure 1C). A further study confirmed that miR-361 was predominantly located in the nucleus, and miR-484 was predominantly located in the cytoplasm (Figure 1D). We test whether nuclear miR-361 may directly affect the expression of miR-484. The results showed that enforced expression of miR-361 could reduce mature miR-484 levels (Figure 1E). Furthermore, miR-361 transgenic mice (Figures S1B and S1C) demonstrated reduced levels of miR-484 in the animal model (Figure S1D and Figure 1F). Taken together, it appears that miR-361 is predominantly located in the nucleus and is able to regulate mature miR-484 levels in the cellular and the animal model.
To understand the mechanism by which nuclear-located miR-361 regulates the levels of cytoplasmic mature miR-484, we tested whether miR-361 is able to affect the levels of pri-miR-484 located in nucleus. We compared the sequences of miR-361 with that of pri-miR-484 using the bioinformatics program RNAhybrid and noticed that miR-361 is complementary to pri-miR-484 (Figure 2A). Their complementary sequences led us to consider whether miR-361 can directly interact with pri-miR-484 and inhibit its processing into pre-miR-484 in the nucleus. We demonstrated that enforced expression of miR-361 resulted in the strong accumulation of pri-miR-484 (Figure 2B) and the reduction of pre-miR-484 (Figure 2C). Knockdown of miR-361 resulted in the reduction of pri-miR-484 (Figure 2D) and the increase of pre-miR-484 (Figure 2E). Thus, it appears that miR-361 prevents the processing of pri-miR-484 into pre-miR-484 in nucleus.
Mature miRNAs are generated via a two-step processing by Drosha and Dicer. The initial processing that occurs in the nucleus is catalyzed by Drosha. The Drosha complex cleaves pri-miRNA into pre-miRNA. To further verify whether miR-361 prevents the processing of Drosha. We applied Drosha to assay the levels of pri-miR-484 and pre-miR-484. Our data showed that enforced expression of miR-361 prevented the reduction of pri-miR-484 and inhibited the increase of pre-miR-484 induced by Drosha, and knockdown of miR-361 had an opposite effects (Figures 2F and 2G). These data suggest that miR-361 prevents the processing of pri-miR-484 by Drosha into pre-miR-484 in nucleus. To further test whether the miR-361 recognition element on pri-miR-484 was responsible for miR-361 binding and inhibition of processing, we applied a biotin-avidin pull-down system to assess the direct binding of miR-361 to pri-miR-484. The cardiomyocytes were transfected with biotinylated miR-361, biotinylated mutant miR-361 and biotinylated negative control (NC) (Figure S1E). Then the cells were harvested for biotin-based pull-down assay. Pri-miR-484 was co-precipitated, and the levels of pri-miR-484 in the pull-down complexes was analyzed by the qRT-PCR (Figure 2H). As shown in Figure 2H, pri-miR-484 was significantly enriched in the miR-361 pull-down products as compared to the biotinylated mutant miR-361 group and negative control group, indicating that miR-361 can directly bind to pri-miR-484 in vivo. We also employed inverse pull-down assay to test whether pri-miR-484 could pull down miR-361, a biotin-labeled-specific pri-miR-484 probe was used. The results showed that pri-miR-484 (Figure S1F) and miR-361 could be co-precipitated (Figure 2I). Taken together, it appears that miR-361 is able to directly bind to pri-miR-484 and prevent the processing of pri-miR-484 by Drosha into pre-miR-484.
Our previous findings found that miR-484 could inhibit mitochondrial fission and apoptosis in cardiomyocytes [24]. The present result appears that miR-361 can interact with pri-miR-484 and regulate mature miR-484 levels. We thus explored the functional role of miR-361 in mitochondrial fission and apoptosis. To this end, the antagomir of miR-361 was employed to knock down endogenous miR-361. Mitochondrial fission induced by A/R was attenuated by the knockdown of miR-361 (Figure 3A). Concomitantly, apoptosis was reduced in the presence of miR-361 antagomir (Figure 3B). These data indicate that miR-361 can promote mitochondrial fission and apoptosis upon A/R treatment.
To understand the pathophysiological role of miR-361, we detected whether miR-361 is involved in the pathogenesis of myocardial infarction in the animal model. miR-361 was elevated in response to ischemia/reperfusion (Figure S2A). Knockdown of miR-361 resulted in a reduction in mitochondrial fission (Figure 3C, upper panel) and apoptosis (Figure 3C, low panel and right panel). We produced miR-361 transgenic mice, and these mice exhibited increased mitochondrial fission, apoptosis (Figure 3D), myocardial infarction sizes (Figure 3E) and potentiated cardiac dysfunction (Figure S2B) in response to ischemia/reperfusion (I/R). Taken together, it appears that miR-361 is able to promote mitochondrial fission, apoptosis and myocardial infarction. How does miR-361 exert its effect on the mitochondrial network? Because miR-361 is able to reduce mature miR-484 expression as shown in Figure 1, we thus tested whether miR-484 is a mediator of miR-361. To confirm the relationship between miR-361 and miR-484 in mitochondrial fission machinery, we used antagomir to inhibit miR-484 levels, and observed that the inhibitory effect of miR-361 knockdown on mitochondrial fission and apoptosis was attenuated in the presence of miR-484 antagomir (Figure S3A and Figure 3F). Taken together, these data suggest that miR-361 targets miR-484 in the cascades of mitochondrial fission and apoptosis.
Recent studies have suggested that lncRNAs may act as endogenous sponge RNA to interact with miRNAs and influence the expression of miRNA [9], [25]–[27]. To explore the underlying mechanism responsible for miR-361 upregulation in response to A/R treatment, we tested whether lncRNA could regulate miR-361 expression. We carried out qRT-PCR to detect lncRNAs levels in response to A/R treatment. LncRNAs were chosen from the lncRNA array published online by Fantom company. Among 100 lncRNAs, AK009271 which we named mitochondrial dynamic related lncRNA (MDRL), was substantially reduced (Figure 4A). The MDRL is 1039 nt in length and the subcellular location showed that MDRL was expressed both in nucleus and cytoplasm (Figure S3B). Further, our results showed that miR-361 levels were elevated in the cells upon knockdown of endogenous MDRL (Figures 4B and 4C). To know whether MDRL can affect miR-361 activity, we constructed miR-361 sensor (with a perfect miR-361 binding site). The lucifease activity of miR-361 sensor was decreased in cells treated with MDRL siRNA (Figure 4D), suggesting the induction of miR-361 activity. Enforced expression of MDRL induced a reduction in miR-361 expression (Figure 4E) and activity (Figure 4F). To further test whether MDRL may act as a miR-361 sponge, we transfected the miR-361 sensor luciferase reporter, along with adenoviral miR-361, MDRL or β-gal. The luciferase activity showed that MDRL counteracted the effect of miR-361 (Figure 4G), suggesting that MDRL is a functional sponge for miR-361. Taken together, these data suggest that MDRL is able to regulate miR-361 levels and activity.
To understand the mechanism by which MDRL regulates the levels of miR-361, we tested whether MDRL can interact with miR-361. We compared the sequences of MDRL with that of miR-361 using the bioinformatics program RNAhybrid and noticed that MDRL contains a target site of miR-361 (Figure 5A). The wild type luciferase construct of MDRL (Luc-MDRL-wt) and a mutated form (Luc-MDRL-mut) were produced by inserting the sequence of putative miR-361 binding site into the report constructs (Figure 5B, upper panel). Luciferase assay revealed that miR-361 could suppress the luciferase activity of MDRL, but it had less effect on the mutated form of MDRL compared to the wild type (Figure 5B). Our results further showed that the mutated form of MDRL had no effect on miR-361 activity (Figure S3C) and it also lost the ability to counteract miR-361 (Figure S3D). These results revealed that MDRL may interact with miR-361 by this putative binding site.
Further, we applied a biotin-avidin pull-down system to test if miR-361 could pull down MDRL. Cardiomyocytes were transfected with biotinylated wild type miR-361, biotinylated mutant miR-361 and biotinylated miR-NC (Figure S4A). We found these transfections did not change MDRL levels (Figure S4B). Then, Cardiomyocytes were harvested for biotin-based pull-down assay. MDRL was pulled down by wild type miR-361 as analyzed by qRT-PCR, but the introduction of mutations that disrupt base pairing between MDRL and miR-361 (Figure 5C) led to the inability of miR-361 to pull down MDRL (Figure 5D), indicating that the recognition of miR-361 to MDRL is in a sequence-specific manner. We also employed inverse pull-down assay to test if MDRL could pull down miR-361, a biotin-labeled-specific MDRL probe was used. The results showed that MDRL (Figure S4C) and miR-361 (Figure 5E) could be co-precipitated. Taken together, it appears that MDRL is able to directly bind to miR-361 in vivo.
We further tested whether MDRL is able to regulate miR-484 expression and influence its processing. Our results showed that enforced expression of MDRL (Figure S4D) resulted in the decrease of pri-miR-484 (Figure 5F) and the accumulation of pre-miR-484 (Figure 5G). Knockdown of MDRL induced the increase of pri-miR-484 (Figure S4E) and the decrease of pre-miR-484 (Figure S4F). Knockdown of MDRL inhibited the decrease of pri-miR-484 (Figure 5H) and the increase of pre-miR-484 (Figure S4G) induced by Drosha, and the inhibitory effect of MDRL knockdown was reduced in the presence of miR-361 antagomir (Figure 5H and Figure S4G). Thus, it appears that MDRL promotes the processing of pri-miR-484 into pre-miR-484 through targeting miR-361.
Our present results have demonstrated that MDRL could promote the processing of pri-miR-484 by Drosha. Thus, we tested whether MDRL is able to regulate mature miR-484 levels. Knockdown of MDRL reduced miR-484 levels (Figure 6A), while overexpression of MDRL resulted in up-regulation of miR-484 expression (Figure 6B). And MDRL counteracted the effect of miR-361 on miR-484 expression (Figure 6C). Our previous report has demonstrated that Fis1 is a downstream target of miR-484. The current data showed that MDRL could regulate Fis1 expression by miR-484 (Figure S5A). These results indicated that MDRL may act as endogenous sponge “antagomir” of miR-361 to regulate the processing of pri-miR-484 and expression of miR-484.
To explore the functional role of MDRL, we tested that enforced expression of MDRL inhibited mitochondrial fission (Figure 6D) and apoptosis (Figure 6E and Figure S5B) induced by A/R. We also demonstrated that knockdown of MDRL induced mitochondrial fission (Figure S5C) and apoptosis (Figure S5D). In the animal model, administration of MDRL attenuated mitochondrial fission, cell death (Figure 6F) and myocardial infarction sizes (Figure 6G) in response to ischemia/reperfusion (I/R). Administration of MDRL also ameliorated cardiac function (Figure S5E). Taken together, it appears that MDRL is able to prevent mitochondrial fission, apoptosis and myocardial infarction. Since MDRL is able to elevate miR-484 expression as shown in Figure 6B, we thus tested whether miR-484 is a mediator of MDRL in the mitochondrial network. To confirm the relationship between MDRL and miR-484 in mitochondrial fission machinery, we used miR-484 antagomir, and observed that the inhibitory effect of MDRL on mitochondrial fission and apoptosis was decreased in the presence of miR-484 antagomir (Figure 6H). Taken together, these data suggest that MDRL targets miR-361/miR-484 in the cascades of mitochondrial fission and apoptosis (Figure S6).
Our present work revealed that miR-361 directly binds to pri-miR-484 and prevents its processing by Drosha into pre-miR-484. miR-361 is able to regulate mitochondrial fission and apoptosis, and this regulatory effects on mitochondrial fission and apoptosis is through targeting miR-484. Our study further revealed that MDRL directly binds to miR-361 and acts as its “sponge”, promoting the processing of pri-miR-484. And MDRL can inhibit mitochondrial fission and apoptosis though targeting miR-361 and miR-484. Our results provide novel evidence demonstrating that MDRL, miR-361, miR-484 constitute an axis in the machinery of mitochondrial network.
LncRNAs have been defined to have important functions in specific cell types, tissues and developmental conditions such as chromatin modification [6], RNA processing [1], structural scaffolds [28] and modulation of apoptosis and invasion, etc [4]. Despite the biological importance of lncRNAs, it is not yet clear whether lncRNAs is involved in the processing of primary transcript and the regulation of mitochondrial network. Our present work for the first time reveals a novel function of lncRNA participating in regulating the processing of miR-484 primary transcript and mitochondrial dynamics. Our results may provide a new clue for the understanding of lncRNAs-controlled cellular events.
It has been shown that lncRNAs may act as endogenous sponge RNAs to interact with miRNAs and influence the expression of miRNA target genes. A recent report shows that the H19 lncRNA can act as a molecular sponge for the major let-7 family of miRNAs [27]. Other report demonstrates that a muscle-specific long non-coding RNA, linc-MD1, governs the time of muscle differentiation by acting as a competing endogenous RNA (ceRNA) in mouse and human myoblasts [9]. Highly up-regulated liver cancer (HULC) may act as an endogenous ‘sponge’, which down-regulates miR-372 leading to reducing translational repression of its target gene, PRKACB [26]. Transient knockdown and ectopic expression of HSUR 1 direct degradation of mature miR-27 in a sequence-specific and binding-dependent manner [25]. Our present study reveals that lncRNA (MDRL) sponges miR-361 and promoting the processing of pri-miR-484, which inhibits mitochondrial fission and apoptosis. The discovery of a long non-coding RNA in miRNA primary processing and mitochondrial dynamics may shed new lights on understanding the complex molecular mechanism of mitochondrial network.
Many research works reveal that mature miRNAs execute their functions mainly in the cytoplasm. Recently, it has been reported that miRNA also functions in nucleus [12], [18]–[23], but the function of nuclear miRNA remains to be fully unveiled. Mature miRNAs are generated via a two-step processing by Drosha and Dicer. The initial processing that occurs in the nucleus is catalyzed by Drosha. The Drosha complex cleaves pri-miRNA into pre-miRNA. The precise processing is pivotal to ensure the production of mature miRNA. This present work reveals that miR-361 in the nucleus can directly bind to the pri-miR-484 and prevent its processing by Drosha into pre-miR-484, and then further inhibit the biological function of miR-484. This finding may provide a new clue for the understanding of miRNAs-controlled gene expression.
Emerging data suggest that changes in mitochondrial morphology may be relevant to various aspects of cardiovascular biology including cardiac development, heart failure, diabetes mellitus, and apoptosis. The heart function stringently depends on the ATP-generating pathways [29], and cardiomyocytes are a good model to study mitochondrial dynamics because of the abundant existence of mitochondria. So far, it remains unclear whether lncRNA is involved in the regulation of mitochondrial dynamics. Our present work indicated that lncRNA (MDRL) can inhibit mitochondrial fission and apoptosis through regulating miR-361 and miR-484. The involvement of MDRL, miR-361 and miR-484 in regulating mitochondrial networks shed new lights on the understanding of mitochondrial integrity and cardiac pathophysiology.
In summary, our present study reveals that miR-361 located in nucleus can directly bind to pri-miR-484 and prevent its processing by Drosha into pre-miR-484. miR-361 reduces mature miR-484 levels and affects mitochondrial apoptotic pathway through targeting miR-484. Moreover, we demonstrated that MDRL acts as endogenous sponge RNA and inhibits miR-361 expression. MDRL is able to inhibit mitochondrial fission and apoptosis through targeting miR-361 and miR-484. Thus, modulation of MDRL and miR-361 may represent novel approaches for interventional treatment of cardiac disease. This finding may provide a new clue for the understanding of lncRNAs and miRNAs-controlled cellular events.
We declare that all experiments were performed according to the protocols approved by the Animal Care Committee, Institute of Zoology, Chinese Academy of Sciences, China.
Neonatal mouse cardiomyocytes were isolated and prepared as we described [30]. In brief, after dissection the hearts were washed, minced in HEPES-buffered saline solution containing 130 mM NaCl, 3 mM KCl, 1 mM NaH2PO4, 4 mM glucose and 20 mM HEPES (pH adjusted to 7.35 with NaOH). Tissues were then dispersed in a series of incubations at 37°C in HEPES-buffered saline solution containing 1.2 mg/ml pancreatin and 0.14 mg/ml collagenase (Worthington). After centrifugation the cells were re-suspended in Dulbecco's modified Eagle medium/F-12 (GIBCO) containing 5% heat-inactivated horse serum, 0.1 mM ascorbate, insulin-transferring-sodium selenite media supplement, 100 U/ml penicillin, 100 µg/ml streptomycin, and 0.1 mM bromodeoxyuridine. The dissociated cells were pre-plated at 37°C for 1 h. The cells were then diluted to 1×106 cells/ml and plated in 10 µg/ml laminin-coated different culture dishes according to the specific experimental requirements. Anoxia/reoxygenation was performed as follows. Briefly, cells were placed in an anoxic chamber with a water-saturated atmosphere composed of 5% CO2 and 95% N2. Cells were subjected to 6 hours of anoxia followed by 12 hours of reoxygenation (95% O2 and 5% CO2).
For creating miR-361 transgenic mice, a DNA fragment containing murine miR-361 was cloned to the vector, pαMHC-clone26 (kindly provided by Dr. Zhong Zhou Yang), under the control of the α-myosin heavy chain (α-MHC) promoter. The primers used to generate miR-361 transgenic mice include, forward primer: 5′-AGAATGAGGCTAACAGGTGAGTCATC-3′; reverse primer: 5′-TGACTGGCAGACACTGGTTTCAGGTGTTAC-3′. Microinjection was performed following standard protocols.
Mitochondrial staining was carried as we and others described with modifications [19], [26]. Briefly, cells were plated onto the cover-slips coated with 0.01% poly-L-lysine. After treatment they were stained for 30 min with 0.02 µM MitoTracker Green (Molecular Probes). Mitochondria were imaged using a laser scanning confocal microscope (Zeiss LSM510 META).
Data are expressed as the mean ± SEM of at least three independent experiments. We evaluated the data with Student's t test. We used a one-way analysis of variance for multiple comparisons. A value of p<0.05 was considered significant.
|
10.1371/journal.pgen.1002042 | DNA Damage, Somatic Aneuploidy, and Malignant Sarcoma Susceptibility in Muscular Dystrophies | Albeit genetically highly heterogeneous, muscular dystrophies (MDs) share a convergent pathology leading to muscle wasting accompanied by proliferation of fibrous and fatty tissue, suggesting a common MD–pathomechanism. Here we show that mutations in muscular dystrophy genes (Dmd, Dysf, Capn3, Large) lead to the spontaneous formation of skeletal muscle-derived malignant tumors in mice, presenting as mixed rhabdomyo-, fibro-, and liposarcomas. Primary MD–gene defects and strain background strongly influence sarcoma incidence, latency, localization, and gender prevalence. Combined loss of dystrophin and dysferlin, as well as dystrophin and calpain-3, leads to accelerated tumor formation. Irrespective of the primary gene defects, all MD sarcomas share non-random genomic alterations including frequent losses of tumor suppressors (Cdkn2a, Nf1), amplification of oncogenes (Met, Jun), recurrent duplications of whole chromosomes 8 and 15, and DNA damage. Remarkably, these sarcoma-specific genetic lesions are already regularly present in skeletal muscles in aged MD mice even prior to sarcoma development. Accordingly, we show also that skeletal muscle from human muscular dystrophy patients is affected by gross genomic instability, represented by DNA double-strand breaks and age-related accumulation of aneusomies. These novel aspects of molecular pathologies common to muscular dystrophies and tumor biology will potentially influence the strategies to combat these diseases.
| All kinds of muscular dystrophies (MDs) are characterized by progressive muscle wasting due to life-long proliferation of precursor cells of myo- (muscle), fibro- (connective tissue), and lipogenic (fat) origin. Despite discovery of many MD genes over the past 25 years, MDs still represent debilitating, incurable diseases, which frequently lead to premature death. Thus, it is imperative to gain novel insights into the underlying MD pathomechanisms. Here, we show that different mouse models for the most common human MDs frequently develop skeletal musculature-associated tumors, presenting as complex sarcomas, consisting of myo-, lipo-, and fibrogenic compartments. Collectively, these tumors are characterized by profound genomic instability such as DNA damage, recurring mutations in cancer genes, and aberrant chromosome copy numbers. We also demonstrate the presence of these cancer-related aberrations in dystrophic muscles from MD mice prior to formation of visible sarcomas. Moreover, we discovered corresponding genomic lesions also in skeletal muscles from human MD patients, as well as stem cells cultured thereof, and show that genomic instability precedes muscle degeneration in MDs. We thus propose that cancer-like genomic instability represents a novel, unifying pathomechanism underlying the entire group of genetically distinct MDs, which will hopefully open new therapeutic avenues.
| Muscular dystrophies (MDs) comprise a group of inherited disorders, characterized by progressive muscle wasting and weakness, frequently causing premature death due to lack of effective therapies. More than 150 years ago, Edward Meryon was the first to characterize the detrimental “fatty degeneration of the voluntary muscles” in Duchenne MD (Meryon E. Lancet 2:588, 1851). Today it is well accepted that the progressive loss of functional muscle tissue and its replacement by adipose and fibrous tissue represent a pathology common to all MDs despite their heterogeneous genetic etiology. Most MDs are caused by gene mutations that lead to absence or dysfunction of structurally and/or functionally important molecules of the muscle fiber [1], [2]. The sarcoplasmic spectrin-related protein dystrophin is thought to structurally stabilize the muscle fiber sarcolemma by linking the actin-based cytoskeleton to the extracellular matrix via interaction with the dystroglycan (DG)-complex. Lack or vast reduction of dystrophin causes severe Duchenne muscular dystrophy (DMD) [3] in humans and myopathy in corresponding mouse models, such as the mdx [4] or mdx-3Cv [5] mouse. Mutations in several glycosyltransferase-encoding genes, such as the fukutin related protein (FKRP) or LARGE lead to defective glycosylation of the α-subunit of DG. This molecular defect underlies the second most common group of MDs, the so called “secondary dystroglycanopathies”. Numerous other MD-related molecules are not known to directly interact with the DG complex, such as dysferlin or calpain-3. Defective expression of dysferlin, a ubiquitously expressed 230-kDa transmembrane protein that has been shown to be involved in resealing muscle fiber membranes, causes limb-girdle muscular dystrophy type 2B (LGMD2B) or Miyoshi-myopathy in humans [6], [7]. An inbred mutation in the murine dysferlin (Dysf) gene makes the SJL-mouse a naturally occurring animal model for the human dysferlinopathies [8]. Mutations of the CAPN3 gene encoding the muscle-specific calcium-activated neutral protease calpain-3, a proteolytic switch in muscle remodeling [9], cause LGMD2A, a MD with a wide clinical spectrum [10]. Again, the corresponding animal model, the Capn3-deficient mouse is only affected by a mild progressive muscular dystrophy [11]. Given the diverse and obviously unrelated functions of these proteins, whose absence or dysfunction causes MDs, a common pathomechanism driving the complex events of parallel muscle regeneration and degeneration and progressive proliferation of fibrous and fatty tissue seen in all MDs is likely but still remains elusive. In the light of the fact that nearly 25 years ago the DMD gene was identified as the molecular basis for Duchenne MD, the lack of causative therapies has dampened earlier therapeutic promises based on the discovery of molecular defects underlying several MDs and underscores the imperative need for a comprehensive understanding of pathology involved in these rare but lethal diseases.
When starting to study age-related phenotypes of murine MDs, we have observed the frequent and spontaneous occurrence of skeletal muscle-derived tumors in our colony of C57BL/10-mdx mice, suggesting a tumor-suppressive role of dystrophin in mice. Therefore we extended our studies to other dystrophin mutations, mouse strains, and even to other MD-mouse models for the most frequent MDs in humans, like dysferlin, calpain-3 and Large, respectively. We show that all of these MD-mouse lines are prone to develop mixed soft-tissue sarcomas containing tumor elements displaying histological and molecular characteristics of rhabdomyo-, fibro-, and liposarcoma. These MD-associated tumors share complex, non-random genomic alterations affecting well-known tumor suppressor as well as oncogenes and these cancer signatures are already detectable in dystrophic muscle tissue, independent of the underlying mutation. Consequently, we show that genomic instability and DNA damage are present also in muscle of human MD patients. Collectively, these data strongly support an unprecedented general link between muscular dystrophy and cancer, driven by the accumulation of DNA damage, chromosome copy number aberrations, and finally the origin of cell clones harboring cancer-like mutations in dystrophic muscle tissue. We propose that - similar to pre-neoplastic lesions - the dystrophic muscle is characterized by genomic instability, which contributes to a common hyperproliferative pathomechanism promoting the degenerative process in human MDs and favoring age-related tumorigenesis in the respective mouse models.
During the last two decades we have observed the spontaneous occurrence of soft tissue tumors arising from various skeletal limb and trunk muscles in our dystrophin-deficient C57BL/10 mdx-mouse [4] cohort. These tumors arose in aged mdx mice (mean age-of-onset: ∼540 d) with an incidence of almost 40%, whereas we never observed the occurrence of such tumors in our C57BL/10 wild-type mice. In our colony of another dystrophin-deficient mouse line, mdx-3Cv, which lacks both the muscle 427 kDa and non-muscle 71 kDa dystrophin isoforms due to a mutation at the intron-exon 66 junction [5], we observed the spontaneous occurrence of skeletal muscle-derived tumors indistinguishable from those observed in C57BL/10-mdx mice. However, mdx-3Cv developed skeletal muscle-tumors at a significant older age (∼660 d) and a decreased incidence of only 5% as opposed to our mdx colony. Because we could not figure out if these differences were due to the different genetic backgrounds (mdx: C57BL/10, mdx-3Cv: C57BL/6 x B6C3Fe) or due to the different dystrophin-mutations, we generated two novel mdx inbred strains, i.e. BALB/c-mdx, and C3H-mdx, respectively, and further studied mdx-mice on mixed C57BL/6 x BALB/c and C57BL/10 x B6C3Fe backgrounds. Indeed we observed the spontaneous occurrence of skeletal muscle-tumors also in these mdx-mice, underlining a strain-independent tumor-suppressor role of dystrophin. Mean ages-of-onset, incidences and gender distributions of tumor-formation were strongly strain-dependent, whereby the C57BL/10 background was most tumor-susceptible (Table 1). The spontaneous occurrence of skeletal muscle-associated tumors in different dystrophin-deficient mouse lines independent of the underlying dystrophin gene mutation supported a candidate tumor suppressor role of dystrophin.
In order to learn whether other MD-genes, which are not directly related to dystrophin, might also suppress tumor formation, we studied mice lacking dysferlin (DysfSJL mutation [8]; Dysf −/−), calpain-3 (Capn3 −/−; knockout [11]), or Large (Largemyd mutation [12]; Large −/−). In a colony of Dysf −/− mice inbred onto C57BL/10 (n = 151), we also observed high incidence (23%; male-to-female ratio ∼3:1) of age-related sarcomas (∼640 d), which mainly arose from proximal hind limb muscles. Also for Dysf −/− mice a strain-dependent effect with respect to mean age of sarcoma-onset was detected, which was more than 100 days later (∼755 d) when the mutation was bred on a mixed C57BL/10 x B6C3Fe background, whereas the sarcoma incidence remained unchanged (22%). Notably, dysferlin-deficiency on the mixed C57BL/10 x B6C3Fe background resulted in a predominant abdominal wall location of sarcomas (Figure 1A, Table 1).
Based on the spontaneous occurrence of skeletal muscle-tumors in mice deficient for the so far molecularly unrelated genes dystrophin and dysferlin, we hypothesized that MD-genes in more general might act as tumor suppressors. To this end, we conducted a life-span study with mice lacking calpain-3, the animal model for LGMD2A in humans. Indeed, also Capn3 −/− mice developed skeletal muscle-derived sarcomas at an incidence of 5%. Finally, we also observed the rare occurrence of sarcoma formation even in myd mice (representing a model for a severe congenital MD in humans), in spite of their considerably short lifespan (Table 1).
In order to test if dystrophin, dysferlin and calpain-3 have tumor-suppressor effects in vivo, we generated double-mutant mouse lines, i.e. dystrophin-deficient (mdx) mice with additional lack of either dysferlin (Dmd −/− Dysf −/−) or calpain-3 (Dmd −/− Capn3 −/−). Dmd −/− Dysf −/− mice (C57BL/10) clinically presented with significant weakness characterized by severe dystrophic signs in the skeletal muscle (R.B., manuscript in preparation), and had a severely reduced life-span of ∼13 months. Remarkably, malignant skeletal muscle-derived sarcomas (Figure 1B) constituted the main cause of premature death in this condition. While penetrance was sharply increased in male mice, 63% of which developed sarcomas, a dramatic decrease in tumor latency was observed in both genders, with the mean age-of-onset reduced to ∼390 d (compared to 540 d in Dmd −/− and 640 d in Dysf −/−).
The combined effect of Dmd −/− Capn3 −/− in double-knockout mice, which also presented with a severe MD-phenotype leading to a shortened life-span of ∼13 months (R.B., manuscript in preparation), resulted in spontaneous sarcoma-formation in 44% of the animals with a mean-age of onset of ∼390 days (Figure 1C). Thus, additional loss of both dysferlin and calpain-3 in dystrophin-deficient mdx mice dramatically reduced sarcoma latency (Figure 1D).
Because the macroscopic appearances of the skeletal muscle-derived tumors showed areas of different colorings and varying consistencies (Figure 1E), we speculated that this might be due to a mixed composition of diversely differentiated tumor-cell lineages. Indeed, careful histopathological examinations revealed that all tumors independent from the underlying MD mutation(s) resembled mixed sarcomas, comprising variably sized coexisting compartments of rhabdomyosarcoma (RMS), fibrosarcoma (FS) and liposarcoma (LS), respectively (Figure 1F–1K). Histopathology of RMS mainly presented as embryonic (ERMS) or spindle-cell tumors, which expressed myogenic factors to various degrees, such as myogenin (Figure 1G), Myf5, or desmin (not shown). Also at the ultrastructural level, these tumor compartments were composed of cells with myofibrils, which were partly arranged in sarcomeric manner (Figure 1L). Tumor-compartments identified as FS displayed bundles of collagen fibres and immature, proliferating fibroblasts, which were arranged in a typical herringbone pattern, hereby recapitulating the histopathological hallmark of human FS (Figure 1F, 1I). The third identifiable compartment was LS consisting of lipocytes, which showed both well-differentiated and de-differentiated morphologies. The unifying characteristics of all LS-cells were positivity for lipid staining by Sudan Black (Figure 1H) and, moreover, immunoreactivity for Cdk4 (a human LS biomarker) (Figure 1K). By electron microscopy, LS-cells characteristically contained numerous fat droplets (Figure 1M). In line with the histological findings, propagated tumor cell cultures also revealed co-existence of different cell types (Figure 1N), most prominently myogenin-positive cells and lipogenic but myogenin-negative cells (Figure 1O), providing further support that tumors arising in MD-mice are mixed-type sarcomas. Because these findings disclosed all MD-tumors as complex, mixed sarcomas, we next studied the expression of select human sarcoma-related genes [13]. Indeed we found increased expression levels for RMS-markers (Myog, Myl4, Igf2, Prox1), a FS-gene (Vcan), and LS-related genes (Pparg, Myo1e, Hoxa5, Plau), which further established the MD-tumors as mixed sarcomas consisting of RMS, FS and LS-compartments (Figure S1).
In order to characterize the emerging link between MD and sarcoma susceptibility, we investigated genetic lesions in tumors originating in our MD-mouse strains. DNA extracted from solid tumors from Dmd −/− or Dysf −/− mice was subjected to an arrayCGH-based screen (n = 8), which revealed that the majority of Dmd −/− tumors were characterized by multiple segmental chromosomal changes, chromosome number aberrations, and amplification of loci harboring the Met (encoding the Met proto oncogene hepatocyte growth factor receptor) or Jun oncogene, while tumors from Dysf −/− mice typically displayed less genomic instability (Figure 2A). Frequent disruption of the tumor suppressor loci Cdkn2a, encoding p16INK4a and p19ARF, Nf1, encoding neurofibromin 1, and Trp53, together with whole chromosome 8 and/or 15 gains represented key non-random alterations of sarcomas in both MD models.
Quantitative PCR (qPCR) experiments (Figure 2B and 2C) of DNA extracted from tumors of Dmd −/−, Dysf −/−, Dmd −/− Dysf −/−, and Dmd −/− Capn3 −/− mice (n = 98) revealed that these genetic lesions were common but occurred at different degrees, depending on the specific gene defect(s). Frequent amplification of Met or Jun oncogenes was observed in Dmd −/− (41%) and Dmd −/− Dysf sarcomas (44%). In contrast, amplifications of Mdm2 and/or Cdk4 (which were additionally tested because of their frequent amplification in human sarcomas, most prominently liposarcomas [14]) were rare (<5%; not shown).
Lesions of the Nf1 gene (exons 23 and/or 56) were more frequently found in Dmd −/− (34%) and Dmd −/− Dysf −/− sarcomas (31%) as opposed to Dysf −/− (14%) or Dmd −/− Capn3 −/− (18%). Conversely, exon 2 of the Cdkn2a tumor suppressor gene, which encodes parts of both p16INK4a and p19ARF, was reduced in 73% of Dysf −/− tumors whereas ∼50% of Dmd −/− Dysf −/− and Dmd −/− Capn3 −/− tumors carried this deletion. Notably, many of the qPCR-ratios obtained for Cdkn2a and Nf1 were consistent with losses throughout the tumor. In 25% of DNA samples from sarcomas with qPCR values indicating Cdkn2a loss, exon 2 copy numbers were <0.2, which suggested the presence of a homozygous deletion in ∼80% of tumor cells, compatible with an early event in tumorigenesis.
Based on the arrayCGH-findings we screened a large cohort of tumors also for chromosome 8 and 15 copy number aberrations. We found gains of either or both chromosomes in the vast majority (80%) of sarcomas. While ∼40–60% of tumors from all MD models displayed gains of both chromosomes, chromosome 8 alone was preferably gained in Dysf −/− and chromosome 15 in Dmd −/− tumors indicating a probable MD-specific preference (Figure 2C).
More than 50% of the measured chromosome 8/15 ratios were consistent with gains throughout the tumor, implying the presence of trisomies in more than 90% of tumor cells. This suggested that together with losses at Cdkn2a and Nf1 loci the recurrent duplications of these chromosomes belong to early events in sarcoma development. Thus, we argued that such events might occur in skeletal muscles of MD-mice prior to formation of clinically identifiable tumors.
To test this hypothesis, we assessed chromosome 8 and chromosome 15 copy numbers in DNA samples extracted from a panel of typically tumor-prone limb muscles (n = 101), which were obtained from different animals (n = 31) that were sacrificed at advanced ages (comparable to the mean age of mice with sarcomas in the respective MD models) but had not developed visible tumors until then (Figure 3A). We found elevated levels of chromosome 8 and/or 15 in ∼30% of muscles from MD-mice but never in wild-type mice (Figure 3B). Also occasional copy number aberrations of the Cdkn2a, Nf1, Met and Jun genes were detected in dystrophic muscles (∼12%). Because the extents of some of these findings were clearly compatible with the presence of malignant cell clones within the tested muscles, we next analyzed these muscles microscopically. Indeed we found variably sized microscopic tumor masses residing between muscle groups and within single muscle fascicles (Figure 3C). Immunohistochemical examination of these tumors in situ revealed intense staining of cell proliferation markers (p27, PCNA) as well as Cdk4 (Figure 3D), compatible with high proliferative activity.
These findings clearly showed (i) that tumor pre-stages and pre-neoplastic lesions are already present in dystrophic muscle and (ii) that the actual sarcoma incidence of MD-mice is much higher than that solely based on the occurrence of visible tumors. Because none of these DNA-abnormalities were present in non-muscle tissues (i.e. brain, liver, and lung) we concluded that these somatic aberrations are specific to dystrophic skeletal muscle.
To address whether aneuploidy affects also human MD, we analyzed primary myoblast lines from DMD and LGMD2B patients. In myoblast DNA samples from a DMD and a LGMD2B patient as well, arrayCGH revealed profiles indicating borderline gains of several chromosomes. In particular, aberration scores indicated gains of chromosome 19 (Figure 4A). To confirm this finding, interphase fluorescent in situ hybridization (I-FISH) experiments were performed on cytospin preparations from early-passage myoblast cell cultures of DMD (n = 4), and LGMD2B (n = 3) patients, as well as healthy donors (n = 2). In contrast to normal cells, myoblasts from DMD and LGMD2B patients frequently harbored tetrasomies of chromosome 19 (13–27%; Figure 4B). Additional analyses for other chromosomes revealed multiple aneusomies, such as tri- and tetrasomies of chromosomes 1 (5–8%), 2 (3–6%), and 8 (4–20%; Figure 4C). In a DMD myoblast cell line, metaphase spreads displayed formation of diplochromosomes (i.e. pairs of sister chromosomes, generated by endoreduplication) (Figure 4D), which are indicative for heterogeneous chromosomal instability and aneusomies. DNA content analyses by FACS profiling of propidium iodide-stained cells revealed that myoblasts from DMD and LGMD2B patients contained abnormally high proportions of nuclei with aberrant DNA-content, indicated by prominent G0+ peaks (Figure 4E). Targeted FISH analysis of nuclei isolated through sorting of such G0+ peaks verified the presence of genomes harboring chromosome 8 aneusomies (Figure 4E, insets). Moreover, the occasional presence of micronuclei implied the continual induction of numerical or structural chromosomal damage in MD-myoblast lines.
In order to preclude that the observed chromosomal copy number aberrations had been acquired or at least amplified in vitro, as reported for embryonic stem cells [15] and committed progenitor cells [16], we asked whether aneusomies also represent an in vivo genotype and do exist in skeletal muscle tissue of MD patients. To this end, interphase nuclei from frozen muscle biopsies from human MD-patients were isolated and probed by I-FISH. We detected tri- and/or tetrasomies of chromosomes 2 and/or 19 in ∼5–12% of the nuclei isolated from DMD muscle (n = 4) (Figure 4F). In contrast, counts of chromosome 13, for which normal copy numbers were found in myoblasts, were readily comparable to control muscles (Table 2). Similarly, aberrant chromosome 2 and 19 counts were detectable in muscle biopsies from patients with LGMD2A (n = 3, CAPN3 mutations), LGMD2I (n = 3, FKRP mutations), as well as LGMD2B (n = 1, DYSF mutations) (Figure 4G). Notably, LGMD2A muscles exhibited slightly aberrant counts also for chromosome 13 (4.6% versus 1% in controls). Generally, poly-/aneusomic nuclei further displayed features like enlargement, more irregular shape, and micronucleus formation, when compared to disomic nuclei. I-FISH signals in nuclei with aneusomic configurations frequently appeared either as highly condensed doublet signals (in particular for chromosome 19) or as bizarre structures with highly elongated conformation, indicating increased variability of differential (probably abnormal) states of chromatin condensation (Figure 4F, 4G). In order to learn if the degree of aneusomies correlates with the disease progression of muscular dystrophies, we also studied fetal muscle obtained during autopsy of aborted fetuses with prenatal diagnosis of DMD or MDC1C. Indeed, these fetal muscle tissues contained much less chromosomal copy number aberrations (chr2: ∼1% versus 0.2% in controls; chr13: 0.6% versus 0%; chr19: ∼3% versus 1%). Thus, compared to age-matched control muscles, we observed an age-dependent increase of the frequency of aneusomic nuclei in MD patients (Figure 4H, 4I).
The finding of cancer-like mutations and somatic aneuploidy in dystrophic muscle prompted us to speculate that this might be caused by damage to DNA induced e.g. by oxidative or replication stress. The formation of interstitial deletions and intrachromosomal amplifications, which we found in pre-neoplastic lesions and sarcomas arising in murine MDs, belong to typical genetic aberrations that result from unrepaired DNA double-strand breaks (DSB) [17] and represent early events in the development of cancer [18]. To explore whether damage to genomic DNA precedes sarcoma development, we studied the canonical DNA damage response pathways in skeletal muscle from dystrophic mice. When analyzing muscle tissue from Dmd −/− mice, pronounced activation of the two major DNA damage response pathways was observed, characterized by high expression of Ser1981-posphorylated ATM (p-ATM, ataxia-telangiectasia mutated kinase) and Ser428-posphorylated ATR (p-ATR, ATM and Rad3-related), and of their downstream signaling targets Chk1 and Chk2 (not shown). We next investigated histone H2A.x, which represents a target of the ATM pathway that signals the presence of DSBs and constitutes a key protein of the DNA damage response by accumulating at large stretches of chromatin surrounding DSBs and recruiting repair factors [19]. In contrast to normal controls, muscle from MD mice was characterized by intense immunoreactivity with an antibody specifically detecting Ser139-phosphorylated histone H2A.x (γ-H2A.x), similar to the reactivity observed in sarcomas (Figure 5A–5C).
We then examined the DNA damage response in muscle biopsies obtained from human DMD patients. In contrast to healthy control muscles, γ-H2A.x immunostainings revealed high levels of DSBs in muscle biopsies from all DMD patients (n = 4) tested, with multiple nucleoplasmic foci formation belonging to muscle fibres and moreover to non-muscle cells within the endomysial connective tissue, such as interstitial fibroblasts and endothelial cells (Figure 5D, 5E). We further found that DNA-damage response was already present in pre-pathologic muscle from very young patients (9–11 months) and a DMD fetus, which suggested that DNA-double strand breaks very likely occur prior to clinical onset of muscle weakness, wasting, and the concomitant inflammatory response. Also muscle tissue in samples from LGMD2A (CAPN3, n = 3), LGMD2I (FKRP, n = 2), MDC1C (FKRP), and LGMD2B patients (DYSF) exhibited intense γ-H2A.x immunoreactivity and multiple nucleoplasmic foci formation (not shown). That both muscle-fiber nuclei and non-muscle cell nuclei displayed massive γ-H2A.x accumulation prompted us to specifically assess the DSB response in myogenic precursor cells. We investigated primary muscle cell cultures generated from DMD and LGMD2B patients. In contrast to myoblasts from healthy donors, nuclei from DMD and LGMD2B myoblasts showed pronounced accumulation of γ-H2A.x foci (Figure 5F–5I). The formation of distinct nuclear immunofluorescent foci was observed in 49% of cells from DMD and 59% from LGMD2B myoblasts (compared to 24% in controls) and the number of cells with multiple (≥3) foci was also markedly increased (DMD: 32%; LGMD2B: 45%; controls: 10%).
Here we show that different types of MD mouse models develop with increasing age mixed soft-tissue sarcomas (STS), presenting as rhabdo-fibro-liposarcomas. While the spontaneous occurrence of RMS has been previously reported in mdx mice [20] and in addition in mice deficient of α-sarcoglycan [21] (Sgca −/−, a model for the human LGMD2D), this is the first report of sarcomas in mice lacking dysferlin, calpain 3, or Large. Our work further shows for the first time that also mice lacking dystrophin due to other mutations than mdx and on different genetic backgrounds are prone to develop age-related STS. In contrast to the previous reports, we found that virtually all sarcomas from MD mice histologically present as mixed sarcomas consisting of RMS and of two additional components with fibro- and liposarcomatous differentiation. Macroscopically, sarcomas feature considerable heterogeneity regarding visual appearance and consistency of tumor mass. Similar to the high complexity and histological diversity inherent to human sarcomas, we found it extremely difficult to exactly stage individual tumors due the highly complex and heterogeneous structure and significant sectional plane divergence. Therefore, our finding of mixed sarcomas in mdx and other MD mice rather extends than rebuts the previous reports by Chamberlain et al. [20], who reported alveolar RMS in mdx, and Fernandez et al. [21], who described embryonal RMS in mdx and also Sgca −/−mice. As a further difference, sarcoma incidence in our C57BL/10 mdx mice (39%) was clearly higher compared to the previously reported RMS incidences (∼6–9%). It remains elusive if these differences are due to different housing conditions or other unknown environmental or strain-specific factors.
It is, however, remarkable that the three main components of malignant cell-types, i.e. myo-, fibro-, and lipocytes, which we observed in our MD-mouse tumors, correspond exactly to the same cell- and tissue types that are crucially characterized by progressive proliferation in MDs. Thus, the MD-associated proliferation of fat and connective tissue might create the molecular context permitting sarcoma development arising from a multipotent mesenchymal or muscle-derived stem cell.
Several observations in our study lend support to the speculative view that MD-genes might have a role as tumor suppressors. We found that strain backgrounds with C57BL/6 proportions obviously exerted protective effects with regard to tumor latency and that tumor penetrance was lower in Dmd −/− mice on C3H or BALB/c backgrounds compared to C57BL/10. In line with our observation, C57BL/6 is known for its resistance to Ptch1+/−-induced rhabdomyosarcomas [22]. Genetic background also clearly influenced tumor gender specificity in Dmd – mice (male preference in BALB/c, female in C3H) and tumor site predilection in Dysf −/− mice (∼60% abdominal wall tumors in C57BL/10 x B6C3Fe compared to ∼20% in C57BL/10). Such strain-specific modulation of incidence, latency, location spectrum, and gender preference has been well documented for other cancer models, such as the p53-deficient mouse [23]. The significantly reduced sarcoma latency in double-mutant Dmd −/−Dysf−/−- and Dmd −/−Capn3 −/− mice also resembles a common feature of tumor suppressor mouse models, as exemplified by the synergistic effect of a combined loss of p53 and Nf1, which accelerates soft-tissue sarcoma development [24]. Thus, the effects we observed for MD-gene losses represent classical credentials of tumor suppressor genes. In support of this view, dystrophin has been linked to human cancer, as its frequent inactivation was shown to be involved in the pathogenesis of malignant melanoma [25]. Notably, in melanoma cell lines dystrophin knock-down enhanced migration and invasion, whereas re-expression attenuated migration and induced a senescent phenotype, fully in line with a tumor suppressor role of dystrophin [25]. Moreover, utrophin, the highly related autosomal paralogue of dystrophin, represents a tumor suppressor candidate, owing to its frequent disruption in human malignant tumors and its capability to inhibit breast cancer cell growth [26]. Notably, aberrations of the DG have been associated with several types of human cancer [27]–[30], suggesting a potential role also in tumorigenesis. In particular, a tumor suppressor function has been suggested for laminin-binding glycans on α-dystroglycan [31], whose loss can be caused by silencing of the LARGE gene in several metastatic epithelial cell lines [30]. For both, dystrophin [32] and dysferlin [33] interactions with the microtubule network have been recently described, which suggests their hypothetical implication in microtubule-mediated cell functions, such as mitosis and cell migration. Future studies will be needed to clarify whether MD-genes act as tumor suppressors, which is suggested but not proven by our data.
We found that murine sarcomas from MD-mice frequently harbor non-random, recurrent genetic lesions that provide links to human mesenchymal cancers. The pivotal p53 and retinoblastoma (RB) cell cycle control pathways were frequently incapacitated by the disruption of the Cdkn2a locus, which encodes two different tumor suppressors, the Cdk4 kinase inhibitor p16INK4a and the Mdm2-p53 regulator p19ARF, both of which play an important role in the development and progression of many human cancer types. Deletions at Trp53 and Nf1 loci established a genetic link to human soft-tissue sarcomas, which are characterized by frequent p53 mutations [34]–[36], as well as to syndromes associated with increased RMS incidence due to germ-line disruption of these tumor suppressor genes (Li-Fraumeni, TP53; Neurofibromatosis type I, NF1) [37]. More recently, human myxofibrosarcoma and pleomorphic liposarcomas were shown to frequently harbor NF1 mutations [14]. Thus, the disruption of Nf1 in sarcomas from MD-mice parallels specific - non myogenic - subtypes of human soft-tissue sarcomas and suggests a more general role for Nf1-lesions in the genesis of mesenchymal cancers. A high fraction of sarcomas from MD-mice harbored amplifications of the Met or Jun oncogenes. The Met oncogene amplification constitutes a critical path to aberrant activation of the Hgf/c-Met axis, which is known to promote tumorigenesis and to be involved in the progression and spread of multiple human cancers. Amplification of the JUN oncogene has been reported in human liposarcomas [38]–[39], in sound accordance with herein discovered frequent Jun amplification in MD mixed sarcomas.
Our finding of recurrent chromosome 8 and/or 15 gains in MD sarcomas provides a link to other murine cancers. Chromosomes 8 and/or 15 are frequently duplicated in T cell tumors [40]–[41] or transgenic mouse models of acute promyelocytic leukemia [42], and probably contribute to elevated expression of the Junb and/or Myc oncogenes, as suggested for Myc in T cell lymphomas [41]. Notably, several human malignancies, amongst them myxoid/round cell liposarcoma [14], are known to harbor recurrent gains of chromosome 8. Most importantly, the human chromosome 8 harbors multiple regions that are syntenic to both murine chromosomes 8 and 15, which we found to be regularly gained in the MD-sarcomas.
We discovered that the most frequent and most prominent genetic alterations that characterize full-blown skeletal muscle-derived sarcomas are already present in dystrophic skeletal muscle of clinically tumor-free mice. We also demonstrated DNA damage and showed that skeletal muscle of MD-mice harbors microscopic tumor infiltrates prior to the development of macroscopically visible tumors. In particular, our findings suggested that somatic aneuploidy, indicated by recurrent gains of chromosomes 8 and 15, contributes to sarcoma susceptibility in murine MD. Thus, the frequent occurrence of chromosome 8/15 gains together with specific losses at the Cdkn2a locus might represent early events occurring in cancer pre-stages and promoting malignant transformation [43]. Importantly, these findings also suggested that the actual sarcoma incidence of MD-mice is much higher than that solely based on the occurrence of visible tumors. In the light of our results, sarcoma formation might be regarded as the disease end-stage of a MD in mice.
The finding of cancer-like genomic aberrations and DNA damage in the skeletal muscle from MD-mice inspired us to search for such aberrations in skeletal muscle of human MD patients. We focused on DMD and LGMDs caused by DYSF, CAPN3, or FKRP mutations, representing the most frequent MDs, and found that all of them are associated with somatic aneuploidy and widespread DNA damage in skeletal muscle tissue in vivo. Also in vitro, cultured myogenic stem cells from DMD and LGMD2B patients exhibited DNA damage and aneuploidy. In our study, somatic aneuploidy appeared to be a feature concurring with the outbreak of pathology in dystrophic muscle and to increase with age in human MD patients. In contrast, high levels of DSBs were already evident in fetal muscle from DMD and MDC1C individuals and in muscle biopsies from DMD infants (<1 year), which suggested that DNA damage precedes the clinical manifestation and therefore cannot be solely related to replication stress. While somatic aneuploidy has been reported in multiple human pathologies, such as Alzheimer's disease, this is the first report on gross somatic aneuploidy in MDs. Genomic instability has been reported in laminopathy-based premature ageing [44], a condition caused by mutations in lamin A/C, notably another MD-related molecule. DNA damage was shown recently in Friedreich's ataxia, a neurodegenerative disorder [45].
Depending on the context, aneuploidy not only can promote tumorigenesis [18], [46]–[47], but also can impair proliferation, cause premature replicative senescence [48], or can even suppress tumorigenesis [49]. Under the assumption that aneuploidy affects cells destined for muscle regeneration and/or function, aneuploidy could therefore represent an important pathological feature causing a propensity for malignant transformation in murine MDs and contributing to tissue malfunction and diminished regenerative capacity in human MDs. Unrepaired DNA damage activates cellular senescence [50] and could therefore be also associated with the known generalized diminished replicative capacity of DMD myoblasts [51], contributing to the progressive exhaustion of the muscle's regenerative potential [52]. In the murine condition, senescence could also underlie sarcoma susceptibility as secreted senescence associated factors can contribute to a pro-tumourigenic inflammatory environment [50], thereby promoting the occurrence of age-related cancer [53]. In this context, it will be interesting to study sarcoma formation in mdx mice lacking the RNA component of telomerase (mdx/mTR) that have very recently been shown to have shortened telomeres in muscle cells and a severe progressive muscular dystrophy [52].
For the time being, we have no answer for why murine MDs frequently end up in sarcoma formation while in human MD patients increased muscle-tumor susceptibility has not been reported. But it is interesting to note that it is also not fully understood why loss of dystrophin causes a fatal MD in humans while only a mild myopathy in mice. Also, Dysf −/− and Capn3 −/− deficient mice are largely spared the severe symptoms of the patients with LGMD due to defects in these two genes. We speculate, however, that essential differences in tumor biology between men and mice could account for this difference: while humans are prone to epithelial carcinomas, mice commonly develop mesenchymal sarcomas, which might be due to profound differences in telomere biology between the two species [54]. Also, fewer genetic events are required to induce malignant transformation in mice compared to humans [43], [55].
Collectively, our findings that genetically distinct MDs in mice and humans share a common molecular pathology characterized by DNA damage and genomic instability similar to pre-cancerous lesions suggests the existence of a novel, unifying pathomechanism that might contribute to disease progression through erosion of the replicative capacity of muscle stem cells and could therefore help to explain the common fatal progression of degeneration and wasting in MD. This is a novel aspect, which contributes to our understanding of MD, and moves an orphan disease close to the common disease cancer, thereby hopefully opening novel therapeutic avenues.
Samples for this study were collected from diagnostic skeletal muscle biopsies, which had been conducted in patients assigned for evaluation of musculoskeletal disorders at our department. Patients or their legal guardians gave informed consent for scientific purpose use of left-over tissue samples. DMD patients included in this study had a confirmed molecular diagnosis of DMD, ascertained by lack of dystrophin staining in immunohistochemistry (IH) and Western Blot (WB), and in most cases a genetic diagnosis. Muscle biopsy samples used in this study were from a total of n = 6 different DMD patients: M2006 (age at muscle biopsy: 9 m; DMD gene mutation: c.3053_3087del), M2008 (11 m; c.8669-1G>T), M1633 (6 a; c.858T>G p.Tyr286X), M1994 (7a; unknown DMD mutation), M1895 (8 a; dup_ex3-7), M1959 (15 a; del_ex17), and two samples from aborted fetuses with DMD. LGM2DA patients (n = 3) had a genetic diagnosis and WB exhibited absence of calpain-3 specific bands in muscle tissue: M1883 (9 a; c.550delA p.Thr184ArgfsX36), M2207 (13 a; c.550delA), M2219 (25 a; c.1342C>T p.Arg448Cys). LGMD2I patients (n = 3) had a confirmed diagnosis by FKRP gene sequencing: M1787 (10 a; c.854A>C p.Glu285Ala), M2190 (28 a; c.826C>A, p.Leu276Ile), M2166 (prenatal; c. [962C>A]+[1086C>G] p.Ala321Glu + p.Asp362Glu). LGMD2B in one patient was confirmed by reduced dysferlin reactivity in IH and WB, and DYSF gene sequencing: M2057 (62 a, c.509C>A p.Ala170Glu). All tissue samples were snap-frozen in dry ice-cooled 2-methylbutane within 1 h after biopsy and stored at -80°C until use.
Primary myoblast cultures were obtained from the Muscle Tissue Culture Collection, Friedrich-Baur-Institute, Department of Neurology, Ludwig-Maximilians-University Munich (Germany). DMD: “Essen 88/07” (14 a, del45_50); “72/05” (7 a, dup_ex8-29); “Essen 8/02” (4 a, del_ex51-55); “166/00” (6 a, 2bp-deletion in exon 6); LGMD2B: “90/01” (36 a, female, c. [638C>T]+ [5249delG]); “176/01” (32 a, male, c. [2367C>A]+ [5979dupA]); “362/03” (male, 33 a, c. [exon 5 p.Pro134Leu]+ [5022delT]); controls: “363/07” (21a, male); “179/07” (21a, female). Cells were maintained in Ham's F-12 medium supplemented with 15% fetal bovine serum, GlutaMax (L-glutamine 200 mM), glucose (6.6 mM), fetuin (0.47 mg/mL), bovine serum albumin (0.47 mg/mL), dexamethasone (0.38 µg/mL), insulin (0.2 µg/mL), epidermal growth factor (10 ng/mL), Pen-Strep (penicillin G 5000 units/mL, streptomycin 5 mg/mL), and fungizone (amphotericine B 0.5 µg/mL) at 37°C in a humidified atmosphere of 5% CO2: 95% air. For experimental purposes, cells were harvested after 3 or 4 passages. DNA was isolated using the QIAamp DNA Mini Kit (Qiagen, Hilden, Germany) according to the manufacturer's recommendations. RNA was isolated using TRI Reagent (Sigma-Aldrich, St. Louis, MO). Cells were stained with BD Cycletest Plus, DNA Reagent Kit for DNA content analysis by flow cytometry (BD Biosciences, San Jose, CA).
Mice stocks were maintained at the Division for Laboratory Animal Science and Genetics (Medical University Vienna, Himberg, Lower Austria) under institutionally approved protocols for the humane treatment of animals. Mice were cared for in our facilities under conventional housing conditions and received food and tap water ad libitum. Mice presenting with weakness received intensive care, were fed with food pellets soaked in tap water, and were examined daily. Aged mice were checked daily for the development of tumors. In general, tumors were characterized by rapid growth, necessitating the killing of affected mice within few days after visual identification of sarcomas. The Dmdmdx C57BL/10 (mdx), Dmdmdx-3cv C57BL/6 (mdx-3cv), SJL Dysfim (SJL-Dysf), and B6C3Fe Largemyd (myd) mice were originally obtained from The Jackson Laboratory (Bar Harbor, ME). To study dystrophin deficiency on other strains, we inbred the Dmdmdx mutation to C3H and BALB/c (>20 consecutive backcross generations; residual heterozygosity <0.01). Some mdx mice were maintained on a mixed C57BL/6 x BALB/c background. Further, we inbred the SJL-Dysf mutation onto the C57BL/10 background, where a prolonged life span compared to SJL was observed, which enabled us to study late-onset stages of dysferlin-deficiency. Capn3 knockout mice (Capn3tm1Jsb) [11] on the 129/Sv x C57BL/6 background were obtained from Isabelle Richard and crossed to C57BL/6 mice. To generate Dmd −/− Dysf −/− and Dmd −/− Capn3 −/− double-mutants, mdx mice were crossed to SJL-Dysf (C57BL/10) and Capn3 knockout mice. Some mdx C57BL/10, mdx-3cv C57BL/6, as well as SJL-dysf C57BL/10 mice were crossed to B6C3Fe myd mice. Of these animals, only Large+/− and Large+/+ mice were used for analysis of tumorigenesis, which were indistinguishable from pure mdx, mdx-3cv, and SJL-dysf mice, respectively. In all cases, Large+/− heterozygosity had no influence on tumorigenesis and conferred no overt additional phenotype with regard to muscle pathology. After sacrifice by cervical dislocation, mice were dissected, and muscles, other tissues, and (where applicable) tumors were excised and snap-frozen in dry ice-cooled 2-methylbutane. All samples were stored at -80°C.
Using sterile techniques, parts of excised tumors were washed in PBS, cut into small pieces, and cultured in primary medium, containing DMEM (Dulbecco's Modified Eagle's Medium, 4.5 g/L glucose; PAA Laboratories, Pasching, Austria), 20% fetal bovine serum (FBS “GOLD” Origin: USA; PAA Laboratories), 200 U/l PenStrep (Penicillin, Streptomycin; Lonza, Cologne, Germany) and 2.5 µg/ml Fungizone (Gibco, Invitrogen Ltd, Paisley, UK). After sporadic adhesion of tumor cells, remaining tissue parts were removed, and the primary medium was replaced by growth medium (DMEM, 20% FBS, 50 U/l PenStrep). To activate differentiation, FBS was replaced by 2% horse serum (PAA Laboratories). DNA/RNA extraction was performed as described above for cultured myoblasts.
DNA was isolated from serial 5 µm-cryosections prepared from dissected skeletal muscle (∼5 mg) or tumor (∼10 mg) specimens. Reference sections from the sampling procedure were HE-stained for histomorphological examination. Sections were stored at −80°C and then subjected to tissue lysis and nucleic acid purification according to the QIAamp DNA Mini Kit protocol (Qiagen). Mouse tail DNA was isolated using the same protocol starting from lysates prepared by directly lysing 2–3 mm tail tips. DNA concentrations were measured using the NanoDrop spectrophotometer (Peqlab, Erlangen, Germany), DNA samples were diluted (10 ng/µl) and stored at −20°C until use. RNA was extracted from serial 10 µm-cryosections by lysis in 1 TRI Reagent (Sigma-Aldrich), chloroform extraction, and precipitation with isopropanol. RNA samples were measured by spectrophotometry (NanoDrop) and quality controlled using BioAnalyzer LabChips (Agilent Technologies, Santa Clara, CA).
Cryosections were stained with haematoxylin and eosin (HE). Sudan Black B was used for lipid staining. For immunohistochemistry, 10 µm cryosections were fixed using 3.7% paraformaldehyde (5 min), treated with 0.1% Triton-X100 (5 min), rinsed in PBS, and subsequently incubated with primary antibodies. For immunocytochemistry, cytospins were prepared from cell suspensions and subjected to methanol/acetic acid (3:1) fixation before antibody incubation. Primary antibodies used in this study were as follows: Myogenin (Santa Cruz Biotechnology, CA; sc-576), Myf-5 (sc-302), desmin (Millipore, Billerica, MA, MAB3430), Cdk4 (sc-260), PCNA (sc-7907), p27 (sc-776), p-Ser1981-ATM (Cell Signaling Technology, Danvers, MA; #4526), phopsho-Ser428-ATR (#2853), p-Ser296-Chk1 (#2349), p-Thr68-Chk2 (#2661), p-Ser139-Histone H2A.X (#9718). Secondary antibodies were conjugated to Alexa-Fluor 488, Alexa-Fluor 594 (Molecular Probes Invitrogen, Carlsbad, CA), Cy3 (Dianova, Hamburg, Germany), or to horseradish peroxidase. Where indicated, immunostained sections were counterstained with 4′,6-diamidino-2-phenylindole (DAPI) and then analyzed by confocal microscopy using either Olympus Fluoview or Zeiss Axioplan2 microscopes. Fully automated software-assisted quantification of DNA damage (γ-H2A.x foci) in myoblasts was performed using the software Metafer (MetaSystems, Altlussheim, Germany). Graphical representations (plots of fluorescence intensity versus foci numbers) were generated in R.
Matched pairs of sarcoma and tail-tip DNA (as reference) samples from the same mice were analyzed using the Agilent mouse genome CGH 44K (design ID 015028) and 244K (014695) oligonucleotide microarrays (Agilent Technologies). Human myoblast DNA samples were analyzed on Agilent human genome CGH 44K arrays (014950), using as reference human genomic DNA from multiple anonymous male donors that was purchased from Promega (G147A; Madison, WI). Labeling and hybridization procedures were performed according to the instructions provided by Agilent. In brief, 200 ng of test and reference DNA were digested with AluI and RsaI (both Promega) and then subjected to differential labeling by random priming with incorporation of either Cyanine 3- or Cyanine 5-dUTP (PerkinElmer, Waltham, MA) using the BioPrime Array CGH Genomic Labeling System (Invitrogen, Carlsbad, CA). After purification with Microcon YM-30 centrifugal filter units (Millipore), the labeled products were combined, mixed with blocking agent, Hi-RPM hybridization buffer (both included in the Oligo aCGH/ChIP-on-Chip Hybridization Kit, Agilent), human Cot-1 DNA (Roche Diagnostics, Mannheim, Germany) or mouse Cot-1 DNA (Invitrogen), and hybridized onto respective microarray slides. Hybridization was carried out for 48 h at 65°C in a hybridization oven. Slides were washed according to the protocol by Agilent, scanned using the Agilent Technologies Scanner G2505B and analyzed using the Feature Extraction and Genomic Workbench 5.01 (formerly DNA Analytics 4.0) software.
To screen for Met (chromosome [chr] 6) and/or Jun (chr 4) as well as Cdk4 and/or Mdm2 (chr 10) oncogene amplification, tumor DNA samples (25 ng) were subjected to a quantitative endpoint PCR, consisting of 0.4 µM each primer, 0.2 mM dNTPs, 1.5 mM MgCl2, (NH4)2SO4-containing amplification buffer, and 0.5 units Taq DNA polymerase (reagents from Fermentas, St. Leon-Rot, Germany). Competitive co-amplification of internal control targets (with similar amplicon size) allowed the unambiguous determination of ≥4-fold amplification levels. Primer sequences (5′->3′) were as follows: Met_f AAC TGT TCT TGG AAA AGT GAT CGT; Met_r TTT GAA ACC ATC TCT GTA GTT GGA; S100a8_f CGT TTG AAA GGA AAT CTT TCG TGA; S100A8_r TAT CCA GGG ACC CAG CCC TA; Jun_f AAA GCA GAC ACT TTG GTT GAA AG; Jun_r CGC TAT TAT AAA TAT GCA CAA GCA A; Mdm2_f CAT CGC TGA GTG AGA GCA GA; Mdm2_r AAG ATG AAG GTT TCT CTT CTG GTG; Cdk4_f AGT TTC TAA GCG GCC TGG AT; Cdk4_r TCT CTG CAA AGA TAC AGC CAA C; Lig3_f AGG AGA GAA GCT GGC TGT GA; Lig3_r AGC TTT CCT TCC TCT TTG CC. After cycling (3 min 95°C, 30× [40 sec 95°C, 40 sec 60°C, 1 min 72°C], 3 min 72°C), 5 µl aliquots of reaction products were analyzed on ethidium bromide-stained 1.5% agarose gels and quantified from captured images using Image J. Relative Met, Jun, Cdk4, and Mdm2 copy number levels were calculated by normalization to the internal standard (S100a8 on chr3 and Lig3 on chr11, respectively). Tumor samples with copy numbers indicating oncogene amplification were also subjected to verification by real-time PCR (see below). Deletions at the Cdkn2a (chr 4) and Nf1 (chr 11) loci were measured using a quantitative real-time PCR (qRT-PCR) SybrGreen assay (ΔCt method), involving separate amplification of target genes and an internal reference (Lig3). Primers were designed for Cdkn2a exon 2, which encodes parts of both p16INK4a and p19ARF. CGH 244K data from one Dysf -/- tumor revealed a compound loss at the Nf1 locus consisting of a large ∼0.4 Mbp deletion encompassing the whole gene and a smaller ∼42 kb deletion spanning exons 9–28. To screen for Nf1 deletions in other tumors, two different exons (23 and 56) were chosen as qRT-PCR targets. Primer sequences were as follows: Cdkn2a_f: GTA GCA GCT CTT CTG CTC AAC TAC; Cdkn2a_r AAT ATC GCA CGA TGT CTT GAT GT; Nf1_I22_f TGA TGA AGT AGT TTG CCA TTG TTT; Nf1_E23_r TTG CCA TCA TGA CTT CAA CTA ACT; Nf1_I55_f CTC TCG CTC TTC ATT TCA TCT TCT; Nf1_E56_r GCC ATA AGC CAT TAA AAC CAA AAC. Met and Jun targets were the same as above. 15 ng of DNA template were amplified in the presence of 0.5 µM primers and components of the SensiMixPlus SYBR universal mix (Quantace, London, UK) using the Stratagene Mx3005P cycler (Agilent). Cycling conditions: 10 min 95°C, 40× [40 sec 95°C, 40 sec 60°C, 1 min 72°C], followed by a dissociation segment for melting curve analysis. Chromosome 8 and 15 gains were also assessed by qRT-PCR choosing Junb (chr 8) and Myc (chr 15) as targets, respectively. Gene dosage was normalized to an arbitrary gene on chr 12 (Prima1), whose copy number appeared widely stable in the CGH screen. Primer sequences were as follows: Junb_f GCA GCT ACT TTT CGG GTC AG; Junb_r GTG GTT CAT CTT GTG CAG GTC; Myc_f CCA CCT CCA GCC TGT ACC T; Myc_r GTG TCT CCT CAT GCA GCA CTA; Prima1_f GTT TCC ATA TCT GCA GGT GAC A; Prima1_r CTC TCG TTC ATC AGC TGT TCC T. Reactions were carried out as above but with 30 sec extension steps. Fluorescence data were analyzed using the MxPro 4.1 software (Stratagene). After verification of primer performance, relative quantification was obtained using the threshold cycle method; ΔCt values were calibrated to wild-type (C57BL/10) tail-tip DNA. Plots of ΔΔCt (sarcomas) values were done in R and graphical representations of ΔCt values from skeletal muscle DNA samples were made in Microsoft Excel 2007.
To study whether mixed sarcomas from MD-mice express select human sarcoma-related genes, we subjected RNA isolated from primary tumor samples as well as from tumor cell cultures to quantitative RT-PCR. Total RNA (1 µg) was reverse-transcribed by standard oligo-dT primed cDNA synthesis using M-MuLV Reverse Transcriptase in a reaction buffer containing 50 mM Tris-HCl (pH 8.3 at 25°C), 50 mM KCl, 4 mM MgCl2, 10 mM DTT, and 1 mM dNTPs (Fermentas). An aliquot corresponding to 10 ng of the initial RNA sample was subjected to a quantitative endpoint PCR, consisting of 0.4 µM each primer, 0.2 mM dNTPs, 2 mM MgCl2, (NH4)2SO4-containing amplification buffer, and 0.25 units DreamTaq Green DNA Polymerase (reagents from Fermentas) in a 25 µl reaction volume. Primer sequences for the human rhabdomyosarcoma-marker genes (Myog, Myl4, Igf2, Prox1), a fibrosarcoma gene (Vcan), and liposarcoma-related genes (Pparg, Myo1e, Hoxa5, Plau) are available from the authors on request. After cycling (3 min 95°C, 35× [20 sec 95°C, 20 sec 60°C, 40 sec 72°C], 3 min 72°C), 10 µl aliquots of reaction products were analyzed on ethidium bromide-stained 1.5% agarose gels and quantified from captured images using Image J. Relative RNA abundance was calculated by normalization to the Gapdh transcript levels and compared to skeletal muscle samples isolated from wild-type and mdx mice. RNA abundance in tumor cell lines was compared to murine C2C12 myoblast cells. Visualization of gene expression was accomplished by heatmaps made in R using the heatmap.2 function.
For I-FISH experiments on myoblasts, cells were fixed using 4% formaldehyde. FISH analysis on interphase nuclei extracted from cryofixed tissues was performed according to a previously published protocol with modifications [56]. In brief, thirty 20 µm-cryosections were fixed in PBS-buffered 4% paraformaldehyde (2–3 h at ambient temperature), rinsed twice with 0.9% NaCl and stored at 4°C overnight until further use. Fixed tissue sections were then transferred into a 90 µm nylon mesh and subjected to proteinase K digestion (0.05%; 10–15 min 37°C). After harvesting by cytospinning through the mesh, nuclei were air-dried, fixed with paraformaldehyde solution (4% in PBS), washed with 1× PBS (2×3 min), pre-treated with sodium thiocyanate (1 M, 80°C 1 min), and subjected to digestion with proteinase K (1 min at 37°C). After fixation, slides were air-dried, followed by heating to 78°C (8 min) for denaturing. Slides were then incubated with digoxigenin or biotin labeled chromosome probes (2p, 18cen, 19q from Dr. M. Rocchi, Molecular Cytogenetic Resource Centre, Bari, Italy; chr1 from Dr. Howard J. Cooke [57]; 8cen purchased from Kreatech Diagnostics, Amsterdam, The Netherlands; chr13 FKHR and 19p/19q from Vysis, Abbott Laboratories, IL) for hybridization overnight at 37°C. Slides were washed in 2× SSC 50% formamide, and 2× SSC at 42°C, and incubated with Cy3-labelled anti-biotin (Dianova, Hamburg, Germany) or FITC-labeled anti-digoxigenin antibodies in 2% BSA for 30 min at 37°C in a humid chamber. After washing in 4× SSC 0.1% Tween-20 (2×7 min at 42°C), slides were incubated with secondary antibodies labeled with Cy3 or FITC (Dianova) in 2% BSA for 30 min at 37°C, washed again as above, ethanol-dried, and mounted using Vectashield with DAPI (Vector Laboratories, Burlingame, CA). Slides were analyzed using an Axioplan2 (Zeiss) microscope and I-FISH signals were captured using the ISIS software and quantification of the I-FISH spots was achieved with the Metafer software (both from, MetaSystems, Altlussheim, Germany). For each sample 300 nuclei were automatically detected by the software and subsequently visually inspected by two independent investigators. Data presented were calculated from an average of 200 nuclei eligible for analysis.
|
10.1371/journal.ppat.1001282 | The C-Terminus of Toxoplasma RON2 Provides the Crucial Link between AMA1 and the Host-Associated Invasion Complex | Host cell invasion by apicomplexan parasites requires formation of the moving junction (MJ), a ring-like apposition between the parasite and host plasma membranes that the parasite migrates through during entry. The Toxoplasma MJ is a secreted complex including TgAMA1, a transmembrane protein on the parasite surface, and a complex of rhoptry neck proteins (TgRON2/4/5/8) described as host cell-associated. How these proteins connect the parasite and host cell has not previously been described. Here we show that TgRON2 localizes to the MJ and that two short segments flanking a hydrophobic stretch near its C-terminus (D3 and D4) independently associate with the ectodomain of TgAMA1. Pre-incubation of parasites with D3 (fused to glutathione S-transferase) dramatically reduces invasion but does not prevent injection of rhoptry bulb proteins. Hence, the entire C-terminal region of TgRON2 forms the crucial bridge between TgAMA1 and the rest of the MJ complex but this association is not required for rhoptry protein injection.
| Invasion by the obligate intracellular parasites, Toxoplasma and Plasmodium, requires the formation of a ring of contact between parasite and host plasma membranes, the so-called moving junction (MJ), that the parasite migrates through during entry. The MJ is a complex of secreted parasite proteins including AMA1, on the parasite surface, and several rhoptry neck proteins (RONs), which are reported to associate with the host plasma membrane. The precise nature of the interaction that causes these two membranes to be so tightly apposed has not yet been elucidated. Here we report that the carboxy-terminal region of Toxoplasma (Tg)RON2 is exposed to the extracytosolic face of the MJ and that two short domains (D3 and D4) within this region independently and efficiently interact with the exposed ectodomain of TgAMA1. As recombinant D3, representing just 54 amino acids from TgRON2, efficiently blocks invasion, this interaction represents the crucial linkage for the MJ complex. Interestingly, D3 does not prevent injection of a rhoptry reporter protein demonstrating that invasion, and specifically a functional MJ, is not required for such injection. Our results suggest that the D3–D4 subregion of RON2, which is conserved across the Apicomplexa, will be a potent addition to current, AMA1-based control strategies for malaria.
| Protozoan parasites are a significant cause of morbidity and mortality in humans worldwide. Among the most devastating and globally prevalent parasites are the members of the phylum Apicomplexa, which includes the etiological agents of malaria, cryptosporidiosis, and toxoplasmosis. Apicomplexans are related by an anterior complex of specialized secretory organelles that secrete molecules necessary for active host cell invasion and subsequent development of the parasitophorous vacuole (PV) around the penetrating parasite [1]. Given the obligate intracellular nature of these organisms, invasion of host cells is a critical event in the host-parasite interaction.
In contrast to many intracellular pathogens that use conventional host-uptake pathways to enter a target cell, apicomplexans actively invade in a rapid, multi-step process that is dependent on the parasite actinomyosin machinery [2], [3]. A distinctive feature of this process is the formation of a close apposition between the parasite and host plasma membranes that is reminiscent of a tight junction in mammalian cells [4], [5]. Beginning with its apical end, the parasite moves through this ring-like structure which is referred to as the ‘moving junction’ (MJ) and which functions to generate the PV membrane from the invaginated host plasma membrane [6]. As invasion proceeds, the MJ also appears to act as a molecular sieve that somehow excludes certain host membrane proteins from the forming PV membrane [7].
The identified heteromultimeric protein complex that forms at the MJ is derived from two distinct secretory organelles of the parasite: the micronemes and the rhoptry neck compartment [8], [9], [10], [11], [12], [13], [14]. The micronemal protein AMA1, which has a type I transmembrane topology in the parasite plasma membrane, is the most well characterized molecule of the MJ complex. The importance of this apicomplexan-specific protein in the invasion process has been directly demonstrated in several members of the phylum, including T. gondii [15], [16], Babesia bovis [17], and Plasmodium spp. [18], [19]. In Toxoplasma, incubation of parasites with antisera specific for the large ectodomain of TgAMA1 reduces the frequency of invasion by ∼40% [15] and depletion of TgAMA1 expression using a conditional knockout strain virtually eliminates invasion [16]. Plasmodium AMA1 is a leading malaria vaccine candidate on the basis of several reports demonstrating that antisera targeting the ectodomain of Plasmodium AMA1 block erythrocyte invasion [19], [20], [21] and immunization with recombinant derivatives of Plasmodium AMA1 confer protection against the blood stage in animal models (reviewed in [22]).
Co-immunoprecipitation studies have led to the identification of TgRON2, TgRON4, TgRON5 and TgRON8 as members of the TgAMA1-associating MJ complex [8], [10], [12], [13]. Visualization of TgRON4/5/8 at the MJ has been confirmed [8], [10], [12], [13] but the subcellular localization of TgRON2 during invasion has been enigmatic. Despite biochemical evidence that is consistent with its localization to the MJ, the only visualization of TgRON2 outside of the rhoptry necks is as a secreted form that localizes to the tip of cytochalasin D-treated parasites (cytochalasin D acts to disrupt actin filaments, which are needed for parasite motility, and in this way blocks invasion but does not affect release of rhoptry proteins) [13]. Identification of P. falciparum AMA1-associating proteins demonstrates that the MJ complex composition is conserved, at least in part, in P. falciparum [9], [11], [23], with the notable exception that an orthologue of TgRON8 has not yet been identified in Plasmodium spp.
Characterization of the topology of the RONs during invasion and identification of the specific RON that binds AMA1 are crucial to gaining a better understanding of this unique complex. To this end, it has been reported that TgRON4/5/8 are secreted into the host cell where they localize to the cytosolic face of the host plasma membrane [12], [13]. TgRON2 has three predicted hydrophobic helices that have been postulated to span membranes [8], [10], [13], [24]. Combined with co-immunoprecipitation studies showing that TgRON2 is capable of independently associating with TgAMA1 [13] and TgRON4 [8], these results have led to the model that TgRON2 is the link between TgAMA1 on the parasite surface and the rest of the MJ complex, possibly spanning the host plasma membrane [12], [13].
To determine the role of TgRON2 in the function of the MJ, we generated transgenic parasites that could be used to visualize this protein at the MJ and used subdomains to better understand the TgRON2-specific interactions within the MJ complex. We demonstrate that a small, carboxy-terminal region of TgRON2 that spans what had been thought to be the third transmembrane domain associates with the ectodomain of TgAMA1 and that this interaction is critical for T. gondii invasion of, but not injection into, host cells.
Initial attempts to localize TgRON2 within the MJ complex relied on antibodies to the native protein but were unsuccessful (data not shown; [13]). As an alternative strategy, we introduced an HA-tag to the C-terminus of TgRON2 in RHΔhxgprt parasites by using a vector to homologously recombine a cassette that contains the coding sequence of the HA tag and the HXGPRT gene flanked by targeting sequences from upstream and downstream of the 3′ end of the TgRON2 gene. Following selection for the HXGPRT marker using mycophenolic acid/xanthine, clones were identified and demonstrated by PCR to have the correct, homologously integrated sequences (data not shown). Examination of the resulting TgRON2-HA-expressing parasites by western blot using an HA-specific antibody demonstrated the expression of a single HA-tagged protein at the expected size of mature TgRON2 at ∼145 kDa (Figure 1B). Examination of the TgRON2-HA-expressing parasites by immunofluorescence (IFA) using an HA-specific antibody demonstrated that this fusion protein correctly localizes to the rhoptry necks of intracellular parasites as confirmed by co-localization with TgRON4 (Figure 1C).
To determine the localization of TgRON2 during invasion, TgRON2-HA-expressing parasites were used to infect human foreskin fibroblast (HFF) monolayers under conditions that synchronize invasion [25]. Parasites were permitted to invade for a short time (∼45 seconds) and then infected HFF monolayers were formaldehyde-fixed and permeabilized in triton X-100. Examination of TgRON2-HA parasites that were stalled in a partially invaded state using an HA-specific antibody demonstrated that TgRON2-HA can be visualized at the MJ, as determined by co-localization with the known MJ protein TgRON4 (Figure 1 D and E). These results provide direct confirmation that TgRON2 localizes to the MJ, thus confirming the biochemical data obtained previously that implicated TgRON2 in the MJ complex [8], [10], [12], [13].
Like most of the other identified MJ components, TgRON2 orthologues are present within all Apicomplexa that show MJ formation during invasion, suggesting a conserved function for this protein. To identify regions of TgRON2 that are crucial for interacting with the other members of the MJ complex, therefore, we generated a multiple sequence alignment of TgRON2 and its orthologues in Neospora caninum and several Plasmodium species. We observed that the greatest sequence conservation among these orthologues is in the C-terminal-most third of the protein (Figure 2). To determine if this conserved region of TgRON2 is important for interactions with the remaining members of the MJ complex we generated protein fusions with glutathione S-transferase (GST) and used these in co-affinity purification studies. We chose to use regions outside of the hydrophobic helices as we anticipated that fusion proteins including these regions would not be soluble in E. coli, a prediction that was subsequently confirmed experimentally (data not shown). GST fusions with portions of TgRON2 N-terminal to the second putative hydrophobic helix (HH2), which spans residues 1277 to 1296, were also refractory to purification under soluble conditions (TgRON2 residue numbers are from Genbank accession HQ110093 with residue 1 as the start methionine); however, fusion of GST with TgRON2 amino acids 1293 to 1346 and 1366 to 1479, generating recombinant proteins GST-domain 3 (D3) and GST-domain 4 (D4), respectively, were successfully expressed and purified from E. coli (Figure 2 and data not shown).
To determine if GST-D3 and GST-D4 are sufficient to co-purify any members of the MJ complex, molar equivalents of the GST fusion proteins as well as GST alone were bound to glutathione sepharose and then incubated with NP-40-soluble, RHΔhxgprt parasite extracts in GST pull-down experiments. A sample of the input material and the co-purified, eluted material was then analyzed by immunoblotting using antisera for the different members of the MJ complex. Antisera to the abundant surface antigen, TgSAG1, or the abundant rhoptry protein, TgROP1, were also used to assess the specificity of the co-precipitations. We observed that both GST-D3 and GST-D4 but not GST alone efficiently co-precipitated TgAMA1 as determined by immunoblotting with the TgAMA1-specific monoclonal antibody B3.90 (Figure 3A, lanes 2–4). There was no detectable co-precipitation of the two negative controls (TgSAG1 and TgROP1), indicating that both the D3 and D4 domains of TgRON2 are independently capable of a specific interaction with TgAMA1.
Immunoblotting for TgRON4, TgRON5 and TgRON8 demonstrated that these proteins were not detected in the GST-D3 co-purified material (Figure 3A, lane 3) but a trace amount of at least TgRON4 and TgRON5 was consistently observed in the GST-D4 co-purified material (Figure 3A, lane 4). These results suggest that under these conditions GST-D4 may be able to interact with one portion of TgAMA1 without completely disrupting the native TgRON2:TgAMA1 interaction (via D3) such that a small amount of the entire complex is pulled down. As discussed further below, the D4 interaction may be too weak to hold the native complex together so that no native TgRON2 remains bound to the TgAMA1 in the GST-D3 co-precipitation.
To confirm that the associations between TgAMA1 and GST-D3 or GST-D4 were direct and not occurring via binding to a TgRON2:TgAMA1 complex, we repeated the GST pull-down experiments using the TgRON2-HA parasites followed by immunoblotting with an HA-specific antibody. The results show that, similar to TgRON4, TgRON2-HA was not detected in the material co-purifying with GST-D3 but trace amounts of it could be detected in the GST-D4 material (Figure 3B, lanes 3 and 4). Collectively, these results demonstrate that GST-D3 and GST-D4 are independently sufficient to affinity-purify TgAMA1 and that these associations are generally not as part of a complex with the other identified members of the MJ although GST-D4 may be able to associate with the complex without disrupting it completely.
To better characterize the associations between GST-D3/D4 and TgAMA1 we repeated the co-purification studies using more stringent conditions (i.e., in the presence of the ionic detergent sodium deoxycholate) that were shown to still yield intact MJ complexes in previous co-immunoprecipitation studies [8], [10]. Immunoblotting analysis of the co-purified material with the TgAMA1-specific antibody B3.90 demonstrate that while the association between GST-D3 and TgAMA1 was maintained, the association between GST-D4 and TgAMA1 is not seen under these conditions (Figure 3C, lanes 3 and 4). As sodium deoxycholate is considered to be more denaturing than NP-40 alone, these results suggest that the interaction between TgRON2 domain 3 and TgAMA1 is considerably stronger than the interaction involving domain 4.
Current models of the assembly of the MJ complex predict that TgRON2 is associated with the host cell where it acts as a receptor for the ectodomain of TgAMA1 [12], [13]. To test this hypothesis and determine whether D3 and/or D4 are the domains of TgRON2 that are functioning in this association we used GST-D3 and GST-D4 in GST pull-down experiments with parasite culture supernatants containing the shed, N-terminal ectodomain of TgAMA1 [15], [26]. To discriminate between the shed and the intact, full-length forms of TgAMA1, immunoblotting was performed using monoclonal antibodies specific for either the C-terminal intracellular domain of TgAMA1 (CL22; [15]) or the N-terminal extracellular domain (B3.90; [26]). The results showed that both GST-D3 and GST-D4 but not GST alone efficiently co-precipitate the more rapidly migrating (shed) form of TgAMA1 (Figure 4A). The identity of this as the shed form was confirmed by its failure to react to CL22 (Figure 4B). A trace amount of contaminating, intact, full-length TgAMA1 in the supernatant material was also co-precipitated, as expected, since this also includes the entire ectodomain (Figure 4 A and B, lanes 4 and 5). The specificity of these co-purification studies was confirmed by a complete lack of enrichment for the abundant surface antigen TgSAG1 (Figure 4). These results demonstrate that both GST-D3 and GST-D4 independently and specifically interact with the ectodomain of TgAMA1.
To determine whether GST-D3 and GST-D4 can interact with the ectodomain of TgAMA1 on intact RH parasites, we used IFA and anti-GST antibodies. Freshly prepared, extracellular parasites were pre-incubated with molar equivalents of GST or GST-D3 (see below for details on GST-D4) then permitted to settle on HFF monolayers under conditions that were not permissive for invasion, followed by a brief shift to invasion-permissive conditions. Infected monolayers were fixed and then stained in the absence of any added permeabilizing agents using a GST-specific antibody and the TgAMA1 ectodomain-specific monoclonal antibody B3.90. The results showed that parasites pre-incubated with GST-D3, but not GST alone, exhibit surface staining with the GST-specific antibody (Figure 5 A and B). As a control to confirm that the parasite plasma membrane was not permeabilized in the fixation process we also stained parasites with the monoclonal antibody CL22, which is specific for the intracellular C-terminus of TgAMA1 (Figure 5C). No staining with CL22 was observed, confirming the intactness of the membranes. These results indicate an association between GST-D3 and the exposed, ectodomain of TgAMA1, which under these conditions is distributed over the entire surface of the parasite (Figure 5 A and B, second panel; [15], [26].
To verify that binding of GST-D3 to the parasite surface is via TgAMA1, we used the conditional TgAMA1 knockout strain (RHΔama1/AMA1-myc) [16] grown in the presence of anhydrotetracycline (Atc), which turns off TgAMA1 expression. Freshly prepared, extracellular RHΔama1/AMA1-myc parasites grown under such conditions were incubated in the presence of GST-D3 and stained, as described above. Staining for TgAMA1 using the monoclonal B3.90 confirmed that these parasites were not expressing detectable levels of TgAMA1 (Figure 5D, second panel) and that, in contrast to parasites that express wild-type levels of TgAMA1, there was no detectable binding of GST-D3 (Figure 5D, third panel). These results demonstrate that the surface staining observed with GST-D3 is dependent on TgAMA1.
Similar analyses were conducted with GST-D4. By IFA, we observed that multiple, independent preparations of GST-D4 stain the surface of RH parasites, as predicted, but these preparations were toxic, causing what appears to be membrane blebs and, perhaps as a result, alteration of TgAMA1 staining in more than 50% of the parasites treated (Figure 5 E and F). In contrast, there was no detectable binding of GST-D4 on most (>80%) of RHΔama1/AMA1-myc parasites that were grown in the presence of Atc (Figure 5G). However, we did observe a toxic effect, as well as GST staining, in less than 20% of the RHΔama1/AMA1-myc parasites (data not shown). While there was no detectable TgAMA1 on these parasites as determined by B3.90 staining (data not shown), it is possible that residual surface-localized TgAMA1, while below the limit of detection, may have been sufficient for binding by GST-D4. Collectively, these results indicate that binding of GST-D4 to the surface of Toxoplasma occurs in an AMA1-specific manner but this binding elicits a morphological change in the parasite plasma membrane that precluded further functional analyses.
Given that GST-D3 associates with the ectodomain of TgAMA1 on the surface of intact parasites, we predicted that pre-incubation of extracellular parasites with GST-D3 would result in an invasion-inhibitory phenotype. To test this hypothesis, equivalent numbers of freshly prepared, extracellular RH parasites were pre-incubated in medium supplemented with molar equivalents of GST-D3, GST alone, or a buffer control and then allowed to invade host cells using a temperature-shift assay to synchronize the process. Following ∼15 minutes at an invasion-permissive temperature, infected monolayers were fixed and analyzed by IFA. Extracellular vs. intracellular parasites were identified by sequential staining for TgSAG1, before and after detergent-permeabilization of the host cells. While treatment of RH parasites with GST alone did not significantly alter the number of intracellular parasites compared to the buffer-treated control parasites, treatment with GST-D3 resulted in a dose-dependent decrease of up to ∼55% (Figure 6A and Supplemental Figure S1A). This level of inhibition is similar to that previously reported with anti-TgAMA1 antibodies [15] and is consistent with disruption of TgAMA1 function through binding of GST-D3.
Previous studies demonstrated that the levels of TgAMA1 in wild-type parasites are apparently present in excess: RHΔama1/AMA1-myc parasites that, even in the absence of Atc, only express ∼10% of the wild-type levels of TgAMA1 are fully competent for invasion [16]. Given this, we postulated that the excess TgAMA1 on the surface of RH parasites might act to absorb GST-D3 thus decreasing the effect of this protein on the biologically-relevant minority of TgAMA1 actually involved in invasion. To test this hypothesis and to confirm that the invasion-inhibitory phenotype we observe with GST-D3 treatment of RH parasites is dependent upon TgAMA1, we repeated the invasion assays using the RHΔama1/AMA1-myc parasites. As seen with the RH parasites, pre-incubation of RHΔama1/AMA1-myc parasites with GST alone did not affect the number of intracellular parasites, as compared to the buffer control (Figure 6B). Strikingly, however, we observed that pre-incubation of the RHΔama1/AMA1-myc parasites with GST-D3 resulted in a dose-dependent decrease of up to ∼87% in invasion efficiency (Figure 6B and Supplemental Figure S1B). To confirm the specificity of this effect, RHΔama1/AMA1-myc parasites were also incubated with a GST protein fused to a scrambled version of the D3 peptide sequence (designated GST-D3scramble), which does not affinity-purify TgAMA1 from parasite extracts or culture supernatants (data not shown). The percentage of intracellular parasites following treatment with GST-D3scramble was similar to that seen for parasites incubated with the buffer or GST alone controls (Figure 6B) demonstrating that the invasion-inhibitory phenotype observed with GST-D3 is indeed highly specific. A specific interaction with TgAMA1 is argued by the fact that the effect of GST-D3 treatment was much more pronounced for the RHΔama1/AMA1-myc parasites than it was for wild-type.
To determine if treatment of parasites with GST-D3 affects host cell attachment we tested the ability of the RHΔama1/AMA1-myc parasites treated with GST alone or GST-D3 to attach to formaldehyde-fixed HFF monolayers. T. gondii will attach to fixed monolayers but cannot invade thus permitting us to test attachment independently of invasion [27]. Using this assay we did not observe any statistically significant difference between the number of attached parasites following treatment with GST alone, GST-D3, or the buffer control (Figure 6C). To determine if treatment of parasites with GST-D3 affects gliding motility, we examined the number and type of “gliding” trails deposited on glass coverslips by RHΔama1/AMA1-myc parasites treated with GST-D3 or GST alone. Again, we saw no detectable difference in the number or form of such trails following either treatment (data not shown). Collectively, these results demonstrate that GST-D3 impedes invasion of host cells, rather than attachment or gliding motility.
During or soon after invasion, Toxoplasma tachyzoites inject soluble rhoptry bulb proteins (“ROPs”) into the host cell as a means to co-opt host cell functions [28], [29], [30]. To determine if treatment with GST-D3 affects the injection of ROPs, we analyzed “SeCreEt” parasites that have been engineered to inject a protein consisting of the soluble ROP, toxofilin, fused to Cre recombinase [31]. Upon invasion of a Cre-reporter host cell by a single SeCreEt parasite, Cre-mediated recombination results in host cell expression of GFP. It should be noted that under normal conditions, Cre-mediated recombination can be observed in occasional uninfected cells; these uninfected, GFP-positive cells are thought to result from abortive invasion events or cells that divide following invasion by the parasites [31]. To determine if GST-D3 treatment affects injection of the toxofilin-Cre fusion, equivalent numbers of freshly prepared, extracellular SeCreEt parasites were pre-incubated in medium supplemented with molar equivalents of GST-D3, GST alone, or a buffer control and then allowed to invade Cre-reporter host cells using a temperature-shift assay to synchronize the process. Following ∼30 minutes at an invasion-permissive temperature, infected monolayers were washed three times and then incubated in media supplemented with GST-D3, GST alone, or a buffer control for an additional ∼24 hours. Infected monolayers were then analyzed for the number of GFP-positive host cells. Using this assay, we saw no significant difference in the total number of GFP-positive host cells following treatment with GST-D3, GST, or a buffer control (Figure 7A). While the total number of GFP-positive cells did not differ, the percentage of the GFP-positive cells that were infected with SeCreEt parasites was significantly reduced following treatment with GST-D3, as expected (Figure 7B). The extent of this reduction in invasion efficiency was less than that observed in the experiments reported in figure 6, likely as a result of the much longer time needed for the Cre-reporter assay (∼24 hours vs. ∼15 minutes); the additional time may allow some number of the GST-D3-inhibited parasites to eventually enter. Regardless, these results clearly demonstrate that while GST-D3 binding to AMA1 can block invasion of host cells, it has little if any effect on the injection of the rhoptry bulb contents.
Due to its toxic effect on the parasites, similar analyses could not be attempted with GST-D4. It is unclear if this effect on the parasites is biologically relevant but, regardless, it precluded attempts to determine if preincubation with GST-D4 has an effect on T. gondii invasion.
The results presented here demonstrate that TgRON2 localizes to the MJ of invading parasites and that at least two regions within the last ∼200 amino acids of TgRON2 independently and specifically associate with the ectodomain of TgAMA1. A short sub-region of just 54 amino acids (D3) was found to block invasion, also in a TgAMA1-dependent manner. The invasion-inhibitory effect of this interaction is unlikely to be due simply to the non-specific presence of GST-D3 on the surface as coating Toxoplasma with antibodies to the abundant surface antigen TgSAG1 does not impair host cell attachment and invasion [27], [32]. Instead, our results strongly argue for a critical role for the association of TgAMA1 and TgRON2 in productive invasion of host cells.
The D3 and D4 regions of TgRON2 identified as interacting with TgAMA1 are separated by 19 amino acids that were previously predicted to span a membrane [8], [10], [13], [24]. As both regions interact with the ectodomain of TgAMA1, however, our data clearly indicate that the entire C-terminus of TgRON2, encompassing D3, HH3, and D4, is on the extra-cytosolic region of the MJ during invasion; i.e. HH3 does not span a membrane. We note that both HH2 and HH3 exhibit a high degree of identity between TgRON2 and its orthologues in other Apicomplexa, thus indicating a key role for these regions beyond simple hydrophobicity. Given the conserved nature of the hydrophobic cleft in AMA1 from Toxoplasma and Plasmodium and the importance of this region in PfAMA1 binding to the MJ complex [11], [14], [33], our results suggest the possibility that HH3, along with D3 and D4, specifically associates with the TgAMA1 face that includes the hydrophobic cleft via hydrophobic∶hydrophobic interactions. These results also provide a likely molecular basis for recent reports using P. falciparum where the invasion-inhibitory R1 peptide and anti-PfAMA1 antibody 4G2 were shown to inhibit invasion [11], [14]; R1 has substantial similarity to a portion of PfRON2 overlapping HH3 and the N-terminal end of domain 4 (VFAEFLP-LFSKF-GSRM-HILK for the R1 peptide vs. LFASIGPYLFAPMAGLAVWNILK for PfRON2 with similar residues underlined and identical residues underlined and in bold). Analysis of the association between the C-terminus of RON2 and the ectodomain of AMA1 by co-crystallization will be needed to resolve the exact nature of their association.
While this manuscript was in preparation, we learned of similar results by others [34]. Their data corroborate ours and also show that the N-terminus of TgRON2 is inside the host cell. Combined with the results presented here showing that the D3 and D4 domains are both outside the host cell, it would appear that TgRON2 spans the host plasma membrane an odd number of times. Given that HH3 is apparently not a TM domain, it is most likely that only one of HH1 and HH2 can be a true trans-membrane domain although we cannot exclude the possibility that this topology is accomplished through the existence of a cryptic, third such span. The fact that TgRON2 and TgRON4, which is reported to be inside the host cell plasma membrane [13], also appear to form a tight interaction with one-another [8] supports a model wherein TgRON2 spans the host membrane with D3/D4 binding to TgAMA1 on the parasite surface and with some other portion of TgRON2, N-terminal of D3, binding TgRON4 inside the host cell.
While it seems simplest to explain the invasion-inhibitory effect of GST-D3 as being due to a steric block in the binding of TgRON2 to TgAMA1, we cannot exclude the possibility that the effect is also a result of interference with a key signaling event. GST-D3-binding did not detectably affect the ability of the parasites to inject rhoptry bulb proteins. This suggests that rhoptry neck proteins (e.g., TgRON2) and at least some amount of rhoptry bulb proteins are injected at the earliest stages of invasion. This is not unexpected based on published data showing injection of ROP9 by a parasite that is less than a third of the way into the host cell [28]. The fact that the RONs are presumably injected before MJ formation (since they must gather into a complex inside the host cells as part of MJ assembly) and the fact that there is no known separation of the bulb and neck compartments also makes the failure of GST-D3 to block ROP injection not unexpected. Hence, it appears that rhoptry bulb and neck release are not separable phenomena, although by virtue of their physical proximity to the apical end of the rhoptries, the RONs are likely injected before the ROPs.
Overall, and regardless of whether the GST-D3-mediated block in MJ formation is steric or through aberrant signaling, our data clearly indicate that the TgAMA1/TgRON2 interaction is key for invasion. It is reasonable to expect, therefore, that synthetic compounds that block this interaction will be potent anti-parasitic agents. Similarly, the D3–D4 region of RON2 should receive serious consideration as a possibly synergistic addition to current AMA1-based vaccines.
Human foreskin fibroblasts (HFFs) were cultured in complete Dulbecco's modified Eagle's medium (DMEM; Invitrogen, Carlsbad, CA) supplemented with 10% heat-inactivated fetal calf serum (Hyclone, Logan, UT), 2 mM L-glutamine, 100 U ml−1 penicillin and 100 µg ml−1 streptomycin.
Toxoplasma gondii (RH strain) lacking a functional hypoxanthine-xanthine-guanine phosphoribosyltransferase (HXGPRT) gene (designated RHΔhxgprt) [35] were cultured by serial passage on confluent monolayers of HFFs in complete DMEM at 37°C with 5% CO2. Selection and maintenance of transfected parasites expressing the HXGPRT cassette were cultured in complete DMEM supplemented with 50 µg/ml of mycophenolic acid and 50 µg/ml of xanthine (MPA/XAN).
All primers used in these studies are listed in Table S1.
To generate a Toxoplasma strain where the endogenous TgRON2 locus was replaced with a copy of TgRON2 fused to a C-terminal hemagglutinin (HA) epitope tag, a plasmid with ∼2.5–3 kb of 5′ and 3′ homologous TgRON2-targeting sequences, flanking the HA tag coding sequence and the HXGPRT gene, was constructed. The 3′ targeting sequence, including the TgRON2 3′UTR, was PCR amplified from Toxoplasma genomic DNA (RH strain) using primers A and B, and then introduced into the pTKO vector (kindly provided by G. Zeiner, Stanford University, Stanford, CA) using the NheI and ApaI sites. The 5′ targeting sequence encompassing the last ∼3 kb of TgRON2, but excluding the native stop codon, was PCR amplified from Toxoplasma genomic DNA using primers C and D. The resultant PCR product, which also contained the coding sequence for the HA tag and a downstream stop codon in frame with the last codon of TgRON2, was introduced into the derivative of pTKO that contained the 3′ targeting sequence using the KpnI and EcoRV sites, to generate the complete pT-TgRON2-HA tagging vector.
The pT-TgRON2-HA vector was introduced into RHΔhxgprt parasites by electroporation as described previously [36], selected for by passage in complete DMEM supplemented with MPA/XAN, and then individual Toxoplasma clones were isolated by serial dilution. Insertion of the HA tag and the HXGPRT gene into the TgRON2 locus was confirmed by PCR using two primers sets; each set included one primer that hybridized to the Toxoplasma genome outside of the targeting regions cloned into the pT-TgRON2-HA and one primer that hybridized to a unique sequence present in the targeting vector that was retained in the Toxoplasma genome after the double-recombination event. Proper integration at the 5′ or 3′ ends was confirmed using primers E and F or G and H, respectively.
To generate TgRON2 fusion proteins with a N-terminal GST tag, subregions of the TgRON2 cDNA were PCR-amplified from a Toxoplasma RH strain cDNA library and introduced into pGEX-6P1 (Agilent) using the BamHI and EcoRI sites. To generate GST-D3, the coding sequence for TgRON2 amino acids 1293–1346 were PCR amplified using primers I and J. To generate GST-D4, the coding sequence for TgRON2 amino acids 1366–1479 was PCR amplified using primers K and L. An unbiased scrambled derivative of D3 was generated computationally with the RandSeq tool (ExPASy proteomics Server, Swiss Institute of Bioinformatics) and the resulting amino acid sequence, designated D3scramble, was reverse translated to a coding sequence for the purpose of primer design. The primers used to generate GST-D3scramble (primers M and N) were designed with 16 base-pairs of overlapping sequence so that they served both as primers and template in a PCR reaction. The amplified product was inserted into pGEX-6P1 as described above.
Production and purification of the GST proteins from E. coli strain Rosetta (Novagen) were done essentially as described previously [37]. Concentrated, purified proteins were stored at −80°C in buffer containing 10 mM Tris-HCl pH 8.0, 150 mM NaCl, and 10% glycerol.
To identify Toxoplasma proteins that interact with GST-D3/D4, approximately 4×108 extracellular RHΔhxgprt or TgRON2-HA parasites were washed three times in 1× phosphate-buffered saline (PBS) and then lysed on ice in 1 ml of lysis buffer (10 mM Tris-HCl pH 8.0, 150 mM NaCl, 1 mM EDTA, 0.1% NP-40) supplemented with Complete EDTA-free protease inhibitors (Roche). The cleared, NP-40-solublized lysate was divided equally into three tubes and each fraction was supplemented with 4B Glutathione-sepharose beads (GE Healthsciences) that were pre-bound with 0.5 µM of GST, GST-D3, or GST-D4. The lysate suspensions were rotated at room temperature for approximately two hours and the bound beads were then washed three times in lysis buffer, followed by elution of the GST fusion proteins and any co-purified parasite proteins by boiling for ∼5 minutes in 2× SDS sample buffer (125 mM Tris-HCl pH 7.0, 4% SDS, 20% glycerol, 0.005% bromophenol blue) supplemented with 10% β-mercaptoethanol. Lysate samples and eluted material were separated on 4–12% gradient Bis-Tris gels (Invitrogen) and analyzed by western blot using the appropriate antibodies.
To better understand the nature of the interaction between TgAMA1 and GST-D3 or GST-D4, the GST pull-down experiments were conducted as described above in buffers supplemented with sodium deoxycholate at a final concentration of 0.25% (w/v).
To determine if the GST fusion proteins bind to the shed ectodomain of TgAMA1, freshly extracellular RHΔhxgprt parasites were pelleted and washed three times in Hanks Buffered Saline Solution (HBSS) supplemented with 20 mM HEPES pH 7.4. Washed parasites were passed through a 5 µm filter, pelleted, and then resuspended at a density of ∼3.5×108 parasites per ml in serum-free DMEM supplemented with 20 mM HEPES pH 7.4 and 2 µM ionomycin (Sigma). Following incubation at 37°C for approximately 30 minutes, the parasites were pelleted. The resulting pellet was washed twice in PBS and then resuspended in one ml of SDS sample buffer (without reducing agents) and saved as the “pellet” fraction while the resulting supernatant was spun again at 100,000×g for 45 minutes at 4°C. A sample of the cleared “supernatant” fraction was saved and the remainder was supplemented with NaCl (150 mM final) and NP-40 (0.1% (v/v) final), then equally divided into three tubes followed by incubation with 0.5 µM GST, GST-D3, or GST-D4 proteins pre-bound to glutathione sepharose. The GST pull-down experiments were conducted as described above. Elution of the GST fusion proteins and any co-purified parasite proteins was by boiling for ∼5 minutes in SDS sample buffer without any reducing agents.
To detect TgRON2-HA by western blot, intracellular RHΔhxgprt or TgRON2-HA-expressing parasites were released from infected HFFs by scraping and passage through a 27 gauge needle. Released parasites were washed twice in 1× PBS and then counted. Lysates were generated for each strain by pelleting 106 parasites and resuspending each pellet in 15 microliters of 2× SDS sample buffer supplemented with 10% β-mercaptoethanol. Following boiling for 5 minutes, 10 microliters of each lysate were separated by SDS-PAGE and TgRON2-HA was detected using the anti-HA rat monoclonal antibody 3F10 conjugated to horseradish peroxidase (HRP) (Roche).
In western blot analyses of the GST pull-down material, TgRON4/8 were detected using rabbit or mouse polyclonal antisera, respectively [8], [12]. TgRON5 was detected using rabbit polyclonal antisera specific for the N-terminus of the protein (kindly provided by the Bradley laboratory, University of California-Los Angeles, Los Angeles, CA). TgAMA1 was detected with mouse monoclonals B3.90 [26] or CL22 [15]. TgSAG1 was detected using rabbit polyclonal sera (a gift from M. Grigg, NIH). TgROP1 was detected using the mouse monoclonal antibody Tg49 [38]. All goat anti-mouse and anti-rabbit secondary antibodies were HRP-conjugated (Biorad).
All secondary AlexaFluor-conjugated antibodies were obtained from Molecular Probes.
To visualize the TgRON2-HA fusion protein in the rhoptry necks, HFF monolayers infected with TgRON2-HA-expressing parasites were fixed in 1× PBS containing 2.5% formaldehyde (EM Biosciences) ∼16 hours post-infection. These fixed HFF monolayers were then permeabilized using 100% methanol and stained with the HA-specific rat monoclonal 3F10 (Roche) and rabbit anti-RON4 polyclonal sera followed by AlexaFluor488-goat-anti-rat antibody and AlexaFluor594-goat-anti-rabbit antibody, respectively. Coverslips were mounted onto glass slides using Vectashield (Vector Laboratories) and then examined using 100× oil-immersion lens on an Olympus BX60 upright fluorescent microscope. All digital images were obtained using Image-Pro Plus and the same exposure parameters were used for all comparison sets.
To visualize TgRON2-HA at the MJ of partially invaded parasites, extracellular parasites were prepared for synchronous invasion of HFF monolayers using high potassium buffers essentially as described [25]. Approximately 45 seconds following the addition of invasion-permissive media, infected monolayers were fixed in 1× PBS containing 2.5% formaldehyde, permeabilized with 1× PBS containing 0.2% triton X-100 and then the HA-tagged proteins and TgRON4 were detected as described above.
To visualize binding of GST proteins to the surface of intact parasites, ∼108 freshly extracellular (by natural egress) RH or RHΔama1/AMA1-myc parasites (grown for 48 hours in the presence of anhydrotetracycline (Atc) as described [16]) were washed twice in DMEM with 2% FBS and then passed through a 5 µm filter. Parasites were incubated in DMEM containing 2% FBS that was supplemented with either 5 µM GST, 5 µM GST-D3, or 5 µM GST-D4 by rotation at 37°C for ∼20 minutes. Parasites were then chilled in an ice water bath to prepare them for temperature-based synchronized invasion. This chilled Toxoplasma suspension was then added to pre-chilled HFF monolayers grown on glass coverslips. Parasites were permitted to settle onto the HFFs in an ice water bath for ∼12 minutes prior to invasion. To initiate invasion, the plate was then transferred to a 37°C water bath for ∼1 minute. Infected monolayers were washed twice in 1× PBS and then fixed in 1× PBS containing 2.5% formaldehyde. Unless otherwise stated, all staining steps were conducted in 1× PBS supplemented with 3% bovine serum albumin (BSA; Sigma) in the absence of added permeabilization agents. TgAMA1 was visualized using the mouse monoclonal antibodies B3.90 or CL22 and AlexaFluor594-goat-anti-mouse. GST proteins were visualized using rabbit anti-GST antisera (ICL) and AlexaFluor488-goat-anti-rabbit.
The preparation of Toxoplasma for the invasion assays was conducted essentially as described above for the IFA analysis of GST-treated parasites with minor modifications. Briefly, washed RH or RHΔama1/AMA1-myc parasites were incubated in DMEM with 2% FBS supplemented with 2.5–5 µM GST, 2.5–5 µM GST-D3, 5 µM GST-D3scramble (RHΔama1/AMA1-myc parasites only) or 1.5% volume of a buffer control (10 mM Tris-HCl pH 8.0, 150 mM NaCl, 10% glycerol). Using temperature-based synchronization as described above, invasion was permitted for approximately 15 minutes at 37°C, and then infected HFF monolayers were washed gently once in 1× PBS and fixed in 1× PBS containing 2.5% formaldehyde. To stain only the extracellular parasites, fixed monolayers were stained with rabbit anti-SAG1 antiserum followed by AlexaFluor594-goat-anti-rabbit. To stain all parasites, the infected monolayers were then permeabilized with 1× PBS containing 0.2% triton X-100 followed by staining with the anti-SAG1 mouse monoclonal antibody DG52 [39] and AlexaFluor488-goat-anti-mouse. The numbers of green (intracellular and extracellular) and red (extracellular) Toxoplasma were counted in 15 (RH) or 20 (RHΔama1/AMA1-myc) randomly selected fields on each of three separately mounted coverslips for each condition and visualization was performed using a 20× lens on a Nikon Eclipse TE300 microscope. All digital images were obtained using Image-Pro Plus and parasites were quantified using ImageJ. Results are representative of data from at least three independent experiments.
Extracellular RHΔama1/AMA1-myc parasites were prepared and treated with the GST proteins as described above. Following treatment with the GST proteins, the parasite suspensions were added directly to HFF monolayers on coverslips that had been formaldehyde-fixed prior to the addition of Toxoplasma, essentially as described [27] and permitted to settle for 15 minutes at 37°C. Monolayers of formaldehyde-fixed HFFs with attached parasites were washed once (for consistency with the invasion assays) and then fixed in 100% methanol. Parasites were stained with rabbit anti-SAG1 antisera followed by AlexaFuor594-goat-anti-rabbit. The number of attached parasites was imaged and quantified as described above. Results are representative of data from at least three independent experiments.
Toxoplasma SeCreEt parasites were used to infect Cre-reporter host cells essentially as described [31] with minor modifications. Extracellular SeCreEt parasites were prepared and treated with the GST proteins as described above. Approximately 30 minutes following invasion, infected monolayers were washed three times in complete media and then incubated at 37°C for 24 hours in complete media supplemented with a buffer control, 5 µM GST, or 5 µM GST-D3. Infected monolayers were washed once in 1× PBS and then fixed in formaldehyde as described above. The number of GFP-positive host cells were counted in 15 randomly selected fields on each of four separately mounted coverslips for each condition and visualization was performed using a 20× lens on a Nikon Eclipse TE300 microscope. The number of infected or uninfected GFP-positive host cells was counted in randomly selected fields on each of four separately mounted coverslips for each condition and visualization was performed using a 100× oil-immersion lens on an Olympus BX60 upright fluorescent microscope. For each condition, 100 GFP-positive cells were examined. Results are representative of data from two independent experiments.
The Graphpad Prism program was used to determine if the differences observed in the data for the invasion, attachment, and rhoptry secretion assays were statistically significant. Data sets were analyzed using the unpaired Student's t-test (two-tailed), under the assumption of equal variance. A p-value of less than 0.05 was considered to be statistically significant in the tests.
|
10.1371/journal.pcbi.1000134 | Predicting Human Nucleosome Occupancy from Primary Sequence | Nucleosomes are the fundamental repeating unit of chromatin and comprise the structural building blocks of the living eukaryotic genome. Micrococcal nuclease (MNase) has long been used to delineate nucleosomal organization. Microarray-based nucleosome mapping experiments in yeast chromatin have revealed regularly-spaced translational phasing of nucleosomes. These data have been used to train computational models of sequence-directed nuclesosome positioning, which have identified ubiquitous strong intrinsic nucleosome positioning signals. Here, we successfully apply this approach to nucleosome positioning experiments from human chromatin. The predictions made by the human-trained and yeast-trained models are strongly correlated, suggesting a shared mechanism for sequence-based determination of nucleosome occupancy. In addition, we observed striking complementarity between classifiers trained on experimental data from weakly versus heavily digested MNase samples. In the former case, the resulting model accurately identifies nucleosome-forming sequences; in the latter, the classifier excels at identifying nucleosome-free regions. Using this model we are able to identify several characteristics of nucleosome-forming and nucleosome-disfavoring sequences. First, by combining results from each classifier applied de novo across the human ENCODE regions, the classifier reveals distinct sequence composition and periodicity features of nucleosome-forming and nucleosome-disfavoring sequences. Short runs of dinucleotide repeat appear as a hallmark of nucleosome-disfavoring sequences, while nucleosome-forming sequences contain short periodic runs of GC base pairs. Second, we show that nucleosome phasing is most frequently predicted flanking nucleosome-free regions. The results suggest that the major mechanism of nucleosome positioning in vivo is boundary-event-driven and affirm the classical statistical positioning theory of nucleosome organization.
| Inside the nucleus, DNA is wrapped into a complex molecular structure called chromatin, whose fundamental unit is ∼150 bp of DNA organized around the eight-histone protein complex known as the nucleosome. Understanding the local organization of nucleosomes is critical for understanding how chromatin impacts gene regulation. Here, we describe a computational model that predicts nucleosome placement from DNA sequence. We train the model using data derived from human cell lines, and we apply the model systematically to 1% of the human genome. We show that previously described models trained from yeast data correlate strongly with the human-trained model, suggesting a common mechanism for sequence-based determination of nucleosome occupancy. In addition, we observe a striking complementarity between models trained using data from weakly and strongly digested samples: one type of model recognizes nucleosome-free regions, whereas the other identifies well-positioned nucleosomes. Finally, our analysis of predicted nucleosome positions in the human genome allows us to identify common features of nucleosome-forming and inhibitory sequences. Overall, our results are consistent with the classical statistical positioning theory of nucleosome organization.
| Nucleosomes are the fundamental repeating unit of chromatin, and the positioning of nucleosomes along the genome has been a topic of long-standing interest. The prevailing “statistical positioning” theory of nucleosome organization was first proposed by Kornberg more than 25 years ago [1]. This theory, for which considerable experimental evidence exists [2], posits that nucleosomes are stochastically positioned along the genome and are distributed between boundary events that comprise nucleosome-free regions, such as those known to be found at the promoters of active or poised genes.
According to statistical positioning theory, the repetitive nucleosomal structure is dynamically punctuated by short regions where regulatory factors bind in place of canonical nucleosomes. Whether a particular genomic position is occupied by a nucleosome may therefore vary from cell to cell within a population of cells and between different cell types. However, it is expected that the vast majority of the genome at any given time is covered by nucleosomes. The observation that specific DNA sequences favored the formation of nucleosomes [3]–[10] raises the possibility that sequence plays a significant role in organizing nucleosomal arrays in vivo.
The determination of nucleosome placement along the genome presumably depends upon a variety of factors, including properties of the sequence itself, physical constraints, and epigenetic factors such as ATP-dependent chromatin remodeling or alterations in the biochemical composition of the histone octamer. MNase cleaves chromatinized DNA preferentially in inter-nucleosomal linker regions; with robust digestion chromatin can be reduced to mononucleosomes and their associated ∼147 bp DNA fragments, which can in turn be mapped to the genome to reveal nucleosome positions using either tiling DNA microarrays or sequencing assays.
Recently, Segal et al. [11] proposed a computational model for sequence-based prediction of nucleosome positioning in S. cerevisiae. The model is generative and, motivated by previous work [12], uses dinucleotide frequencies collected from a training set of aligned nucleosome-bound sequences. Segal et al. used dynamic programming to identify the highest-scoring series of nonoverlapping nucleosome positions along a given chromosome, and demonstrated that, for a test set of nucleosomal sequences not used in the training procedure, the model places 54% of the nucleosomes within 35 bp of their true locations. This is 15% higher than would be expected by random placement of the same number of nucleosomes along the genome.
Subsequently, Peckham et al. [13] proposed a complementary computational model that is discriminative, rather than generative, and that focuses only on sequences that show the strongest signals of nucleosome occupancy or vacancy in a microarray-based assay of S. cerevisiae [14]. Briefly, the assay consists of cleaving nucleosomal and bare genomic DNA with MNase, and then measuring the abundance of uncleaved products using a tiling microarray with 50-mer probes. The model of Peckham et al. is trained to discriminate between microarray probe sequences that showed the highest and lowest fluorescence log-ratios. The resulting model correctly predicts 50% of well-positioned nucleosomes within 40 bp of their correct positions. Like the model of Segal et al., this represents an improvement of 15% relative to random placement of nucleosomes along the genome. Recently, two variants of the MNase microarray assay were applied to human DNA [15],[16]. Here, we demonstrate that the discriminative model described by Peckham et al. can be successfully trained from these two human datasets. We chose to use the discriminative approach because the current understanding of chromatin biology suggests that there are, indeed, genomic sequences which ensure nucleosomal occupancy [11],[14], as well as sequences which nucleosomes seem to avoid. The latter include nucleosome-free regions enriched in gene regulatory regions, CpG islands, and transcription termination sites [14], [17]–[19], nucleosome-free regions flanked by H2A.Z [20],[21], nucleosome-free regions ∼200 bp upstream of the start codon of Pol II transcribed genes in yeast [14],[21], polycomb response elements [22],[23], and P53 binding sites [24]. Because several lines of inquiry indicate that primary genomic sequences can positively or negatively influence whether a particular locus is nucleosomally occupied, we felt justified in using the SVM to discriminate between probes in our datasets that were more or less MNase sensitive and thus more or less likely to be occupied by nucleosomes.
We find that, for both of the human MNase array datasets that we investigated, the SVM model achieves a comparable level of cross-validated accuracy as the yeast-trained model. Furthermore, we observe that the qualitative behavior of the learned model depends strongly the amount of MNase digestion applied to the given sample. At lower levels of digestion, the model excels at recognizing regions that are protected from MNase cleavage; conversely, with stronger digestion, the model accurately identifies regions of accessibility to MNase cleavage. Essentially, the former model predicts the presence of a strongly positioned nucleosome, whereas the latter predicts the absence of a nucleosome. Next, we apply both types of trained model to the ENCODE regions of the human genome, and we observe close agreement with the yeast-trained models of Peckham et al. and Segal et al., suggesting a shared mechanism for sequence-based determination of nucleosome occupancy. Additionally, our model indicates that nucleosomal occupancy is primarily determined by short genomic sequences, with C- and G-containing dinucleotide pairs strongly represented in the nucleosome-forming sequences. We also show that the models predict a dip upstream and a peak downstream of annotated transcription start sites in the ENCODE regions, suggestive of nucleosome placement at the TSS with nucleosome absence immediately upstream. The model was also used to understand the characteristics of sequences around nucleosome-forming and nucleosome inhibitory probes. We find that nucleosome-forming probes do not necessarily prescribe a translationally positioned nucleosome, while sequences surrounding nucleosome inhibitory sequences have SVM discriminant scores that indicate translationally positioned nucleosomes. Finally, the model can be used to make predictions of nucleosomal occupancy on DNA fragments used in in vitro experiments. We can use these predictions to refine our understanding of the primary sequence contribution to ATP-dependent remodeling experiments.
Overall, our results suggest that, to the extent that nucleosome positioning signals exist, they exist in tandem with nucleosome-free regions. This is fundamentally consistent with the statistical positioning theory, in which nucleosome positioning and nucleosome-free region boundary events are mechanistically linked.
We began our analyses by training a model using the microarray data from Dennis et al. [16]. This dataset contains three microarrays, each containing ∼120,000 probes. The probes cover 25 kb regions upstream of 42 genes, using 50-mer probes tiled every 20 bases. For all of our analyses, we omitted probes that overlap repetitive elements, as identified by RepeatMasker [25]. Each 50-mer is included on the array three times, as is its reverse complement. Thus, three replicate arrays yield 18 measurements per 50-mer. To train the model, we followed the protocol described by Peckham et al., first identifying the top 1,000 and bottom 1,000 probes by log ratio (see Methods). Each probe sequence is then represented as a 2,772-element vector, in which each entry is a normalized count of the occurrences of a particular k-mer or its reverse complement, for k = 1 up to 6. These vectors were used to train a support vector machine [26], which is a powerful discriminative classification algorithm that is widely used in diverse bioinformatics applications [27].
We evaluated the quality of the resulting classifier using a cross-validation procedure. In this procedure, the dataset is divided at random into ten subsets. An SVM is trained on 90% of the data (i.e., using 1,800 probes) and tested on the held-out probes. This train–test procedure is repeated ten times using a different hold-out set each time.
Figure 1 shows that the SVM trained on the microarray data from Dennis et al. is strongly predictive. The figure shows a receiver operating characteristic (ROC) curve [28], which plots the rate of true positives as a function of the rate of false positives for varying classification thresholds. To produce the ROC curve, each probe in the test set is ranked according to the discriminant score produced by the SVM. In this ranked list, a random classifier would not successfully separate positive from negative probes and would therefore produce a diagonal line at approximately y = x. On the other hand, a perfect classifier would rank all positive probes above all negative probes, producing a line that travels from the origin vertically to (0,1) and then horizontally to (1,1). The quality of a classifier can be evaluated by computing the area under the ROC curve (the “ROC score”), with 0.5 corresponding to chance and 1.0 corresponding to perfect separation. For the SVM trained on probes from the Dennis et al. dataset, the median ROC score across the ten cross-validations is 0.908, which is significantly better than chance. In a similar ten-fold cross-validation experiment, Peckham et al. report an ROC score of 0.951 from a similar SVM trained on the yeast dataset of Yuan et al. [14].
Next, we repeated the SVM cross-validation testing procedure using data generated by Ozsolak et al. [15]. As before, we selected the top 1,000 and bottom 1.000 probes for training and testing the SVM. Ozsolak et al. performed their analyses on seven different cell lines. Therefore, we trained seven different SVMs using data from each cell line. We also trained an eighth SVM by selecting probes that showed a consistently high or low level of fluorescence intensity across cell lines, using a rank-based selection procedure (see Methods). All eight ROC curves are included in Figure 1A. We compute the statistical significance of differences in ROC scores using the Wilcoxon signed-rank test. Figure 1B shows the results of pairwise comparisons among all of the SVMs, using a threshold of p<0.01.
Each of the eight SVMs performs significantly better than chance. However, the predictive power of the SVMs trained on different cell lines varies dramatically, from a median ROC score of 0.706 for T47D and MCF7 up to 0.880 for MEC. This variance does not correlate with the amount of replicate data available. The Ozsolak et al. dataset contains triplicate arrays for two of the cell lines (MALME and IMR90) and duplicate arrays for one cell line (A375). However, the best-performing SVM—MEC and A375, with median ROC scores of 0.880 and 0.878, respectively—are based on one and two microarrays, respectively. As shown in Figure 1B, these two SVMs perform significantly better than all of the other six SVMs.
Even between these two best-performing SVMs, the qualitative performance of the classifier differs significantly. This difference can be seen in the shapes of the respective ROC curves in Figure 1A. For MEC, the ranked list of test set probes contains a significant proportion of false positives near the top of the list, and very few false negatives near the bottom of the ranked list. Conversely, the SVM trained on data from the A375 cell line maintains a low false positive rate near the top of the list, but suffers a higher false negative rate lower down the list. Thus, the MEC SVM is good at recognizing regions of high accessibility to MNase, whereas the A375 is good at recognizing regions of strong protection from MNase cleavage. This characterization agrees well with the observation that the A375 cell line was digested with a different, less potent batch of MNase than was used for all the other cell lines [15].
Surprisingly, the SVM trained on all seven cell lines performs worse than five of seven SVMs trained on data from individual cell lines. A priori, an attractive model is that local patterns of nucleosome protection and exposure can be divided into tissue-specific and constitutive patterns. Presumably, the nucleosome positioning signal in regions that show consistent patterns across tissues would be encoded in the genome, which is static, whereas extra-genomic signaling would control tissue-specific positioning. Contrary to this model, the large variability in predictability among different cell lines and in particular the relatively poor performance of the SVM trained on probes that are consistently high or low across cell lines suggest that constitutive patterns of nucleosome protection and exposure are not preferentially encoded in the genome.
Overall, the SVM trained on the data of Dennis et al. performs much better than the SVM trained on the complete Ozsolak dataset, but performs comparably to the best-performing SVM trained on a single cell line. To determine whether the strong performance of the Dennis SVM can be explained by a systematic difference between the two datasets, we performed several additional experiments.
One important difference between the two microarray datasets that we investigated lies in the design of their probes. The Dennis et al. data spans −20 kb to +5 kb around the transcription start sites of genes selected based on their transcriptional response to ATP-dependent chromatin remodelers, whereas the Ozsolak et al. data covers only −1,250 to +250 bases around the transcription start sites of human cancer-related and randomly selected genes. If promoter regions contain systematic differences in nucleosome positioning relative to larger upstream regions, then these differences will be apparent in the two datasets.
Therefore, to reduce the differences between the Dennis SVM and the Ozsolak SVM, we trained an additional SVM. This time, we used probes from the Dennis et al. dataset, but we only considered probes that lie in the promoter-proximal region, as defined by Ozsolak et al. (i.e., −1,250 to +250 bases from the transcription start site). To ensure that we retain a similar degree of nucleosome protection or exposure as the initial Dennis SVM training set, we selected only the 150 probes with highest and lowest intensity, rather than 1,000 probes for each class. The resulting promoter-specific SVM performs slightly worse than the original Dennis SVM, achieving a median ROC score of 0.882. This is not significantly different from the ROC score of the original Dennis SVM, and it is comparable to the performance of the MEC, A375 and MALME SVMs.
A second difference between the Dennis and Ozsolak datasets is related to data processing. A significant concern with any microarray study is the possibility of bias resulting from probe hybridization artifacts. To combat this potential bias, Ozsolak et al. employ a wavelet denoising procedure which attempts to identify probes that show significantly higher or lower signal than either of their flanking probes. Dennis et al. did not perform any such denoising. It is therefore possible that our SVM performs well on the Dennis dataset precisely because it is able to learn to recognize these hybridization artifacts.
To further investigate whether the SVM is learning to recognize hybridization artifacts, we trained three additional SVMs. The first SVM is trained on the raw Ozsolak data, prior to wavelet denoising. The second and third SVMs are trained on wavelet smoothed versions of the Dennis dataset (“strong” smoothing and “weak” smoothing; see Methods). In all three cases, wavelet smoothing has no significant effect on the cross-validated ROC score. For the Ozsolak SVM, removing the wavelet smoothing changes the median ROC from 0.737 to 0.739. For the Dennis SVM, weak smoothing causes the median ROC score to decrease slightly, and strong smoothing causes the median ROC score to increase slightly. None of these differences is statistically significant. Thus, it does not appear that probe hybridization artifacts explain the difference in performance between the Ozsolak and the Dennis SVMs.
We next sought to determine whether yeast and human share common nucleosome positioning primary sequence features. Toward this end, we used the ENCODE regions, which span 1% of the human genome and systematically cover a range of gene densities and densities of conserved noncoding sequence. We tiled the ENCODE regions with nonoverlapping 50-mer probes and computed corresponding discriminant scores from the SVM trained using the A375 cell line of the Ozsolak dataset. We used the A375 SVM because, as shown above, this SVM excels at recognizing regions of strong nucleosome protection. We also applied two yeast-based models to these same probes: an SVM trained on the yeast microarray data from Yuan et al. [14], following the protocol of Peckham et al. [13], and the dinucleotide position-specific scoring matrix (PSSM) described in [11]. For the latter, we use 138 bp centered around each 50-mer probe, and we do not include the dynamic programming portion of the Segal et al. method.
The predictions made by the A375 SVM trained on human data correlate strongly with the predictions made by both of the models trained on yeast data. Figure 2 shows density heatmaps of scatter plots of predicted nucleosome positioning for human versus yeast models. The observed correlations for A375 are 0.862 (Peckham model) and 0.849 (Segal model). The corresponding correlations for the Dennis SVM (not shown), which also recognizes nucleosome protected regions, are 0.879 and 0.847. These correlations strongly suggest that the human-trained and yeast-trained models are learning to recognize a common sequence pattern.
DNA primary sequence plays a large role in the conformation of the double helix. More than 20 years ago, Drew and Travers [29] showed that certain dinucleotide frequencies are amenable to deformation around the nucleosome core. Many similar papers have since addressed the location and frequency of dinucleotide pairs in nucleosome formation. In order to understand the features of sequences that make them more or less suitable for organization around the histone octamer, we collected the 1,000 highest scoring probes identified by the A375 SVM within the ENCODE regions, and a complementary set of 1,000 lowest scoring probes identified by the MEC SVM (see Methods for details). These probes represent nucleosome-forming sequences and nucleosome inhibitory sequences, respectively. Compared to the training set probes, the ENCODE 50-mers represent a much smaller proportion of the genome—0.033% of the 30 Mbp ENCODE regions, compared to 0.9% and 1.8% of the probes on the Dennis and Ozsolak arrays, respectively. Also, unlike the microarray probes, the ENCODE 50-mers are not restricted to promoter-proximal regions of the genome. We analyzed dinucleotide usage within each of the ENCODE probe sets (Figure 3). The sequences identified by the A375 SVM as nucleosome-forming sequences are rich in CC/GG dinucleotides (41.8%). Conversely, AA/TT, AT, and TA sequences are severely underrepresented in these probes, collectively accounting for only 1.7% of the dinucleotides. One concern when using datasets generated from MNase cleavage involves the slight sequence bias of MNase for AT sequences (Wingert and VonHippel, 1968). However, this sequence preference is unlikely to account for the vast differences seen in the AT underrepresentation in the nucleosome-forming probes. Among the bottom ranked, nucleosome inhibitory sequences, we find a clear overrepresentation of AC/GT and CA/TG dinucleotides (69.7%).
The observations regarding dinucleotide abundance prompted us to analyze the periodic nature of their occurrence. We computed the distance between a given dinucleotide and each of its identical neighbors and then plotted counts of all pairwise distances in a single histogram for each dinucleotide (Figure 4). Analysis of the top ranked nucleosome-forming sequences reveals a clear 3 bp periodicity of CG and GC dinucleotides with a distance of 3 bp. Three base pairs represents about one third of a helical turn of the DNA helix, and periodic occurrences of GG/CC and GC dinucleotides have been implicated as a nucleosome positioning signal in human DNA [30],[31].
The GC 3 bp periodicity is particularly interesting because many lines of evidence implicate trinucleotide repeats of CTG and CAG as potent nucleosomal occupancy positioning signals. For example, CTG repeats do not appear to have any rotational phasing preference with respect to the core nucleosome [32], and relatively straight DNAs characterized by CTG repeats have a high affinity with the histone octamer in reconstitution experiments [33],[34]. Furthermore, sequences of DNA sequences containing 75 or 130 CTG repeats were shown to form nucleosomes 6 and 9 times more strongly, respectively, than the 5s rDNA, a naturally occurring nucleosome positioning sequence [35]. Surprisingly, DNAse I experiments have indicated that CTG repeats are among the most flexible trinucleotides, indicating that this relatively straight DNA has the potential to bend [36]. CAG repeats, on the other hand, were found to be enriched in nucleosome positioning sequences from the mouse genome [37] and are found frequently within centromeric satellite repeats. Perhaps most significantly, a crystal structure of the nucleosome was solved using a palindromic alpha satellite DNA sequence [38] which contains many degenerate forms of CNG runs. These CNG-type runs occur on one edge of the histone octamer, where they hug the form of the nucleosome along positions in the structure where a sharp DNA-deforming bend is not required. The location of these runs is consistent with all of the above observations. It therefore seems reasonable that the lack of rotational phasing preference, strong binding characteristics and inherent flexibility would make sequences containing this type of repeat nucleosome forming. Additionally, because this type of trinucleotide repeat does not appear to have any rotational phasing with respect to the nucleosome core particle, it is not surprising that our periodicity analysis does not reveal any 10 bp periodicity within these sequences.
Low complexity DNA sequence appears to be the hallmark of the nucleosome inhibitory probes. Eight of the ten dinucleotide periodicities appear as simple dinucleotide repeats. This is interesting in light of the fact that these probes were selected from a repeat masked library, indicating that short stretches of simply repeating sequence may have been evolutionarily constrained to ensure a fluid DNA chromatin structure at particular loci. Additionally, for several of these low complexity repeats—AG, AT, CG, GA, GC, and TA—the simple dinucleotide repeat persists for runs of less than 15 bases, indicating that a relatively short stretch of dinucleotide repeat appears sufficient to destabilize nucleosomal organization of the DNA sequence. This behavior was predicted by Drew and Travers [29] when they postulated that runs of homopolymer AT or GC would be excluded from the central region of nculeosomes due to their relative inflexibility.
We next aimed to characterize the nucleosome formation potential of the sequences flanking the low scoring probes. Nucleosome-free regions in chromatin arise secondary to the cooperative binding of sequence-specific DNA binding factors [39]. The degree to which the primary DNA sequence contributes to this process beyond supplying the protein binding motifs is unknown. If certain DNA sequences are less stable in the context of a nucleosome (“nucleosome-excluding sequences”), then these regions would presumably provide more fertile ground for the nucleation of transcription factors.
Every nucleosome-depleted region should be, by definition, flanked by nucleosome-occupied regions. Indeed, efficient positioning of nucleosomes may potentiate the coalescence of regulatory factors in the intervening region by providing higher-order stability following loss of the central nucleosome. We therefore next investigated nucleosome-disfavoring sequences in the context of their local chromatin environment. In particular, we wanted to investigate if these sequences were flanked by regions with higher than average nucleosome stabilizing sequences. Consequently, for this analysis, we used the A375 SVM, which identifies such nucleosome-forming sequences with high accuracy. We aligned the 1,000 lowest scoring probes from the A375 SVM, averaged the scores and symmetrized them to remove strand-specific artifacts. The results are striking (Figure 5, left panel): the nucleosome inhibitory sequence is flanked by probes whose discriminant scores oscillate, suggestive of two flanking well-positioned nucleosomes. These results echo what has been seen in the nucleosome-free region of yeast promoters: the nucleosome-free region is surrounded on either side by regularly spaced nucleosomes [14]. A similar oscillatory pattern is observed when we produce these plots using the Dennis SVM, but not when we use the MEC SVM (Figure S1). The absence of this pattern in the latter case presumably arises because the MEC SVM only recognizes nucleosome inhibitory sequences, rather than nucleosome-forming sequences.
Because nucleosome inhibitory sequences are flanked by sequences that have high nucleosome formation potential, we next sought to understand the characteristics of sequences adjacent to high scoring probes. In particular, we wanted to investigate if these sequences were set in the center of a particular pattern of nucleosome formation potential. We aligned the top-scoring 1,000 probes collected previously, and plotted average symmetrized scores, as before. In Figure 5, right panel, we see a linear dropoff on either side of the probe and no periodic information. These results indicate that a high scoring probe does not confer a pattern of nucleosomal occupancy in the surrounding sequence. The linear dropoff to baseline, as well as the magnitude of the SVM discriminant score in the flanking region, indicates that these probes are found in regions of relatively higher nucleosomal formation potential. These probes may be critical in facilitating nucleosomal occupancy at these particular locations, with the histone octamer likely adopting many positions over a window greater than 150 bp.
An orthogonal means of validation is to examine the pattern of predicted nucleosome occupancy near genomic landmarks that are known to impact nucleosome positioning. We therefore examined the pattern of predicted nucleosome occupancy near transcription start sites. We expect nucleosome occupancy to decrease on average in promoter regions, because bound nucleosomes can impede promoter activity. We collected 2 kb regions centered around transcription start sites identified in the Gencode annotation [40]. Figure 6 shows, for the A375 and MEC SVMs, the average predicted nucleosome occupancy in these 1 kb regions. The A375 SVM shows a pronounced peak over the start of the gene; conversely, the MEC SVM shows a very strong dip just upstream of the TSS. A pattern similar to that produced by the MEC SVM is observed for the other six cell lines (data not shown), all of which were digested with the same batch of MNase as the MEC cell line. These results are consistent with a model in which the start of the gene is occupied by a well positioned nucleosome, which has the potential to become nucleosome-free upon gene activation. The A375 SVM recognizes the former effect, and the MEC SVM recognizes the latter.
Oszolak et al. further partitioned the genes on their array into those that are expressed and those that are not expressed in A375 cells. When the average log2 nucleosomal/bare genomic ratios of each of these promoter subsets are aligned at transcription start sites an interesting pattern emerges. The unexpressed genes show a nucleosomal occupancy profile very similar to that of the A375 SVM model: higher nucleosomal occupancy at the transcription start site with relatively lower occupancy in the promoter region. It is only in the expressed genes that a chromatin structure reflective of a nucleosome-free region flanked by positioned nucleosomes is seen. This is in keeping with a nucleosome remodeling event regulating gene expression.
We next sought to understand if the high scoring probes were enriched proximal to transcription start sites. In order to accomplish this we used our 1,000 top and bottom scoring probes from the ENCODE regions and plotted their proximity to transcription start sites from the Gencode annotation. For the A375 trained SVM, 589 out of the top 1,000 high scoring probes lie within 1 kb upstream or downstream of a transcription start site.
The above results are in keeping with whole genome studies that find lower nucleosomal occupancy in promoter regions than coding regions [17],[19],[41]. Our results indicate a higher potential for nucleosomal occupancy at and around the transcription start site. Spontaneous histone removal or unwrapping is too slow to account for the tight control and regulation of gene expression. ATP-dependent chromatin remodeling complexes facilitate the relocation of DNA sequence elements relative to the histone octamer, thus playing a critical role in gene regulation [42]. Moreover, recent studies have shown that diverse ATP-dependent remodeling complexes confer a repositioning specificity directed at least in part by DNA sequence information [43].
We next sought to understand whether the sequence features identified by the SVM are consistent with nucleosome positions adopted in in vitro experiments. The specific locations of nucleosomes along the DNA sequence may play both inhibitory and activating roles in nuclear processes. One way to alter the position of nucleosomes relative to DNA sequence is through the action of ATP-dependent remodelers. These remodelers may use cues from the primary sequence to restrict the remodeling to certain genomic loci or confer a directionality upon the movement of the DNA with respect to the histone octamer. Using the positions of nucleosomes and remodeled products on the 202 bp TPT fragment determined by Fan et al. [44], we sought to understand the characteristics of the starting and remodeled product. We generated SVM discriminant scores for the 202 bp TPT fragment by dividing the sequence into overlapping 50-mers. We then took the resulting discriminant scores and overlaid the starting and remodeled positions of the histone octamer along this sequence. The starting position of nucleosome at the 5′ end of the sequence clearly shows higher discriminant scores, indicating that these are indeed nucleosome-forming sequences (Figure 7, solid line). The dominant starting nucleosome position is centered directly over the 50 bases with the highest scoring discriminant score. In this set of experiments the SVM predicts the starting position of the histone octamer on the DNA sequence.
At 1.5 helical turns on either side of the histone octamer dyad axis, nucleosomal DNA is severely bent to accommodate necessary contacts with the H3/H4 tetramer [38],[45],[46]. Extreme DNA curvature is known to potentiate nucleosome positioning in vitro and under physiological conditions [47],[48]; however, such extreme curvature is likely to be rare in vivo. Computational curvature scores have been explored as predictors of nucleosome positioning [47]–[49]. We therefore compared the SVM discriminant score (Figure 7, solid line) to the DNA curvature scores (Figure 7, dotted lines). The two signals do not correlate, indicating that the SVM scores are picking up on signals that discriminate on features not limited to the bending potential of DNA. Thus, information from both the SVM discriminant scores and DNA curvature predictions may be used to understand primary sequence determinants of nucleosome position. On the TPT+45 the highest scoring curvature region occurs just downstream of the nucleosome dyad axis. Thus, it appears for this sequence that the general placement of the nucleosome can be determined by elements of the primary sequence as recognized by the SVM, and the fine position is determined by the potential to deform the DNA around the histone octamer.
Finally, we investigated how remodeled nucleosome states correlate with SVM score by comparing mapped positions of nucleosome starting positions and their remodeled products on the 202 bp TPT fragment [44]. Recently, Rippe et al. [43] proposed a model in which the remodeling activities of different remodelers depends upon sequence features of the DNA sequence of the nucleosome. The Sf2h remodeled position from Fan et al. [44] occupies regions of high SVM discriminant score and shifts the DNA fragment of high curvature to the region just upstream of histone octamer dyad axis. The nucleosome is translationally repositioned to the closest curvature amenable position within the context of the highest scoring region of nucleosome occupancy as predicted by the SVM discriminant scores. Thus, it appears that the primary DNA sequence signals, as determined by curvature and SVM discriminant scores, may be used to predict and understand nucleosomal occupancy of both starting and remodeled nucleosomal states.
Our results suggest that the human genome contains sequence-based signals that contribute to the placement of nucleosomes. By focusing only on the top- and bottom-ranked probes in microarray datasets, we operate under the hypothesis that only a subset of nucleosomes are well-positioned. This hypothesis agrees with the statistical positioning theory of nucleosome organization and is supported by two previous studies in yeast, each of which concluded that a sequence-based model explains only 15% more nucleosome positioning than is explained by a random model [11],[13]. We have shown, using cross-validation on two independent datasets, that a subset of human nucleosome positions can be accurately predicted using an SVM trained only on sequence patterns.
Additionally, the model of nucleosome occupany affirms earlier work suggesting that DNA sequences shorter than the 150 bp canonically protected by a nucleosome may make a significant contribution to nucleosome occupancy [8],[50]. Great strides have recently been made in understanding primary sequence determinants of nucleosome formation in yeast and worm models [11],[51]. Each of these experiments has used entire core sequences from nucleosome that occupy a constant genomic location: the same protection of approximately 150 bases is occupied by the nucleosome in all cells. However, a relatively small percentage of nucleosomes are well positioned in eukaryotic genomes [51]. Our model is trained using primary sequence features of occupancy alone and does not preselect for nucleosomes that occupy a single position in the genome. This may account for the discrepancies between yeast occupancy models and our human model. Our strategy allows for the analysis of short sequence features that determine the nucleosome-forming or inhibitory potential of a given short sequence of DNA.
The likelihood that a short sequence of DNA will be incorporated into a nucleosome depends upon two features of the DNA: the ability of the sequence to form specific interactions with the histone core and a sequence dependent DNA curvature that allows it to wrap around the histone core. Seminal work by Drew and Travers [29] addressed DNA bending and its relation to nucleosome position. Since then, rules and codes for nucleosome positioning have been defined in terms of the intrinsic flexibility of a particular sequence of DNA. Herein, we show that our model accurately identifies the correct nucleosome position on a nucleosome positioning sequence. Moreover, we show that the model recognizes something other than DNA curvature signals. Thus, our model of nucleosome occupancy may be combined with curvature predictions to better understand and predict locations of nucleosomes and how nucleosomal occupancy may be altered by ATP dependent remodelers.
We observed a complementarity in the SVMs trained using strongly versus weakly digested DNA samples. SVMs trained from weakly digested samples, such as the A375 sample from Ozsolak et al. or the data of Dennis et al., accurately predict regions of strong protection from MNase cleavage, which presumably correspond to well-positioned nucleosomes. These models also correlate strongly with two previously described models of nucleosome positioining in yeast. In contrast, SVMs trained from more completely digested DNA samples, such as the MEC data from Ozsolak et al., are very good at recognizing positions of increased MNase cleavage. These occur, for example, in promoter regions, where strong nucleosome placement would presumably impede the transcriptional apparatus.
Nucleosomal occupancy and chromatin structure have functions in the regulation of transcription [42]. The concept that promoters and classical cis-regulatory elements such as enhancers represent nucleosome-free regions was well-established in the literature by the mid-1980s, and derived from chromatin structure assays performed on hundreds of eukaryotic genes from a variety of species [52]. Multiple reports using genomic technologies have recently emerged in support of the generality of this principle across the yeast genome [17],[19],[41],[53]. The low SVM discriminant scores upstream of the TSS indicate that primary sequence features play a role in the regulatory nature of chromatin. These low occupancy scores point toward roles that the primary sequence may play in chromatin remodeling (i.e., the transient repositioning of nucleosomes) and histone eviction. Additionally, recent studies of chromatin remodeling have shed considerable light on the relationship between nucleosome occupancy and competitive interaction with regulatory factors [54]–[57]. Our results suggest a powerful role for the SVM in prediction of nucleoseome-disfavoring sequences, which are fertile ground for regulatory protein interactions.
An advantage of our models of nucleosome occupancy is that we are able to scan for possible nucleosome-free regions in a manner that is unbiased by genomic location or regulatory factor binding. Indeed, multiple laboratories have shown well-positioned nucleosomes flanking regulatory nucleosome-free regions [14],[58]. Thus, the model of nucleosome occupancy may act as a tool for the de novo prediction of regulatory regions within the human genome.
Taken together, our results suggest that we have developed a potentially powerful model of short-range nucleosomal organization that can predict the location of genomic regions that may have an intrinsic predisposition to harboring cis-regulatory elements. Recent work [59] suggests that the nucleosome data of Yuan et al. contains significant long-range correlations. Consequently, a future model that explicitly includes these correlations might successfully capture features of nucleosome organization at a larger scale than were examined in the current study.
We emphasize that our results are fundamentally consistent with the long-standing statistical positioning theory of nucleosome organization, which posits that nucleosome positioning is largely secondary to strong non-nucleosomal boundary events. Our model may be used to generate novel testable hypotheses concerning the role of nucleosome-DNA interactions in transcription and other chromosomal regulatory processes.
The Dennis dataset contains three arrays with each probe spotted three times in each orientation, yielding a total of 18 measurements for each of 56,633 50-mer loci. To identify top- and bottom-ranked probes, we follow a five-step procedure. First, we eliminate all probes that overlap a repetitive element as identified by RepeatMasker. Second, we convert the log-ratio intensity values on each array into ranks. Third, for each locus and each strand, we sum the corresponding nine ranks. Fourth, we sort these rank-sum values into a single list. Fifth, we move a threshold down this list from highest to lowest values, accepting into the list of top-ranked probes any probe whose forward and reverse rank-sums are above the current threshold. We stop the traversal when 1,000 top-ranked probes have been identified. The bottom-ranked probes are identified in a similar fashion, traversing the list in the opposite direction.
Top- and bottom-ranked probes in the Ozsolak data are identified using a different rank-based procedure. The Ozsolak dataset contains seven different cell types. Initially we compute a single value for each probe: for IMR90 and MALME with three samples, this single value is the median; for A375 with two samples, the single value is the mean; the remaining four cell types (MEC, MCF7, PM, T47D) have a single sample each. To identify top-ranked probes, we then produce a single sorted list containing all of these values. Traversing this list from largest to smallest value, we accept a probe into our list of top-ranked probes when 5 of the 7 cell types have been observed in the sorted list, as long as no probe within 50 bp has already been accepted into the list. We continue this procedure until 1,000 probes have been accepted. The list of bottom-ranked probes is derived in a similar fashion, traversing the list in reverse order.
Prior to analysis by the SVM, each 50-mer probe is converted into a vector in which the entries are normalized counts of occurrences of all possible k-mers for k = 1···6. Combining reverse complements leads to a total of 2,772 entries in this vector. For each value of k, the corresponding counts are normalized so that the sum of their squares is 1. Thus, the final vector resides on a sphere with radius 6 in a 2,772-dimensional space.
SVMs are trained using svmvia (http://noble.gs.washington.edu/proj/svmvia), which implements the entire regularization path optimization algorithm [60]. For each training set, the regularization parameter is selected to maximize the ROC score computed on a hold-out set from within the training set. The SVM uses a linear (dot product) kernel.
Smoothed microarray values are generated using wavelets. The maximal overlap discrete wavelet transform (MODWT) [61] is applied to raw replicate array values from each array separately. Given the input resolution of approximately 20 bp, we construct jth-level smoothing for j = 2 and 3, giving an output smoothing window of approximately 80 and 160 bp, respectively. The jth-level smooth is constructed using the MODWT multi-resolution analysis using the “la-8” wavelet family. Calculations are performed in R using the waveslim library http://cran.rproject.org/src/contrib/Descriptions/waveslim.html.
Predicted highly occupied positions are identified by ranking the ENCODE 50-mers by SVM discriminant, and then traversing the ranked list from highest to lowest, accepting any 50-mer that is not marked as repetitive or low complexity and that is more than 250 bp away from any previously accepted prediction and continuing until 1,000 50-mers have been accepted. A similar procedure is carried out in reverse to identify sites of low occupancy.
|
10.1371/journal.pgen.1000152 | The Identification of Zebrafish Mutants Showing Alterations in Senescence-Associated Biomarkers | There is an interesting overlap of function in a wide range of organisms between genes that modulate the stress responses and those that regulate aging phenotypes and, in some cases, lifespan. We have therefore screened mutagenized zebrafish embryos for the altered expression of a stress biomarker, senescence-associated β-galactosidase (SA-β-gal) in our current study. We validated the use of embryonic SA-β-gal production as a screening tool by analyzing a collection of retrovirus-insertional mutants. From a pool of 306 such mutants, we identified 11 candidates that showed higher embryonic SA-β-gal activity, two of which were selected for further study. One of these mutants is null for a homologue of Drosophila spinster, a gene known to regulate lifespan in flies, whereas the other harbors a mutation in a homologue of the human telomeric repeat binding factor 2 (terf2) gene, which plays roles in telomere protection and telomere-length regulation. Although the homozygous spinster and terf2 mutants are embryonic lethal, heterozygous adult fish are viable and show an accelerated appearance of aging symptoms including lipofuscin accumulation, which is another biomarker, and shorter lifespan. We next used the same SA-β-gal assay to screen chemically mutagenized zebrafish, each of which was heterozygous for lesions in multiple genes, under the sensitizing conditions of oxidative stress. We obtained eight additional mutants from this screen that, when bred to homozygosity, showed enhanced SA-β-gal activity even in the absence of stress, and further displayed embryonic neural and muscular degenerative phenotypes. Adult fish that are heterozygous for these mutations also showed the premature expression of aging biomarkers and the accelerated onset of aging phenotypes. Our current strategy of mutant screening for a senescence-associated biomarker in zebrafish embryos may thus prove to be a useful new tool for the genetic dissection of vertebrate stress response and senescence mechanisms.
| By performing genetic mutant screens using senescence-associated biomarkers, we show that the zebrafish is a tractable model system for the study of aging. In vertebrate organisms, it has not previously been possible to carry out systematic screens for genes that are important for stress responses and aging in an unbiased way. However, such vertebrate models are of considerable importance, given the provocative evidence of common biochemical and functional pathways modulating stress responses and lifespan as well as aging in a wide range of organisms. Our present study has successfully employed a colorimetric high-throughput method using a senescence-associated β-galactosidase-based assay to screen for mutations that alter the stress responses in zebrafish embryos, in the hope that these might represent potential aging mutants. Subsequently, the mutations identified by embryonic senescence have indeed displayed adult aging-related phenotypes in zebrafish. Hence, our method for the identification of mutant zebrafish has the immediate potential to accelerate the discovery of novel genes and new functions relevant for our understanding of aging processes in vertebrates. Such knowledge will be essential for the ultimate development of pharmacological, nutritional, and behavioral interventions for the amelioration of oxidative stress- and age-associated diseases and disabilities in humans.
| Chronic oxidative stress has been shown to reduce lifespan in many species and lead to accelerated aging [1]–[3]. It has also been reported that oxidative stress is involved in neurodegeneration, sarcopenia and other muscle wasting conditions, which are accompanied by multiple aging symptoms [4]–[6]. Reactive oxygen species (ROS) are generated during normal cellular metabolism, primarily as a result of inefficiencies in the electron transport chain during mitochondrial respiration. Optimally localized levels of ROS serve functionally in the activation of some signal transduction pathways. However, ROS can also cause damaging chemical modifications of macromolecules such as proteins, lipids and DNA, which can in turn contribute to the progression of neurological diseases and neuromuscular disorders including Huntington's disease, Parkinson's disease, Alzheimer's disease, amyotrophic lateral sclerosis, and ataxia telangiectasia [4],[7].
The genetic regulation of the stress and damage response pathways in vertebrates may be more complex than that seen in simple model organisms such as Drosophila and C. elegans. However, a strong case can be made for repeating the genetic screens performed in these lower organisms, in a vertebrate model, to identify genes regulating oxidative stress. Such analyses have the potential to identify candidate genes related to multiple stress- and age-associated diseases in humans. However, due to the challenges of performing large-scale forward genetic screens in mice, it would be of considerable benefit if the high-throughput screening technology used in simpler organisms (i.e., invertebrates) can be adapted for use in zebrafish (Danio rerio), a vertebrate model in which forward genetic screens are routinely performed [8]–[10]. The zebrafish is inexpensive to maintain and has favorable characteristics for experimentation such as a high fecundity, rapid external development, embryonic translucence, and ease of genetic manipulation. In addition, the sequence of the zebrafish genome, while not yet completely annotated, has already revealed a high degree of similarity between fish and human genes.
Thus far, we and two other groups have mainly contributed to establish important baseline information validating the use of zebrafish as a valuable model for aging studies [11]–[16]. We have extensively searched for various biomarkers of aging in zebrafish [17]. However, to faithfully monitor the wide-ranging in vivo effects of several stresses on senescence and aging in zebrafish in a high-throughput manner, we required a reliable and easily applicable biomarker that robustly indicates presence of oxidative stress during embryonic development as well as symptoms of aging in adults. One obvious candidate was senescence-associated β-galactosidase (SA-β-gal), a marker of cellular senescence in vitro as well as of organismal aging in vertebrates [16], [18]–[21]. Importantly, genes known to cause embryonic senescence can be detected by SA-β-gal in mice [22],[23]. Mounting evidence suggests that the identity of SA-β-gal is in fact the well characterized lysosomal β-galactosidase enzyme, which is most active at a much lower pH, but has some minimal activity at pH 6.0 where it can be detected when abundant [24],[25]. The cellular lysosomal content increases in aging cells due to the accumulation of non-degradable intracellular macromolecules and organelles in autophagic vacuoles [26]. Thus, lysosomal β-galactosidase induction could represent a general adaptive response to cellular senescence. Oxidized protein and lipid by-products that cannot be degraded by lysosomal hydrolases nor be exocytosed accumulate over time in post-mitotic cells, and are not diluted by cell division. One such by-product is lipofuscin, also known as “age pigment” [27]. Lipofuscin is composed of cross-linked protein and lipid residues [28],[29] and is generated by iron-catalyzed oxidative processes as well as by the incomplete degradation of damaged mitochondria [30],[31]. It has previously been demonstrated that both oxidative stress and aging promote lipofuscin accumulation [32].
In our current study, we demonstrate that the level of SA-β-gal is elevated in zebrafish embryos exposed to acute but sub-lethal levels of oxidative stress as well as in aged adults. We then present two genetic screens for mutants in stress responses that might also display altered aging phenotypes. We first examined a collection of retrovirus-mediated insertional mutants for the embryonic induction of SA-β-gal. The screening of these mutants for modified SA-β-gal activity could be performed relatively quickly in the absence of an extrinsic insult and stress, since the homozygous embryos can be identified by their morphology. We found from our results that, of the 306 insertional mutations we examined, at least 11 scored as having significantly elevated SA-β-gal production levels in homozygous mutant embryos, two of which we present in further detail herein. The mutation which resulted in the highest SA-β-gal levels caused an inactivation in a gene encoding the zebrafish homologue of the Drosophila spinster, which is responsible for regulation of aging and lifespan in flies and has been implicated in a lysosomal storage function [33],[34]. One of the other mutants inactivated the telomeric repeat binding factor a (terfa) gene, a zebrafish homologue of the telomeric repeat binding factor 2 (terf2) gene, which plays prominent roles in telomere protection and telomere-length regulation [35],[36].
For our second screen, we developed a new zebrafish mutant screening protocol based upon N-ethyl-N-nitrosourea (ENU) chemical mutagenesis. We performed a sensitized dominant screen in the zebrafish to detect mutations in the heterozygous state by using a chemical sensitizer (rather than a genetic sensitizer). In our pilot screen using this methodology, we obtained eight mutants in two complementation groups that showed altered SA-β-gal activity in response to oxidative stress. Importantly, adult fish that were heterozygous for several of these mutations also showed premature expression of aging markers/phenotypes, and a shorter lifespan. Our new screening strategy using a senescence-associated biomarker during the embryonic stages in zebrafish provides a new tool for the genetic dissection of vertebrate stress responses and aging mechanisms. Moreover, our initial results strongly suggest that genetic lesions in certain early developmental mechanisms lead to late adult-onset phenotypes with age.
To further characterize aging in the adult zebrafish, we have previously examined several potential biological and biochemical markers, including regenerative competence and assays for the oxidative damage of proteins, lipids and DNA [16],[17],[19]. The most reliable and readily detectable age-dependent marker was determined to be a histochemical assay for SA-β-gal activity, which can be quantitatively applied to whole adult zebrafish using X-gal as a substrate at pH 6.0 [16]. In our current experiments, staining for SA-β-gal was found to increase in the skin of zebrafish with age throughout their lifespan (n = 139) (Figure 1A, B), as was previously reported in both humans and zebrafish [16],[18],[19]. To quantitatively examine SA-β-gal levels in vivo, we generated high-resolution digital images that enabled us to select stained pixels using image analysis software and to then calculate the percentage of stained pixels out of the net total in each case (Figure S1). Unlike other markers that tended to vary discontinuously with age, we found that SA-β-gal activity increases linearly with age in adult fish ranging in age from 5 to 57 months (Figure 1C).
We hoped to avoid the need to screen for aging mutants using actual lifespan analyses if we instead screened embryos for mutations that alter the expression of aging markers in response to oxidative stress [17]. For such an approach to succeed, we surmised that the chosen biomarker must respond both to aging in adults and to stress responses during embryonic development. Hence, we tested whether the SA-β-gal assay would also respond to oxidative stress in embryos treated with ROS such as hydrogen peroxide (H2O2) or tert-butyl hydroperoxide (BHP) (Figure 2A–D). Several different doses of these peroxides were used over a developmentally long period to better simulate long-term chronic oxidative stress. An experimental endpoint at 6 days post fertilization (dpf) was chosen to avoid potential spurious effects of caloric restriction or other nutritional deficiencies, as this is the point in larval development at which the supportive yolk has been consumed and the fish begin to eat and rely on oral intake nutrition. The LD50 values for H2O2 and BHP were measured at approximately 300 µM and 1 mM, respectively, under these assay conditions. At sub-lethal doses of peroxides, SA-β-gal levels increased in a roughly linear fashion with increasing concentrations of H2O2 and BHP to the maximum tolerated doses of 150 µM and 500 µM, respectively. Compared with the untreated controls (n = 50) (Figure 2A), zebrafish embryos treated with either 150 µM of hydrogen peroxide (n = 50) or 500 µM of BHP (n = 50) displayed an approximately 3-fold increase in SA-β-gal staining intensity following six days of development (Figure 2B–D). These results suggested that SA-β-gal-based screens of chemically or genetically stressed embryos could indeed be used to identify senescence-related mutants in zebrafish.
To test whether the induction of SA-β-gal activity in zebrafish embryos that have been exposed to oxidative stress occurs in a similar manner to that reported in other organisms, we performed genetic manipulations of a ROS detoxification enzyme in vivo. A number of studies in a variety of species have shown that both catalase and glutathione-peroxidase are responsible for antioxidant protection by limiting the accumulation of hydrogen peroxide [2],[3],[37],[38]. To ascertain the potential importance of catalase in protecting zebrafish embryos from oxidative stress-induced senescence, we altered the expression levels of this enzyme in stressed embryos and measured the effects of this upon SA-β-gal activity. Embryos overexpressing zebrafish catalase were generated by the injection of 300 pg of mRNA encoding this enzyme at the one-cell stage (n = 50). This resulted in a reduction in both hydrogen peroxide- and BHP-induced SA-β-gal activity, compared with control GFP mRNA injections (n = 50) (Figure 2E and data not shown). BHP was used as the oxidative agent throughout the later stages of this study as it is more stable than hydrogen peroxide and produces less variable stress responses. The most dramatic rescue effects were observed when intermediate concentrations of BHP were used in these catalase experiments.
We additionally tested if a reduction in the catalase expression levels would enhance the induction of SA-β-gal activity by oxidative stress. To this end, we injected embryos with an antisense morpholino oligonucleotide (MO) targeting zebrafish catalase and, indeed, observed a marked enhancement in their susceptibility to elevated SA-β-gal activity following exposure to oxidative stress (Figure 2F). The greatest effects were again observed when intermediate concentrations of BHP were used. These results confirmed that the manipulation of a single gene can modulate the SA-β-gal activity levels induced by oxidative stress in zebrafish embryos and prompted us to pursue a genetic screening project to uncover potential aging mutants.
We hypothesized that a loss-of-function (or even partial loss-of-function/decrease-of-function) mutation in certain genes may induce specific stress conditions in mutant embryos. To identify potential aging mutants, we first screened for mutants with an altered production of the stress response marker SA-β-gal in an established zebrafish mutant collection generated by retrovirus-insertional mutagenesis [39]. Currently, the Hopkins' insertional mutant collection contains more than 500 recessive mutants with morphological embryonic phenotypes, which include mutations in 335 different identified genes [10],[40],[41]. We screened unstressed homozygous embryos derived from incrossed heterozygotes from 306 of these lines for SA-β-gal expression at 3.5–5 dpf, depending upon the onset of the morphological phenotype.
In general, the levels of SA-β-gal seen in the homozygous mutants were low, with only 11 mutants clearly scoring robustly higher than wild-type background activity (Figure 3; Figures S3 and S4) (Table 1). It should be noted also that since all of the 306 mutations screened are ultimately homozygous lethal, these data indicate that SA-β-gal production is not a general result of embryonic death. Figure S4 shows several examples of embryonic lethal mutants whose SA-β-gal levels are no higher than (or indistinguishable from) their wild-type siblings despite varying amounts of cell death. Similarly, the cloche (clo) mutant, which has no circulatory system did not show detectable SA-β-gal induction above background activity (n = 54) (Figure 3C). However, for 11 of the lines, the mutant embryos showed significantly stronger SA-β-gal staining than their wild-type siblings. For example, as shown in Figure 3D, an insertional mutation in the atp6v1h gene encoding the V1 subunit H component of vacuolar ATPase (v-ATPase), a multi-subunit enzyme that mediates the acidification of eukaryotic intracellular organelles, is one of the 11 mutants identified that showed robust levels of SA-β-gal induction (n = 45).
We chose to study two out of 11 of the insertional mutants in more detail based upon previous knowledge about the mutated genes in other organisms. Of these insertional mutants, the highest SA-β-gal activity was found to be associated with an insertion in the gene denoted “not really started” (nrs) (currently denoted as zebrafish spinster homolog 1, spns1) (hi891) (nrs mutant, n = 135; wild type, n = 185) (Figure 3B) [42]. The nrsm/m homozygotes die by 4 dpf and show a substantial accumulation of an opaque substance in the yolk (Figure 3E, indicated by a black arrow in the left lower panel). Furthermore, the nrs gene has been identified as the zebrafish homologue 1 (spns1) of the Drosophila spinster (spin) gene. A Drosophila partial loss-of-function (hypomorphic) mutant for the spinster gene accumulates lipofuscin granules in the central nervous system, accompanied by neurodegeneration and abnormal ovary development. Notably, Drosophila hypomorphic spin mutants also have a shortened lifespan [33].
The other mutant which we focused on exhibited an insertion in the “telomere repeat binding factor a” (terfa) gene. The mutant lines hi3678 (n = 155) and hi1182 (n = 120) (Figure 4A, lower image for hi3678 and Figure S2B, lower right panels for hi1182) had significantly higher SA-β-gal activity compared with wild-type controls (n = 100 for each) (Figure 4A, upper image; Figure S2B, lower left panel). The terfa gene is a zebrafish homologue of the human terf2 gene which encodes the telomeric repeat binding factor 2 protein (TRF2). Due to the varied nomenclature for terfa in other species, we denote the zebrafish gene as terf2 hereafter. TRF2 has an essential role in telomere end protection and t-loop formation [35],[43],[44]. Moreover, the disruption of endogenous TRF2 function in human cells by expressing dominant-negative forms of this protein markedly increases the rate of telomere end-to-end fusions and cellular senescence [36]. A deletion of the terf2 gene in mouse embryonic fibroblasts also results in a senescence-like arrest and SA-β-gal induction [45]. Hence, the SA-β-gal induction that we see in our zebrafish terf2 mutant embryos is consistent with the established biological role of this gene and the results of previous studies in other organisms.
Telomeres of homozygous terf2 embryos were visualized by cross mating with transgenic fish expressing a green fluorescence protein (GFP)-tagged human TRF1/Pin2 fusion protein [46]. While many telomere speckles were evident in the wild-type background (n = 20) (Figure 4B, upper left panel), we observed enlarged telomere speckles and abnormal nuclear shapes in terf2m/m fish embryos (n = 24) (Figure 4B, lower left and right panels), which are likely to reflect telomere end-to-end fusions and impaired chromosome integrity. Moreover, homozygous terf2m/m mutant zebrafish embryos showed aggressive neurodegenerative phenotypes in the eye, brain, and spinal cord (n = 32) (Figure 4C, right panel; Figure 7W–Y), compared with wild type (n = 10) (Figure 4C, left panel; Figure 7T–V). In contrast to normal retinal development in wild-type embryos (n = 5) (Figure 4D, left panel), embryonic retinas stained with phalloidin in order to visualize actin filaments in plexiform layers revealed obvious structural defects in terf2m/m mutants (n = 10) (Figure 4D, right panel). Neurodegeneration in the retina was also detected histologically by performing Fluoro-Jade B staining of terf2m/m embryos (n = 10) (Figure 4E, right panel), compared with normal wild-type retinas (n = 5) (Figure 4E, left panel). Significantly, the neural phenotypes associated with a terf2 mutation appear to be consistent with the recent observations of mammalian TRF2 function reported in neural cells in vitro [47],[48]. While we did not examine other mutant lines at this level of detail, it was clear that some of the other insertional mutants with elevated SA-β-gal levels also exhibited widespread cell death in the central nervous system and eyes (Table 1).
To further examine the SA-β-gal induction caused by disruptions of the nrs and terf2 genes, we knocked down the translation of their respective mRNAs using MOs that target the start codon of each gene. Injection of a nrs MO at the single- or two-cell stage resulted in an exact phenocopy of the nrs mutant embryos which manifested obvious yolk opacity. Upon SA-β-gal staining, an extremely high level of induction was observed in the nrs morphants (n = 116 of 120; 97%), identical to the SA-β-gal levels in the nrs mutants (Figure S2A). In contrast, the control embryos did not show any significant SA-β-gal activity (n = 150). We also injected a terf2 MO into zebrafish embryos, and observed robust SA-β-gal induction (n = 191 of 200; 95.5%), and a relatively moderate morphological phenotype similar to the terfahi1182/hi1182 allele, that has a weaker phenotype than the insertional mutant (terfahi3678/hi3678) used above (Figure S2B). This is consistent with the higher residual levels of terf2 mRNA in the morphants and in the hi1182 mutants harboring an insertion in the first intron of the terf2 gene which allows for the production of some wild type transcript, in contrast to the hi3678 allele that has the insertion in the first exon of the terf2 gene (http://web.mit.edu/hopkins/insertion20sites/1182.htm).
We additionally exposed nrs and terf2 mutant embryos to oxidative stress by BHP treatment. The homozygous terf2m/m (terfahi3678/hi3678) mutants (n = 50), but not the heterozygotes, clearly show enhanced induction of SA-β-gal activity with a more severe morphology in the eyes and heads (Figure 4F), whereas no significant difference was observed in the nrs mutant animals of either homozygous or heterozygous backgrounds (data not shown). Taken together, the outcomes from our current screen of 306 lines from the Hopkins' insertional mutant collection serve as a proof of concept for our strategy and the first success in our novel approach to identify potential aging-related genes by examining the senescence-associated biomarker in zebrafish embryos.
Having established that SA-β-gal is induced by oxidative stress caused by BHP treatment, we next performed a new screen for ENU mutant zebrafish which displayed phenotypic alterations arising from genetic mutations in stress response mechanisms. We crossed individual F1 mutant males with wild-type females to produce clutches of F2 embryos, each of which was heterozygous at many loci. By using BHP as a chemical sensitizer, we hoped to identify heterozygous mutants with an altered response to oxidative stress; that is we expected the chemical sensitizer to induce haploinsufficiency in many of the potential target genes. We have denoted this methodology ‘CASH’ (Chemically Assisted Screening in Heterozygotes).
We treated 50 embryos from the resulting clutches with 350 µM BHP from 6 hours post fertilization (hpf) to 6 dpf. In each clutch of embryos, half of the clutch would be heterozygous for any mutant allele, and half would be wild type for that allele. Thus, an F1 male carrying a mutation that alters sensitivity to oxidative stress would produce clutches wherein half the embryos show altered induction of SA-β-gal activity (Figure 5A). We divided SA-β-gal staining intensity in the F2 embryos into discrete quantitative ranges and measured how many embryos fell into each staining intensity range. When we performed this analysis on wild-type embryos, the result was a tight Gaussian distribution (Figure 5B). When we looked at our candidate mutant clutches, in most cases their staining distributions appeared similar to wild type. However, occasionally, nearly half of the embryos were darker than the others and the distributions appeared abnormal (see a dotted red line throughout Figure 5B, 5D, and 5F). The F1 fathers of these clutches were potential carriers of mutations that either enhanced SA-β-gal activity. These are what we will refer to as ‘Class 1’ mutant candidates (Figure 5C and 5D; n = 35 for this tested candidate; P<0.01, Student t-test). We also observed clutches of another class of mutants in which approximately half of the embryos showed a clear morphological abnormality in the presence of BHP (Figure 5E and 5F; n = 45 for this tested candidate; P<0.001, Student t-test). However, when we repeated these outcrosses without exogenous oxidative stress, the resultant clutches appeared morphologically normal. These clutches comprise what we will refer to as ‘Class 2’ mutant candidates. We performed an initial screen in 150 F1 mutagenized genomes, and obtained 8 candidate mutants that bred in a consistent recessive Mendelian fashion through to the F4 generation. Six of the mutants were from Class 1 and two were from Class 2.
When we incrossed F2 siblings (n>30) of each peroxide-sensitive mutant (psm) line, we observed the homozygous phenotypes seen in Figure 6 in about 25% of the embryos of each clutch from about 1 out of every 4 incrosses. This is consistent with the expected Mendelian segregation of a recessive trait. For each psm mutant, over 2 generations and at least 20 fish per generation the fish that transmitted the heterozygous psm phenotype with high SA-β-gal activity also transmitted the recessive morphological phenotype. Seven of the phenotypes were very similar (psm2, 5, 6, 8, 9, 10, 11), with evidence of a moderate dorsal curvature of the trunk, and showed pronounced levels of cell death which was clearly apparent in the brain beginning at 48 hpf (n = 50 for each mutant) (five of these were Class 1 mutants and two were Class 2 mutants). These homozygous mutations were embryonic lethal, with death occurring at around 6 dpf. The remaining psm7 mutant (Class 1) developed a minor protrusion of the jaw and had an opaque yolk that first became apparent at 4 dpf (n = 50), subsequently dying at around 7 dpf.
We proceeded to examine the homozygous phenotype of each psm mutation more closely. All F3 homozygous embryos (n = 50) with abnormal phenotypes were also found to have higher SA-β-gal staining than the wild-type controls (n = 20) in the absence of any exogenous oxidative stress (Figure S5). Punctate SA-β-gal staining was seen throughout the central nervous system in each of the seven mutants we identified (psm2, 5, 6, 8, 9, 10, 11), which showed brain abnormalities and dorsal curvatures (n = 25 for each mutant) (Figure 7B, 7E and 7F; psm6 mutant is shown). Acridine orange (AO), which stains dying cells, produced very intense signals throughout the neural tube and brain tissues of these mutants between 36 and 72 hpf (n = 50 for each mutant at each time point), indicating massive cell death (Figure 7M and 7O; psm6). In vivo staining of the neurodegenerative mutants at 3.5 dpf with dichlorofluorescein diacetate (DCFH-DA), an indicator of ROS, revealed the presence of high levels of ROS in the neural tube, specifically in the dorsal half (n = 20 for each mutant) (Figure 7S; psm6). Interestingly, these phenomena were also true of homozygous terf2m/m mutant embryos that showed high levels of ROS in the neural tube at 3.5 dpf (Figure 7Y), and demonstrated positive AO-staining indicating cell death also at 2 dpf (Figure 7X). Histological analysis of these embryos at 2 and 3 dpf indicated evident abnormalities around the regions of the brain, neural tube and eyes where the accumulation of neuronal cell death products (data not shown). At 4 and 5 dpf, further histological examinations revealed that the brain and neural tubes of the psm mutants were considerably smaller than those of wild types and contained fewer neuronal nuclei (Figures 7P; wild type and 7Q; psm6).
The mutant showing yolk opacity (psm7) also showed mottled SA-β-gal staining throughout the muscle in the trunk (n = 71) (Figure 7C and 7G). Histological sections of mutant animals at 4 dpf revealed the absence of muscle fibers throughout the myotomes (n = 35) (Figure 7I), suggesting muscular degeneration. However, this phenotype does not appear to be the result of defective muscle fiber attachment as reported in another dystrophy-like mutant (Bassett et al., 2003), since whole-mount in situ histochemical analysis of dystrophin expression appeared to be normal (n = 20) (Figure S6), indicating that the fiber loss is not likely related to dystrophin-mediated fiber adhesion. In addition, DCFH-DA staining of the psm7 homozygotes at 3.5 dpf showed high ROS production in many individual muscle fibers (n = 45) (Figure 7K).
It is noteworthy that all of the seven neurodegenerative mutants (psm2, 5, 6, 8, 9, 10, 11) were in the same complementation group while the mutant showing muscle degeneration (psm7) was not linked to any of the neural phenotype mutants (Table S1). Moreover, each of the psm mutants complemented both the nrs and terf2 mutations (Table S1). Thus it is possible that this pilot ENU screen has recovered mutations in only 2 genes, each of which are distinct from the nrs and terf2 genes.
We wondered whether there might be long-term (‘aging’) effects of heterozygosity for the genes identified in our screens stemming from the associated embryonic alterations in senescence markers/phenotypes. We thus measured the premature aging marker levels and pathohistological phenotypes in heterozygous fish as they aged. Two of the 5 tested heterozygous psm mutant lines (psm6, n = 8 and psm9, n = 8) showed significantly higher levels of SA-β-gal activity in the skin at just 1.5 years of age (18 months) compared with their wild-type siblings (n = 10 for each group) (Figure 8A). In contrast, a heterozygous nrsm/+ mutant (n = 12) showed only a modest increase in skin SA-β-gal activity at 2 years but showed a high induction of SA-β-gal at 3.3 years of age (40 months), compared with wild-type siblings (n = 15) (Figure 8A).
In addition to SA-β-gal activity, lipofuscin is often considered to be a hallmark of aging, showing an accumulation rate that correlates with longevity in some tissues [31],[49]. We reported previously that wild-type adult zebrafish are refractory to the accumulation of lipofuscin in muscle by 2 years of age [19]. In our current study we found that there is still no detectable lipofuscin accumulation in wild-type sibling fish (n = 15) by 2.8 years (33 months) of age, but that heterozygous nrsm/+ mutant fish (n = 10) had accumulated a great deal of lipofuscin in the skeletal muscle by this time (Figure 8H). On the other hand, in terms of the liver histology of the wild-type zebrafish, lipofuscin accrual changed dramatically with age (Figure S7A), similar to that reported in mice [50]. Notably also, we found in our current analyses that male psm6 (n = 10) and psm7 (n = 10) heterozygous animals, as well as the male nrsm/+ (n = 8) heterozygotes, accumulated lipofuscin in the liver at an early age compared with their wild-type siblings (n = 10 for each group) (Figure 8B–G). Although heterozygous nrsm/+ and psm7m/+ mutant fish did not display significantly increased SA-β-gal levels at 25 months (2.1 years) and 18 months (1.5 years) of age, respectively (Figure 8A), both of these mutants did show increased lipofuscin accumulation at these ages (nrs, n = 9 and psm7, n = 8) (Figure 8E, G). This indicated that different aging phenotypes may occur independently or that liver lipofuscin buildup may be an earlier or more sensitive indicator of aging in these mutants. Taken together, these results suggest that our newly identified mutants show a robust early-onset expression of aging biomarkers that normally only manifest in much older wild-type animals.
A striking phenotype associated with the terf2m/m homozygous mutant embryos was observed in the central nervous system, including the eye (retina) and brain, as shown in Figure 4C–E and Figure 7W–Y. Given the degenerative phenotype seen in embryonic terf2m/m retinas (Figure 4C–E, right panels), we examined histological sections of heterozygous adult mutants. Histological sections of mutant (n = 23) and wild-type sibling (n = 15) retinas were analyzed at various ages to determine the cellular basis for the observed age-dependent retinal defects. We observed structural abnormalities, principally retinal cell degeneration, in most aged mutant retinas with drusen-like autofluorescent accumulations around the retinal pigment epithelium (RPE). This degeneration was not uniform over the retinas but tended to be patchy. Areas of rods (rod outer segments) degeneration were interspersed between areas where significant numbers of rods remained. Cones (cone outer segments) generally were better preserved than the rods, but areas of cone degeneration were also noted. The observed degeneration was progressive with age. In animals older than 12 months, degeneration was usually seen only in the central-most regions of the retina. By 21 months of age, however, degeneration was observed across much of the retina, although the far peripheral regions tended to be spared. Importantly, in the wild-type zebrafish retinas, these degenerative changes appeared dramatically with advancing age (Figure S7B).
Representative histological sections from a 23-month old wild-type sibling fish and an age-matched heterozygous terf2m/+ mutant fish were compared and are shown in Figure 8I. In zebrafish, the photoreceptors are typically tiered so that in the light-adapted retina, the rods are positioned more distally than the cones. In the 23-month old terf2m/+ retina, the peripheral regions were found to be similar to control retinas, but obvious abnormalities were observed throughout the central retina. In some areas only the rods were affected, i.e., the rod outer segments were disorganized and reduced in length, and autofluorescent accruals in the RPE were increased in number and size (Figure 8I, right panel in terf2m/+). In other areas, the rods were severely degraded but the cones appeared relatively normal. In yet other areas, both the cones and rods were affected, i.e., both were sometimes disorganized and reduced in length (data not shown). Moreover, the inner plexiform layers of the retina usually looked thinner in affected animals, and occasionally some patchy thinning of the inner nuclear layer and empty spaces were also observed (Figure 8I, left panel in terf2m/+), suggesting some loss of inner retinal neurons. To investigate whether there might be loss of elements other than photoreceptors in the mutant zebrafish retinas, we measured the thickness of various retinal layers centrally in control (n = 6) and mutant (n = 8) animals at 23 months of age (Table S2), and also in control (n = 6) and mutant (n = 6) animals at 30 months of age (data not shown). The retinas of mutant animals were clearly thinner than those of the wild-type fish (Figure 8I). Much of this change in thickness was accounted for by a decrease in the thickness of the photoreceptor layer and inner plexiform layer.
Finally, we obtained Kaplan-Meier survival curves for some of our mutant fish. The oldest fish that were heterozygous for psm mutations in our stocks at the time of writing had not reached their maximum lifespan (wild-type zebrafish have a maximum lifespan of roughly 5 years [51]), so we have not yet been able to determine the effects of these mutations upon overall lifespan. We have, however, maintained populations of heterozygous nrs and terf2 mutant fish and their wild-type siblings until death, and conducted observational studies on their lifespan. We scored the cohorts of nrsm/+ fish (no genders identified; n = 148) in comparison with their wild-type siblings (no genders identified; n = 256), and found significant decreases in the lifespan of the heterozygous mutants (P<0.0001, log rank test) (Figure 9A). Moreover, the male terf2m/+ fish cohorts (n = 96) also manifested a shorter lifespan compared with that of their wild-type male siblings (n = 79) (P<0.0001, log rank test) (Figure 9B). In contrast, neither the heterozygous terf2 mutant females nor their wild-type female siblings had reached the end point of their lifespan during the period of our current experiment. These lifespan analyses of the two mutations identified by embryonic senescence phenotypes suggest that screening for the embryonic appearance of ‘aging biomarkers’ may in some cases at least, predict a role for specific genes in the organismal aging process.
The goal of our current study was to test the hypothesis that mutations which enhance the appearance of embryonic stress markers might result in degenerative or even aging phenotypes in adults. We have presented the results of two different screens for potential aging mutants in this report, both utilizing SA-β-gal production in the zebrafish embryo as a key part of the screening process. We first screened a collection of retrovirus-mediated insertional mutants for elevated SA-β-gal production in unstressed homozygous embryos. Notably, this collection of insertional mutants was isolated based on the requirement for homozygous developmental phenotypes. Each of the two mutants out of the 11 candidates from this first screen included a lesion in the nrs gene, the zebrafish homologue of the Drosophila spinster gene a known regulator of aging in flies [33], or the terf2 gene, which plays roles in telomere protection and telomere-length regulation as a component of the telosome/shelterin complex [35]. In the second screen, using an oxidant agent as a sensitizer, we have isolated a series of mutants, denoted psm mutants, which showed elevated SA-β-gal expression in the stressed heterozygous state and additionally in an unstressed homozygous state of embryos. Sensitized screens are much less labor and time intensive than traditional screens for recessive phenotypes in homozygous mutant embryos. Importantly, heterozygous mutant fish from both screens showed elevated expression of aging biomarkers in relatively younger ages. Thus, mutants from both screens showed degenerative phenotypes in homozygous embryos during early development and in heterozygous adults with age. Notably, the heterozygous animals in some cases also appeared to die at earlier ages than controls.
It is possible that the two screens revealed some different classes of mutants. For instance neither of the two insertional mutants which we studied in detail showed statistically significant alterations in their sensitivity to oxidative stress as heterozygotes (data not shown). The heterogeneity of the observed phenotypes in mutants from the 2 screens persisted in the adult heterozygotes. Both nrs and psm mutations showed hastened lipofuscin accrual in the livers of young adult heterozygotes, whereas two psm mutants (psm6 and psm9) showed enhanced SA-β-gal production in relatively young adults (1.5 y) compared with their wild-type siblings, a trait shown only in older nrsm/+ fish (3.3 y). However, since we examined mutants for only two genes in each screen in any detail, it is premature to derive substantial conclusions at this stage.
The accelerated production of lipofuscin in the liver is a phenotype that was found to be common to the adult zebrafish from both screens. Lipofuscin accumulation is a hallmark of aging in many organisms from worms to mammals [32],[52]. Tissues that are traditionally thought to be sensitive to lipofuscin accrual in mammals include the brain, and both skeletal and cardiac muscle. However, previous studies from our laboratory have shown that there is no clear accumulation of lipofuscin inclusions with age in wild-type zebrafish skeletal muscle cells and cardiac myocytes, at least up to 2.5 years of age [13],[19] (unpublished observations). In contrast, age-related changes in liver structure and lipofuscin accumulation have been demonstrated in male mice [50]. We have also observed premature lipofuscin accrual in the adult liver of our heterozygous male psm mutants as well as in male nrsm/+ heterozygotes. It is widely believed that oxidative damage processes underlie sources of lipofuscin production, and that its accumulation may have multiple negative effects, including a further increase in sensitivity to many stress-induced damage responses [32]. Since we observed no obvious response to exogenous (extrinsic) oxidative stress by BHP in nrs mutant embryos, we suspect that differences in intrinsic energy metabolism may underlie lipofuscin appearance in the nrs mutants, although this clearly requires further investigations. In the case of flies, it has also been shown that mutations in the spinster gene cause enhanced lipofuscin production [33], though the precise underlying mechanism also remains unknown. Therefore, parallel studies of this gene product in other organismal aging model systems would be very desirable to provide a more conclusive insight into its mechanistic roles.
There is cumulative evidence to date to suggest that early onset neuronal degeneration phenotypes in homozygous zebrafish mutants are predictive of a late-onset visual impairment in the corresponding heterozygous animals [53]–[56]. Intriguingly, in heterozygous terf2m/+ mutant adult fish, drusen-like autofluorescent accumulation (presumably caused by lipofuscin accrual) is more obvious in the RPE compared with age-matched siblings. In contrast to other tissues, ocular lipofuscin has been identified as N-retinylidene-N-retinylethanolamine (A2E). A2E is a quaternary amine and retinoid by-product of the visual cycle and causes the accumulation of free and esterified cholesterol in RPE cells [57],[58]. Although endogenously produced A2E in the RPE has been associated with macular degeneration, the precise mechanisms are unclear. Therefore, the involvement of the telomeric factor TRF2 in the mechanism of the RPE lipofuscin accrual might provide intriguing new insights into potential novel strategies for the prevention and treatment of neurodegenerative disorders. Alternatively, telomere-associated proteins might be involved in neural differentiation, as homozygous terf2m/m mutant embryos show an aggressive neurodegenerative phenotype in both the eye (retina) and brain. In this regard, divergent molecular and physiological responses to telomere dysfunction in mitotic neural stem/precursor cells and postmitotic neurons appear to regulate the differentiation and survival of neurons as well as RPE cells [47],[48].
Notably, both terf2 (males) and nrs (gender not identified) heterozygous mutant had shorter lifespans in comparison with their control wild-type siblings, suggesting that partial loss-of-function/decrease-of-function in these genes may have systemic effects on physiological aging rather than the organ-specific tissue aging on which we focused in the current study. In addition, there may be gender-dependent differences in longevities and mortality rates in zebrafish when we look at the survival curves of wild-type siblings of terf2 and nrs mutants. Male terf2 heterozygous mutant fish and their wild-type male siblings were kept isolated from females except for the occasional matings performed in our current study. Heterozygous terf2 mutant females and their wild-type female siblings had not reached the end point of their lifespan at the time when the male fish all died out. On the other hand, heterozygous nrs mutant males and females basically co-habited throughout their life, as did their wild-type siblings. When they died, sex determination was therefore not possible due to rapid decomposition (autolysis) in water. Although further studies are definitely needed, social and environmental factors between genders might affect zebrafish lifespan in addition to their genetic background.
Any understanding of the mechanisms by which the psm genes regulates the functions of age accompanied with degenerative phenotypes will require positional cloning of the mutated genes. The psm mutants may also be useful for studying signal transduction pathways and related biological processes in addition to both stress responses and aging. The fact that these mutations regulate embryonic tissue-specific responses to oxidative stress is also interesting per se. It will be of considerable interest to see if these mutations lie in genes already associated with oxidant responses, such as antioxidant genes or genes that maintain mitochondrial functions. It is also possible that the psm mutants which showed neuronal phenotypes may have value as new models for specific disorders such as Huntington's disease, Parkinson's disease, Alzheimer's disease, amyotrophic lateral sclerosis, and ataxia telangiectasia, because several lines of evidence suggest that oxidative stress is associated with the development of these neurodegenerative diseases [59]–[63]. In addition, oxidative stress has been shown to play a role in sarcopenia and other muscle wasting conditions [5],[6], where the psm7 mutant may be involved. Since we have already shown an embryonic haplosensitivity of the psm mutants to oxidative stress by the very nature of the CASH strategy, we are currently further investigating the possibility that more of these mutants may show accelerated onset of aging in the adult heterozygotes, expanding their utility as models of stress-associated pathophysiological aging and degenerative disease.
In the future, we hope to further utilize our technology to identify a real “suppressor” type mutant which would display enhanced embryonic stress resistance and, perhaps, a longer healthy lifespan (‘health span’). Alternatively, it might also be possible to isolate a ‘revertant’ from the background of an accelerated aging mutant, which would restore the original normal phenotype by means of a suppressor mutation.
In summary, our current study has demonstrated for the first time in vertebrates that it is possible to obtain mutations that alter adult aging markers/phenotypes and lifespans by screening mutagenized and sensitized embryos for the extemporal expression of the aging biomarker. It is our hope that this novel tactic of screening for aging biomarkers in zebrafish embryos may open a new avenue for the future genetic dissection of vertebrate aging mechanisms.
Zebrafish of the wild-type strain and mutants were maintained with a 14:10 h light/dark cycle and fed living brine shrimp twice per day. Brine shrimp was given using 1 mL pipettes at an amount of about 0.75 mL per 20 fish. Flake food was also given a few days per week semiquantitatively according to the number of fish in the tanks. A continuously cycling Aquatic Habitats™ system was used to maintain water quality (Apopka, FL, USA) which completely replaces the water in each tank every 6–10 min. Each tank is a baffle/tank system that ebbs water in a circular motion to ensure flushing and water turnover. Ultraviolet (UV) sterilizers (110,000 microwatt-s cm−2) were employed to disinfect the water and prevent the spread of disease in the recirculating system. The water temperature was maintained at 28±0.5°C. The system continuously circulated water from the tanks through Siporax™ strainers, through a fiber mechanical filtration system, and finally into a chamber containing foam filters and activated carbon inserts. Water quality was tested daily for chlorine, ammonia, pH, nitrate, and conductivity under real-time computer monitoring with alarms to signal potential fluctuations. The general health of each fish was observed on a daily basis throughout fish life to monitor longevity, and abnormal looking or acting fish were quarantined into isolated tanks unconnected from the general circulation. The water in these quarantined tanks was treated with methylene blue (0.0001–0.001%). If and when fish recovered, they were returned to their original tanks on the general circulation. All animals showing signs of infectious or parasitic disease that are not alleviated by a 7-day incubation with methylene blue in water were euthanized in a beaker containing tricaine (approximately final 0.5% in water). Animals exhibiting overt tumors or extreme morbidity were also euthanized.
Embryos were collected by natural spawning, raised in 10% Hank's saline with or without 0.003% 1-phenyl-2-thiourea (PTU) (embryo media) and staged according to Kimmel et al. [64]. All embryos were incubated at 28.5°C during development.
Data processing and statistical analyses were performed using Microsoft Excel and Statistical Package for the Social Sciences (SPSS) version 14.0, which were used to generate each of the scatter plots, tables, and graphs shown in the text, performing statistical tests where appropriate. Additional statistical analyses were performed at the Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute. These analyses included survival estimates using the method of Kaplan and Meier, and comparison of survival between mutant fish and their wild type controls using the log rank test.
Zebrafish adults and embryos were fixed in 4% paraformaldehyde in phosphate buffered saline (PBS) at 4°C (for 3 days in adults and overnight in embryos), and then washed 3 times for 1 h in PBS-pH 7.4 and for a further 1 h in PBS-pH 6.0 at 4°C. Staining was performed overnight at 37°C in 5 mM potassium ferrocyanide, 5 mM potassium ferricyanide, 2 mM MgCl2, and 1 mg/ml X-gal in PBS adjusted to pH 6.0. All animals were photographed under the same conditions using reflected light under a dissecting microscope. SA-β-gal activity in each animal was quantitated using a selection tool in Adobe Photoshop for a color range that was chosen by 25 additive blue color selections of regions that showed visually positive SA-β-gal staining. For analyses of embryos, these regions were selected in each embryo proper only and not in the yolk in order to eliminate variability due to differences in initial yolk volume and yolk consumption over time. Since the yolk stains much more intense blue for SA-β-gal at all stages of development than any other embryonic tissues, even under conditions of high oxidative stress, it was desirable to eliminate this as a source of variability. Following pixel selection, a fuzziness setting of 14 was used, and the chosen pixel number was calculated using the image histogram calculation. For adult zebrafish analyses, the trunk area for colorimetric quantitation was chosen by selection of the area between the operculum and the dorsal and anal fins (Figure S1).
The cloning of zebrafish catalase cDNA was performed using the SuperScript One-Step RT-PCR kit with Platinum Taq (Invitrogen) according to the manufacturer's instructions, by using the forward primer 5′-TTTGCCTCGTGTTTTGTCAC-3′ and reverse primer; 5′-GGAGTCAGTGTTGCATTTGCT-3′. These primers were designed using the flanking regions of the zebrafish Ensembl sequence for the predicted human catalase homolog. The full-length cDNA was then cloned into the pCS2+ vector. The resulting plasmid was linearized by digestion at a restriction site immediately after the poly-A signal, and capped mRNAs were transcribed in vitro using the mMachine Kit (Ambion Inc.) following the manufacturer's instructions. The stated amount (300 pg) of mRNA were injected into one-cell stage zebrafish embryos using a gas driven microinjector (Medical Systems Corp.).
Knockdown of zebrafish catalase was performed by injection of 8 ng antisense morpholino oligonucleotide (MO) (Gene-Tools, LLC) with the sequence 5′-TCGACTTTTCTCTGTCGTCTGCCAT-3′ or a control morpholino 5′-CCTCTTACCTCAGTTACAATTTATA-3′. Knockdown of zebrafish nrs/spns1 and terf2/terfa was performed also by injections of MOs (8 ng) containing the sequences 5′-ATCTGCTTGTGACATCACTGCTGGA-3′ and 5′-GGTTCGCAGGGTTTGTCGCTCATTC-3′, respectively.
Heterozygous incrosses of each mutant line were performed, and the resulting embryos were raised in 10% Hank's saline at 28.5°C. The embryos were fixed at either 3.5 dpf or at least 18 h before the occurrence of embryonic lethality after 3.5 dpf for lines whose the homozygotes are known to die. SA-β-gal staining was performed as above, and positive candidates were determined by correlating high SA-β-gal activity with the presence of a homozygous mutant phenotype. Since some of mutants needed to be distinguished by their pigmentation patterns, the entire screening of 306 mutants was performed in these embryos without PTU treatment. Embryos of selected 11 candidates were also processed for single-embryo SA-β-gal quantitation with PTU treatment as described above.
ENU mutagenesis was performed as previously described [65]. Briefly, regularly bred 1-year old *AB strain adult zebrafish males were treated with 3 mM N-ethyl-N-nitrosourea at 20°C for 1 h once a week for 3 times. From five to seven weeks after the final treatment, mutagenized males were outcrossed to wild-type *AB females and F1 progeny were raised to adulthood.
F1 mutagenized males were outcrossed with three wild-type *AB females to maximize the resulting clutch size. 50 embryos from the resultant clutches were then raised from 6 hpf to 6 dpf in 10% Hank's Saline with 0.003% 1-phenyl-2-thiourea (PTU) and 350 µM tert-butyl hydroperoxide (BHP). The media was refreshed every 48 h. Embryos were processed for single-embryo SA-β-gal quantitation as described above. Clutches showing either obvious phenotypic abnormalities in 50% of the embryos, or having staining standard deviations of at least 1.5 times that of wild-type clutches in both of two independent breedings, were considered to be a positive hit. Positive hit F1 males were subsequently outcrossed with wild-type *AB females and the resulting F2 generation was raised to adulthood. F3 embryos derived from these F2 sibling incrosses in each hit family were assessed for their phenotype and SA-β-gal activity levels both in the presence and absence of oxidative stress.
For the in vivo detection of cell death, live 2-day old embryos were incubated in 2 µg/ml acridine orange (AO) (Sigma) in embryo media in the dark for 30 min and washed three times for 5 min in fresh embryo media. Fluorescence was then observed under a 488 nm wavelength excitation. For in vivo ROS detection, live 2–4-day old embryos (2–4 dpf) were incubated in 5 µM 2′,7′-dichlorofluorescein diacetate (DCFH-DA) (Sigma) for 20 min at 28.5°C and washed three times for 5 min with embryo media. Fluorescence was again observed under a 488 nm wavelength excitation. For Fluoro-Jade B histochemical analysis of 2 dpf embryos, adjacent sections were stained using the standard Fluoro-Jade staining procedure as described previously [66].
Embryos and adult tissue samples were fixed in 4% paraformaldehyde in PBS for 48 h at 4°C. Samples were dehydrated in ethanol and infiltrated in JB-4 resin following the manufacturer's instructions (Polysciences Inc.). Specimens were then sectioned at 5 µm using a Jung Supercut 2065 microtome. Histological hematoxylin-eosin (H&E) staining of the sections was subsequently carried out using standard protocols. After digestion with diastase, periodic acid Schiff's (PAS) staining was performed as follows: sections were heat-adhered to slides at 70°C for 5 min and placed in the following solutions at room temperature; 1% periodic acid for 5 min, several water changes over 5 min, Schiff's reagent for 30 min, 0.5% sodium metabisulfite in 1% concentrated HCl 3×2 min, and several water changes for 10 min each.
|
10.1371/journal.ppat.1003669 | Cross-Serotype Immunity Induced by Immunization with a Conserved Rhinovirus Capsid Protein | Human rhinovirus (RV) infections are the principle cause of common colds and precipitate asthma and COPD exacerbations. There is currently no RV vaccine, largely due to the existence of ∼150 strains. We aimed to define highly conserved areas of the RV proteome and test their usefulness as candidate antigens for a broadly cross-reactive vaccine, using a mouse infection model. Regions of the VP0 (VP4+VP2) capsid protein were identified as having high homology across RVs. Immunization with a recombinant VP0 combined with a Th1 promoting adjuvant induced systemic, antigen specific, cross-serotype, cellular and humoral immune responses. Similar cross-reactive responses were observed in the lungs of immunized mice after infection with heterologous RV strains. Immunization enhanced the generation of heterosubtypic neutralizing antibodies and lung memory T cells, and caused more rapid virus clearance. Conserved domains of the RV capsid therefore induce cross-reactive immune responses and represent candidates for a subunit RV vaccine.
| Human rhinovirus infections cause the majority of common colds as well as asthma and chronic obstructive pulmonary disease (COPD) exacerbations. The disease burden attributable to rhinoviruses is therefore huge. Despite this and the fact that human rhinoviruses were discovered over 50 years ago, there are currently no specific antiviral therapies or vaccine available. The lack of a rhinovirus vaccine can at least in part be attributed to the fact that rhinoviruses like other pathogens have high variability in surface antibody binding regions, resulting in >100 serotypically distinct strains. We have defined areas of the rhinovirus polyprotein which are highly conserved across strains and which may therefore induce cross-reactive immune responses capable of providing broader protection. Using a mouse model, we show that immunization with a recombinant rhinovirus capsid protein induces cross-reactive cellular and humoral immune responses. After subsequent infection, immunization enhances both neutralising antibody and lung effector and memory T cell responses, expediting virus clearance. Importantly these effects were evident upon challenge with multiple heterologous rhinovirus serotypes, indicating that immunization with conserved rhinovirus capsid proteins may represent a viable strategy for producing a broadly cross-reactive vaccine.
| Human rhinovirus (RV) infections are the most frequent cause of the common cold [1] and are highly associated with exacerbations of asthma and COPD [2], [3], [4]. Despite the great disease burden and healthcare costs therefore attributable to RV infections, there is currently neither a vaccine nor specific anti-viral therapy available.
The requirements for immunity to RV are poorly understood. Experimental and natural infections induce antibodies which provide some protection against re-infection with the same RV serotype [5], [6], [7]. Intranasal and intramuscular inactivated virus vaccinations similarly induce neutralizing antibodies and provide protection against disease induced with the same RV serotype [8], [9]. There are however greater than 100 serotypes of RV [10], divided into major and minor groups based on receptor usage and A and B groups based on antiviral sensitivity and nucleotide sequence [11], [12], and a further ∼50–60 RV species more recently defined as group C RVs based on sequence data alone [13], [14]. Serological variability amongst RVs therefore means that vaccines designed to generate neutralizing antibodies are unlikely to provide sufficiently broad protection to prevent the frequent infections which occur throughout life.
Alternative vaccination strategies based on inducing T cell responses to conserved antigens have been explored for a number of pathogens, including respiratory viruses [15], [16]. An advantage of this approach lies in the ability of T cells to recognize internal virus proteins which are typically more highly conserved than surface exposed regions containing neutralizing antibody epitopes. T cells are therefore potentially cross-reactive against different virus strains, as has been shown with influenza viruses [17], [18], for which surface antigenic variability is also an obstacle to effective vaccine design.
For RVs, naturally occurring memory T cells can be cross-serotype responsive [19], [20] and immunization with RV peptides has been suggested to be capable of inducing cross-serotype reactive T cells in mice [21]. Most of the naturally occurring RV-specific memory T cells characterized to date have shown a Th1/Tc1 bias [19], [20]. In vitro responses to RV by mixed PBMCs have been associated with virus shedding or cold symptoms after subsequent infection [22] but there is no evidence that naturally occurring RV-specific memory T cells specifically provide benefit in terms of virus control or disease symptoms in vivo. Here we show that a vaccine composition which elicits a Th1/Tc1 biased T cell response to conserved RV antigens could have efficacy.
We took a bioinformatic approach to identify regions of the RV polyprotein which are conserved across A and B group and major and minor receptor binding group viruses, and which might be used as immunogens in a cross-reactive vaccine. As in similar analyses by others [11], we show that areas of the capsid VP0 protein are highly conserved amongst RVs. Immunization with VP0 protein from major group RV16 combined with Th1 promoting adjuvants induced antigen-specific, type I orientated T cell responses in the airways, enhanced neutralizing antibody responses to infection and caused a more rapid decrease in lung virus load in mice. Importantly, these effects were seen in mice infected with heterologous RV strains, indicating that capsid protein immunization could provide broadly cross-reactive immunity against RVs.
Using published amino acid sequences we defined areas of the RV polyprotein which are conserved across A and B species RVs. The methodology for determining amino acid sequence conservation amongst RVs is described in materials and methods. We did not find well conserved sequences covering both A and B species RVs, but within each species three regions were identified as highly conserved in agreement with similar sequence comparisons carried out previously [11] and therefore represented candidate antigens. These were amino acids 1–191 and 243–297 in the N-terminus of the polyprotein, and the C-terminal domain of the RNA polymerase (Fig. S1a). The two N-terminus regions lie within the VP4 and VP2 capsid proteins, of which VP0 is the natural precursor. Because VP0 contains both very highly conserved internal (VP4) and surface exposed regions with neutralizing epitopes (VP2), VP0 was chosen as the antigen for further studies. Sequences from RV16, a major group A species RV were used to allow study of cross-reactivity to minor group RV strains which can infect wild type mice [23]. Figure S1 shows detailed analysis of the high sequence conservation within VP0 (Fig. S1b), the amino acid sequence of the RV16 VP0 immunogen (Fig. S1c) and comparison of RV16 VP0 with VP0 sequences of minor group RVs 1B, 29 and 14 used subsequently (Fig. S1d).
We first assessed the immunogenicity of subcutaneously delivered RV16 VP0 protein. Analysis of antibody responses by western blot showed that RV16 VP0 - specific IgG was detectable in serum 28 days post-immunization (Fig. 1a). In mice immunized with VP0 protein alone, VP0-specific IgG1 and IgG2c, Th2 and Th1 associated IgG isotypes respectively, were detected.
To assess whether a Th1/Tc1 orientated response to infection is associated with improved disease outcome, we attempted to induce a Th1 skewed response to RV16 VP0 using a combination of incomplete freund's (IFA) and CpG adjuvants (IFA/CpG). The addition of IFA/CpG to the immunogen caused a more prominent IgG2c response (Fig. 1a).
Having established that RV16 VP0 is immunogenic, we next assessed the T cell response to immunization by measuring splenocyte cytokine production in response to stimulation with VP0, or control polymerase, peptides (described in Fig. S2). Stimulation with control polymerase peptides did not induce cytokine production (Fig. 1b,c). In both ELISPOT (Fig. 1b) and cytometric bead array (Fig. 1c) assays VP0 peptide pool stimulation induced IL-5, or both IL-5 and IFN-γ production by cells from mice immunized with VP0 protein alone, indicating a Th2 or mixed Th1/Th2 orientated response. As expected, the addition of IFA/CpG adjuvant to the immunogen caused a near complete suppression of IL-5 and substantial increase in IFN-γ responses (IL-5 p<0.01, IFN-γ p<0.001 RV16 VP0+IFA/CpG vs RV16 VP0 treatment for VP0 peptide pool stimulation) (Fig. 1b,c). Importantly, splenocytes from major group A species RV16 VP0 protein immunized mice produced cytokines when stimulated with VP0 peptides based on minor group A species RV1B and major group B species RV14 sequences, indicating cross-serotype reactivity.
We next determined the effect of (major group, A) RV16 VP0 plus IFA/CpG immunization on responses to intranasal challenge with RV1B, a heterologous minor group A virus (Fig. 2a).
We observed no signs of clinical disease in animals which were immunized prior to infection consistent with our previous experience of mouse RV infections. Differential staining of bronchoalveolar lavage (BAL) leukocytes showed a significantly increased magnitude of lymphocyte response to infection in immunized and infected (RV-immunized) vs adjuvant treated and infected (RV-Adjuvant) mice (day 6 post-infection p<0.001) (Fig. 2b). To examine this enhanced lymphocyte response further, T cells in BAL and lung were analyzed by flow cytometry. CD4+ T cell numbers were substantially increased in both BAL and lung, and CD8+ T cell number was increased in BAL of RV-immunized vs RV-adjuvant treated mice on day 6 post-infection (p<0.01 BAL and lung CD4+ T cells, p<0.001 BAL CD8+ T cells) (Fig. 2c). The response in RV-immunized mice was dominated by CD4+ T cells whose number was ∼10-fold greater than CD8+ T cells by day 6 post-infection. In infected mice, the proportion of BAL and lung T cells expressing the activation marker CD69 was also significantly increased by immunization (RV-Immunised vs RV-adjuvant p<0.001 lung CD4+ and CD8+ T cells day 1–14, p<0.05 BAL CD4+ & CD8+ T cells d6 & d14)(Fig. 2d). Immunization-induced increases in T cell number were associated with enhanced levels of T cell chemokine CXCL10 (p<0.001 RV-Immunised vs RV-adjuvant at 24 hrs post infection)(Fig. 2e).
We also examined the effect of immunization with RV16 VP0 on the polarity and antigen specificity of airway T cells after heterologous RV1B challenge. Immunization significantly increased the levels of signature Th1 (IFN-γ), Th17 (IL-17a) and Th2 (IL-4) cytokine mRNAs in lung tissue of RV1B challenged mice (p<0.01 RV-immunised vs RV-adjuvant at 24 hrs post-infection) (Fig. 3a). Consistent with the use of the Th1-promoting adjuvants, this response was dominated by IFN-γ in RV-immunized mice. IFN-γ and IL-17a protein were detected at 24 hrs post-infection only in immunized and challenged mice (p<0.001 vs RV-adjuvant treatment). IFN-γ again dominated with concentrations ∼20× higher than IL-17a (Fig. 3b). IL-4 protein was undetectable in BAL of all groups.
Since immunization generated cross-reactive, VP0-specific cells in the spleen (Fig. 1), we also determined if cross-reactive memory cells were recruited to the airways after infection by measuring IFN-γ production by antigen stimulated lung leukocytes using ELISPOT assays. The frequency of IFN-γ producing lung cells was greatest in mice both immunized and RV challenged (Fig. 3c). Stimulation with homosubtypic immunogen RV16 VP0, with heterotypic RV1B and RV14 VP0 peptide pools, and with live RV1B all induced similar IFN-γ responses (all viral stimuli p<0.001 RV-Immunised vs RV-adjuvant). RV16 VP0 immunization therefore induces cross-reactive Th1/Tc1 responses in the lung in response to RV1B challenge that are of significantly greater magnitude than with RV infection plus adjuvant treatment or immunization with sham infection (Fig. 3c).
RV16 and RV1B belong to different receptor binding groups (major and minor respectively), but are highly related at the nucleotide level [11] and the amino acid level (Fig. S1d) within VP0. To establish if immunization induces more broadly cross-reactive responses we therefore assessed responses to infection with the more distantly related [11] minor group A virus, RV29 (Fig. S1d).
BAL cell staining revealed increased lymphocyte numbers in RV16 VP0 immunized and RV29 infected (RV-immunized) vs adjuvant treated and RV29 infected (RV-adjuvant) mice (p<0.01 day 4, p<0.001 day 7 post-infection)(Fig. 4a). Total and activated CD4+ (Fig. 4b & 4c) and CD8+ (Fig. 4d & 4e) T cell number in BAL and lung tissue were also significantly increased compared to infection or immunization treatments alone. Upon stimulation with RV antigens in ELISPOT assays, IFN-γ producing lung leukocyte frequency was greater in response to challenge serotype (RV29) stimulation in RV-immunized vs RV-adjuvant treated mice (p<0.001)(Fig. 4f). Similar increases were apparent after stimulation with RV1B (p<0.001) and RV14 (p<0.05) derived VP0 peptide pools, again indicating cross-serotype reactivity. We also determined lung T cell-specific IFN-γ production by intracellular flow cytometry staining and observed early (day 1) increases in CD8+ and later (day 6) increases in CD4+ T cells expressing IFN-γ in RV-immunized vs RV-adjuvant, or PBS-immunized treatment groups (RV-immunised vs RV-adjuvant p<0.001) (Fig. 4g).
Significantly increased numbers of activated CD4+ T cells persisted in the lungs of immunized and RV infected mice on day 14 post-infection (Fig. 4c). To determine if this represented enhanced generation of local T cell memory we performed flow cytometric staining for memory markers on lung CD4+ T cells. The proportion and absolute number of CD4+ T cells with a CD44+CD62Llow, effector memory, phenotype was significantly higher in RV29 infected and RV16 VP0 immunized mice compared to either treatment alone (p<0.05). However, no differences were observed between groups in CD44+CD62Lhigh central memory cells (Fig. 4h & 4i).
As neutralizing antibodies are believed important in protection against RV infection, we next investigated the effect of immunization on generation of humoral immune responses by measuring serum and BAL immunoglobulin binding to RVs, and the ability of sera to neutralize RV infection in vitro.
ELISA binding assays showed that immunization with RV16 VP0 in the absence of RV infection weakly induced RV29 and RV1B binding antibodies (Fig. S3a–d). The cross-reactivity of antibodies induced by RV16 VP0 immunization against multiple virus serotypes was also shown by Western Blot (Fig. S3e). When combined with RV1B or RV29 infection in vivo, immunization generated more rapid and greater magnitude of RV-specific serum and BAL IgG responses, and BAL IgA responses, than RV-adjuvant treatment (Fig. S3a–d), indicating that immunization also boosts antibody responses upon subsequent heterotypic RV infection.
We next investigated if enhanced heterotypic antibody responses included boosting of neutralizing activity. Immunization with RV16 VP0 alone did not induce neutralizing antibodies in uninfected mice (Fig. 5a,b). Neutralization of the infecting serotype virus was observed with day 14 post-infection sera of mice treated with adjuvant and infected with RV1B (Fig. 5a), but this was not observed for RV29 (Fig. 5b), suggesting that the neutralizing antibody response to RV in the mouse is either weak or absent. Prior immunization of RV challenged mice however induced both a more rapid induction (day 6) and greater peak titer of neutralizing antibodies (RV1B infection: 50% inhibition dilution [ID50] day 14 RV-immunized 1∶3218 vs RV-adjuvant 1∶160) (Fig. 5a–c). Antibodies induced by RV16 VP0 immunization only neutralized the in vivo infecting RV serotype (data not shown).
These data indicate that immunization with RV16 VP0 is capable of substantially enhancing neutralizing antibody responses to in vivo infection with heterologous RVs.
Finally, we determined whether the Th1 and neutralizing antibody responses induced by immunization conferred any benefit on virus control. Immunization resulted in more rapid virus clearance, as RV1B RNA was undetectable on days 4 & 6 in RV-immunized but not in adjuvant treated mice (Fig. 6).
The unmet medical need attributable to RV infections is enormous but serotypic heterogeneity represents a major barrier to the development of an RV vaccine. We therefore identified regions of the RV polyprotein which are highly conserved amongst RVs to select potential constituents of a broadly cross-reactive subunit vaccine and tested their efficacy in a mouse model. We found that domains of the VP4 and VP2 (VP0) capsid proteins were highly conserved across A and B species RVs.
Immunization with recombinant RV16 VP0 protein increased the magnitude of airway T cell, especially CD4+ T cell, responses to infection consistent with the recruitment to and expansion of immunization-induced memory T cells in the airways. Although the CD4+ T cell dominance of this response contrasts with the prominent CD8+ CTL responses characteristic of other respiratory virus infections [24], [25], [26], there is evidence to suggest this is representative of naturally occurring RV infection [20], [27].
CD4+ T cells provide B cell help and can also possess direct cytotoxic effector function similar to CD8+ CTL and could therefore have both direct and indirect roles in RV control [28], [29], [30], [31].
The observed increases in airway T cell number in immunized and infected mice might in part be explained by the enhanced levels of the T cell recruiting chemokine CXCL10 measured in BAL 24 hrs after infection. Locally induced or systemically transferred memory T cells have previously been shown to increase airway innate immune mediators after influenza challenge via both IFN-γ dependent and independent mechanisms [32]. CXCL10 is an interferon inducible gene and in our studies the increase in CXCL10 might be explained by the enhanced levels of IFN-γ in the lungs of immunized and infected mice at the same timepoint after infection.
There is limited data available regarding T cell polarization during RV infections in humans. In the mouse model little T cell cytokine response was measurable in the airways of infected and adjuvant treated mice, but by combining Th1 promoting adjuvants with the VP0 immunogen we observed a strong type I response to RV challenge and an acceleration of rhinovirus clearance. This is the first clear evidence that such enhancement of type I polarized T cell responses to RV provides benefit in terms of virus control [19], [22], [33]. In addition, asthmatics are a major target group for RV vaccination and Th1/Tc1 responses may also suppress type 2 responses which are associated with increased disease severity during experimental RV-induced disease exacerbations in atopic asthmatics [33], a hypothesis which can now be tested by utilising the mouse RV-induced asthma exacerbation model we have described previously [23]. A key requirement for an RV vaccine is broad cross-reactivity against the ∼150 strains. Human memory CD4+ T cells specific for conserved influenza proteins have been demonstrated to be cross-subtype responsive [17], [18] and we hypothesized that immunization with conserved RV proteins might induce similarly cross-reactive cells. We found that RV16 VP0 immunization induced systemic T cells that were responsive to VP0 peptides from heterologous group A and group B RV serotypes. Following subsequent challenge, cells recovered from the lungs were reactive to the RV16 derived immunogen, to heterologous group A live viruses with which mice were infected and to group B RV VP0 peptides. This cross-reactivity likely represents the recognition of conserved epitopes within VP0, primarily by CD4+ T cells given their greater expansion. Whether this cross-reactivity will be similarly evident in human populations with diverse MHC is not known but a previous study encouragingly showed that VP2 peptides can induce cross-haplotype responses in mice [21]. Further, whilst these studies provide proof of concept for the generation of cross-reactive T cells to RVs, further studies should determine if similar cross-reactivity is seen for the ∼100 other known RV serotypes. Likewise the large number of genetically defined C species RVs which are to date not well characterized [13]. Whether vaccine induced enhancement of Th1 cell responses to RV will prove a safe strategy for preventing RV induced disease awaits confirmation in a clinical trial. However, influenza vaccines are already licenced which use adjuvants which promote strong CD4+ T cell responses and have been shown to be safe [34], [35].
Immunization also induced IgG antibodies which bound multiple RV serotypes and following subsequent infection, enhanced heterologous infection serotype specific antibody levels in serum and BAL. Notably, this included a BAL IgA response which as we have shown previously [36] is otherwise weak or absent after a single infection in this model. Importantly, immunization with RV16 VP0 also enhanced neutralizing antibody responses to infection with the heterosubtypic viruses RV1B and RV29. The fact that generation of neutralizing antibody was dependent upon infection suggests that the effect of immunization on production of serotype-specific neutralizing antibodies following subsequent infection results from B cell help provided by broadly responsive immunization-induced T cells. Enhancement of both the speed and magnitude of antibody responses may provide benefit in terms of accelerating virus clearance and reducing duration of disease caused by naturally occurring infections with virus strains heterologous to that upon which the sequence of the immunogen is based.
Consistent with a role for immunization-induced responses in enhancing virus control, we found that viral RNA was cleared more rapidly from the lungs of immunized mice after subsequent virus infection. This effect was more evident at later stages of infection, which is likely attributable to the fact that virus replication in this mouse model is short-lived compared to human infection, lasting only around 24 hrs [23], [33], [37] and therefore before enhanced T cell responses are apparent. The fact that T cell and antibody responses were able to speed virus clearance in a mouse model where replication is short lived suggests however that in man, where replication is much more robust and of longer duration [33], [37], the magnitude of benefit might be substantially greater.
Pre-existing neutralizing antibodies to RVs provide protection against infection and symptoms in humans [6], [7] and in addition to accelerating virus clearance during the first naturally acquired infection with a given serotype, enhanced neutralizing antibody responses may provide better and more durable protection against future RV infections. Likewise the enlarged effector memory T cell pool in immunized persons, because local memory T cells are likely to respond rapidly to secondary challenge and are proposed to possess more potent anti-viral function than systemic memory cells [38], [39]. Immunization with VP0 may therefore generate serotype specific protective humoral and cross-reactive lung T cell memory responses to natural infection. Because RV infections are frequent throughout life, typically comprising 8–10 per year in young children and 2–5 per year in adults [40], natural infection following immunization could result in protection against a broad range of previously unseen RVs.
In summary, immunization with a recombinant RV capsid protein enhanced airways Th1 cell and airways and systemic antibody responses to infection with heterologous virus serotypes. Immunization also accelerated virus clearance. This study therefore provides proof of principle for a broadly cross-reactive subunit vaccine for RV infections.
All animal studies were conducted according to UK home office legislation (Animals (Scientific Procedures) Act 1986), project licence number PPL 70/7234, or under approval of the Sanofi Pasteur Animal Care Committee protocol numbers F.DI.RVI005.Ms, F.DI.RVI006.Ms and F.DI.RVI007.Ms.
The design of the VP0 immunogen was based on linear sequence conservation amongst RVs. All RV sequences were retrieved from the National Center for Biotechnology Information (NCBI) Genbank database on August 23, 2007 and sequence alignments were generated for all available complete polyproteins from HRV-A and HRV-B using the MUSCLE algorithm [41]. This included 136 polyprotein sequences across 74 A species serotypes and 51 sequences across 25 B species serotypes to take account of variability both between serotypes and between different field strains within serotypes. A phylogenetic tree was elaborated using the maximum likelihood method from the Seaview application [42] and bootstrap values were calculated to assess the robustness of the nodes. A global consensus sequence was generated from the alignments using the Jalview application [43]. Global consensus sequences were extracted from each alignment and frequency of occurrence for each major amino acid was calculated (Fig. S1 a,b).
The VP0 nucleotide sequence was optimized for E. coli expression and synthesized (Life Technologies, Saint Aubin, France). Antigen was expressed as a recombinant protein fused to a SUMO tag using the pET-SUMO vector (Invitrogen, Saint Aubin, France). The Overnight Express Autoinduction System 1 (EMD Millipore, France) was used with BL21λDE3 E. coli transfected with the pET-SUMO plasmid encoding RV16 VP0. As it was expressed into the insoluble fractions as inclusion bodies, purification was then performed according the manufacturer recommendations (Invitrogen) adapted for insoluble proteins.
Briefly, SUMO-fused proteins extracted with Tris/NaCl buffer containing 8M Urea were loaded onto Nickel sepharose columns (Pharmacia) for Immobilized Metal Affinity chromatography (IMAC). Purification was performed by applying an imidazole gradient to the column. Recombinant proteins eluted into the 250 mM imidazole fractions were further dialysed against a digestion buffer (Tris 20 mM, NaCl 150 mM pH 8.0 containing 2M Urea) to cleave the SUMO moiety by the SUMO ULP-1 protease. The RV16 VP0 obtained after digestion was applied onto a second Nickel sepharose column to remove the SUMO moiety, the non-cleaved protein and the protease-containing His tag (Fig. S1e). The cleaved RV16 VP0 obtained after the second purification step was further dialysed against Tris/NaCl buffer (Tris 20 mM, NaCl 150 mM, Arginine 0.5 M, pH 8.0).
Peptide pools for RV1B and RV14 were generated for the VP0 and 3′ polymerase regions. Peptides were synthesized and purified commercially (JPT, Germany). Peptides were 15mers overlapping by 11 amino acids, with each pool comprising approximately 40 peptides. The sequences upon which the respective peptide pools are based are presented in Figure S2.
RV serotype 1B and 29 for in vivo studies were propagated in H1 HeLa cells (American Type Tissue Culture Collection (ATCC) ref CRL-1958) and purified and titrated as described previously [23]. RV stocks were originally obtained from the ATCC. A purified, uninfected HeLa cell lysate preparation was generated as a control for virus-specific immunoglobulin assays.
6–8 week old, wild type, female C57BL/6 mice were purchased from Charles River Laboratories (UK, or Saint Germain sur l'Arbresle, France) and housed in individually ventilated cages.
For immunogenicity experiments (Fig. 1), mice were immunised subcutaneously (s.c.) on days 0 and 21 with 10 µg RV16 VP0 protein, Incomplete Freund's and CpG (IFA/CpG) adjuvant (10 µg CpG 1826 (MWG Eurofins, Germany) and 100 µL IFA), or with adjuvant alone. Further controls received protein buffer (Tris 20 mM, NaCl 150 mM, Arginine 0,5 M pH 8,0) with or without IFA/CpG adjuvant. Mice were culled on day 49.
For RV challenge studies mice were immunised s.c. on days 0 and 21 with a solution containing; 10 µg RV16 VP0 protein, 10 µL CpG oligonucleotide (100 µM ODN 1826) and 40 µL IFA (Sigma-Aldrich, UK) in sterile PBS, or adjuvant alone. On day 51, mice were challenged intranasally with 5×106 TCID50 RV serotype 1B or 29, or mock challenged with PBS, and were culled at the indicated timepoints.
Bronchoalveolar lavage (BAL) was performed and processed as previously described [23]. For lung leukocyte analyses, tissue was homogenized using the GentleMACS tissue dissociator (Miltenyi Biotech, UK) and homogenized tissue was digested in RPMI medium containing 1 mg/ml collagenase type XI and 80units/mL bovine pancreatic Dnase type IV (both Sigma-Aldrich). Red cells were lysed with ACK buffer. For RNA extraction, an apical lobe of the right lung was excised and stored in RNAlater stabilization buffer (Qiagen, UK). Splenocytes were isolated by manually homogenizing spleens through a cell strainer and treating with Hybri Max Red Blood Cell Lysing Buffer (Sigma- Aldrich). Blood was collected from the carotid arteries into ‘microtainer’ serum separation tubes or Vacutainer Vials (both BD Biosciences) and serum was separated by centrifugation.
BAL cells were spun onto slides, stained and lymphocytes were counted as previously described [23]. Counts were performed blind to experimental conditions.
1–10×105 lung or BAL cells were stained with ‘live/dead fixable dead cell stain’ (Invitrogen) and incubated with anti-mouse CD16/CD32 (FC block; BD biosciences). Directly fluorochrome-conjugated antibodies specific for CD3-Pacific Blue (clone 500A2), CD4-APC (clone RM4-5), CD8-PE (clone 53-6.7), CD69-FITC (clone H1.2F3), CD62L-PE (clone MEL-14) and CD44-FITC (clone IM7) (all BD biosciences) were added directly. Cells were fixed with 2% formaldehyde. For intracellular staining, lung cells were stimulated for 4 hrs in media containing 50 ng/mL Ionomycin, 500 ng/mL PMA (Both Sigma Aldrich) and golgi transport inhibitor (Golgi Stop, BD Biosciences). Cells were then surface stained as described, permeablised with 0.5% (w/v) saponin (Sigma-Aldrich) and stained with fluorochrome conjugated anti-IFN-γ-FITC (clone XMG1.2, BD biosciences).
Flow cytometry data was acquired using CyanADP (Dako, USA) and FACSCanto (BD biosciences) cytometers and analysed using Summit software (Dako, USA).
Cytokine and chemokine proteins in BAL were assayed using protocols and reagents from Duoset ELISA kits (R&D systems).
RV-specific IgG and IgA were measured using in-house assays as described previously [36]. 96 well plates were coated overnight with purified RV1B or RV29, as used for in vivo infections, and blocked with 5% milk in PBS-0.05% tween 20. Samples were pooled for each treatment group/timepoint, diluted as indicated in 5% milk blocking solution and plates were incubated for a further 2 hrs at room temperature. Detection antibodies were biotinylated rat anti-mouse IgG1 (clone A85-1), IgG2a/c (clone R19-15) and IgA (clone C10-1) (all BD biosciences) diluted in PBS 1% BSA. Plates were then incubated with spreptavidin-HRP followed by TMB substrate (both Invitrogen) and reactions were stopped by addition of 1M H2SO4 For analysis of IgA in BAL, IgG was first depleted by incubation with protein G sepharose beads (Sigma-Aldrich). Antibody binding to HeLa cell lysate control coated wells was measured in parallel in all assays and values were subtracted from those of virus coated wells during analysis.
4×105 splenocytes per well were distributed in 96 well plates and stimulated with 1 µg/mL of RV peptide pools. Supernatants were harvested after 3 days at 37°C. IL-5 and IFN-γ concentrations were measured using the mouse Th1/Th2 cytokine kit (BD Biosciences) and a Facscalibur cytometer (Becton Dickinson). Data was analyzed on FCAP Array software (Becton Dickinson).
Assays were performed in 96 well multiscreen HA plates (Millipore) coated with purified anti-mouse IFN-γ or IL-5 (BD biosciences). After blocking, 1 or 2×105 lung cells were added, followed by medium containing RV or control stimuli (RV16 VP0 protein (25 µg/mL), live RV1B (2×106 TCID50/mL), RV peptide pools (1 or 4 µg/mL), DMSO peptide pool control, PMA/Ionomycin (50/500 ng/mL)). Plates were incubated for 18 hrs or 3 days at 37°C. Detection antibodies were biotinylated rat anti-mouse IFN-γ or IL-5 (BD biosciences). Plates were subsequently incubated with streptavidin-horseradish peroxidase (Southern Biotech) or extravidin alkaline phosphatase (Sigma-Aldrich) followed by AEC or NBT/BCIP substrate (both Sigma-Aldrich), respectively. Reactions were stopped with water.
In immunogenicity experiments (Fig. 1), IgG responses were analyzed by Western blot of pooled sera. 2 µg of recombinant viral protein and molecular weight standard (SeeBluePlus2, Invitrogen) were run on a 4–12% polyacrylamide SDS gel (Invitrogen). Protein was transferred onto a nitrocellulose membrane (Bio-Rad, USA) and blocked with 5% milk in PBS 0.05% Tween 20. Membranes were probed with (1 in 200) diluted pooled mouse sera followed by HRP-conjugated goat anti-mouse IgG (Jackson ImmunoResearch, UK). Blots were developed colorimetrically using 4-chloro-1-naphthol Opti-4CN substrate (Bio-Rad).
For the study of antibody cross-reactivity (Fig. S3) blots were performed as described but with 1.25 µg virus protein (in vivo inoculum) or 12.5 ng recombinant RV16 VP0. Detection antibody was goat anti-mouse IgG (Santa Cruz biotechnology, USA) and blots were developed using ECL (GE Healthcare, UK).
Neutralisation of RV was measured in Ohio HeLa cells (UK Health Protection Agency General Cell Collection catalogue number 84121901). Sera for given treatment groups/timepoints were pooled and incubated with purified RV at room temperature with shaking for 1 hr, before addition of HeLa cells and further incubation at 37°C for 48–96 hrs. Protection from CPE was measured by crystal violet cell viability assay whereby cells were stained with 0.1% (w/v) crystal violet, washed with water, air dried and crystal violet was solubilised with 1% SDS. Absorbance was measured at 560 nm.
Graphical data is expressed as mean +/− SEM, representative of at least 2 independent experiments. For all data differences between treatment groups were assessed by one or two way ANOVA and if significant (P<0.05) individual differences were identified using bonferroni post-tests.
|
10.1371/journal.pgen.1002322 | A Phenomics-Based Strategy Identifies Loci on APOC1, BRAP, and PLCG1 Associated with Metabolic Syndrome Phenotype Domains | Despite evidence of the clustering of metabolic syndrome components, current approaches for identifying unifying genetic mechanisms typically evaluate clinical categories that do not provide adequate etiological information. Here, we used data from 19,486 European American and 6,287 African American Candidate Gene Association Resource Consortium participants to identify loci associated with the clustering of metabolic phenotypes. Six phenotype domains (atherogenic dyslipidemia, vascular dysfunction, vascular inflammation, pro-thrombotic state, central obesity, and elevated plasma glucose) encompassing 19 quantitative traits were examined. Principal components analysis was used to reduce the dimension of each domain such that >55% of the trait variance was represented within each domain. We then applied a statistically efficient and computational feasible multivariate approach that related eight principal components from the six domains to 250,000 imputed SNPs using an additive genetic model and including demographic covariates. In European Americans, we identified 606 genome-wide significant SNPs representing 19 loci. Many of these loci were associated with only one trait domain, were consistent with results in African Americans, and overlapped with published findings, for instance central obesity and FTO. However, our approach, which is applicable to any set of interval scale traits that is heritable and exhibits evidence of phenotypic clustering, identified three new loci in or near APOC1, BRAP, and PLCG1, which were associated with multiple phenotype domains. These pleiotropic loci may help characterize metabolic dysregulation and identify targets for intervention.
| The metabolic syndrome represents a clustering of metabolic phenotypes (e.g. elevated blood pressure, cholesterol levels, and plasma glucose, as well as abdominal obesity) and is associated with an increased risk of atherosclerosis and type 2 diabetes. Although multiple genes influencing the specific metabolic syndrome components have been reported, few studies have evaluated the genetic underpinnings of the syndrome as a whole. Here, we describe an approach to evaluate multiple clustered traits, which allows us to test whether common genetic variants influence the co-occurrence of one or more metabolic phenotypes. By examining approximately 20,000 European American and 6,200 African American participants from five studies, we show that three regions on chromosomes 12, 19, and 20 are associated with multiple metabolic phenotypes. These genetic variants are highly intriguing candidates that may increase our understanding of the biologic basis of the clustering of metabolic phenotypes and help identify targets for early intervention.
| The metabolic syndrome represents metabolic dysregulation expressed as the clustering of several physiologic risk factors and is associated with an increased risk of atherosclerosis and type 2 diabetes [1]. The core metabolic syndrome domains are abdominal obesity, atherogenic dyslipidemia, elevated blood pressure, elevated plasma glucose, a pro-thrombotic state, and a pro-inflammatory state [2], which are represented to varying degrees in commonly used metabolic syndrome scoring systems [3]–[7].
Several lines of evidence support a genetic basis underlying the core metabolic syndrome domains. Measures of metabolic domains cluster in families [8] and heritability estimates range from 16% for systolic blood pressure to 60% for high-density lipoprotein (HDL) cholesterol [9]. Genome-wide association (GWA) studies have also identified common variants in CETP, LPL, APOA5, and GCKR that influence the co-occurrence of metabolic domain phenotypes [10], [11].
Despite evidence of the clustering of metabolic domain phenotypes, current approaches for identifying unifying genetic mechanisms (i.e. pleiotropy) remain largely focused on clinical categories that do not provide adequate etiological information [12]. As an alternative, a phenomics approach that assembles coherent sets of phenotypic features that extend across individual measurements and diagnostic boundaries creates the opportunity for novel genetic investigations of established biological pathways and complements the traditional GWA study or candidate gene-based strategy focused on individual phenotypes [13]–[15]. In addition to making use of existing knowledge on process-related information or pathways, a multi-phenotype phenomics approach also may provide greater statistical power than analyses of individual phenotypes [16] and improve the ability to detect effects of small magnitude [17]. Although several authors have advocated the use of such strategies [15], [18], [19], the approach is implemented infrequently.
This study evaluated evidence of pleiotropy in clustered metabolic domains using data from five well characterized population-based studies composed of approximately 20,000 European American and 6,200 African American participants: the Atherosclerosis Risk in Communities (ARIC) study, the Coronary Artery Risk Development in Young Adults (CARDIA) study, the Cardiovascular Health Study (CHS), the Framingham Heart Study (FHS), and the Multi-Ethnic Study of Atherosclerosis (MESA). Six phenotype domains (atherogenic dyslipidemia, vascular dysfunction, vascular inflammation, pro-thrombotic state, central obesity and elevated plasma glucose) encompassing 19 quantitative traits were examined. After dimension reduction, we applied a statistically efficient and computationally feasible multivariate approach that related the phenotype domains to 250,000 imputed SNPs. Our approach, which is applicable to studies of heritable, clustered interval scale outcomes, identified several genome-wide significant loci associated with multiple phenotype domains, which may help characterize metabolic dysregulation and identify targets for intervention.
After excluding duplicate samples (N = 56), first- and second-degree relatives (N = 1,152) in all studies except the family-based Framingham Heart Study, and individuals identified as genetic outliers (N = 20), there were 19,468 European American and 6,287 African American Candidate Gene Association Resource Consortium (CARe) participants available for analysis. As expected, CARDIA participants (mean age: 25 years) had better cardiovascular health profiles, including lower low density lipoprotein concentrations, markers of vascular inflammation, and blood pressure levels when compared to the older cohorts (Tables S1, S2, S3, S4, S5).
Eight principle components were used to characterize the six metabolic syndrome trait domains (Figure 1): one principal component each for vascular dysfunction, elevated plasma glucose, pro-thrombotic state and central obesity and two principal components for atherogenic dyslipidemia and vascular inflammation. Correlation between the principal components, which served as the eight phenotypes of interest, was modest and consistent across studies and racial groups. As an example, race- specific results from the ARIC Study are presented in Tables S6, S7.
ARIC and CARDIA were the only studies with full phenotype data for all 19 of the variables used to define the metabolic trait domains. Although apolipoprotein A1 and B measurements were unavailable in three cohorts, the high correlations with high-density and low-density lipoprotein concentrations (r>0.70 in ARIC data, Tables S8, S9) suggested that all five cohorts provided similar atherogenic dyslipidemia phenotypes. A similarly high correlation was observed between von Willebrand factor and factor VIII in the ARIC data, implying a common pro-thrombotic phenotype in studies missing either measurement. The modest correlation between systemic markers of inflammation in the MESA study, which did not measure white blood cell count and uric acid concentration, suggests that this study may contribute a slightly different vascular inflammation phenotype. The MESA study also did not assay factor VII, suggesting that this study also contributed a somewhat different pro-thrombotic phenotype. However, a sensitivity analysis excluding pro-thrombotic and inflammation principal components estimated in the MESA study yielded comparable results.
In European Americans, we identified 606 SNPs representing 19 loci that were associated with at least one metabolic trait domain (Table 1, Figure 2) at the genome-wide significance level (P<2.13×10−7; the SNP with the lowest P – value chosen if multiple significant SNPs were identified for a given locus) and these results were consistent across the multiple large cohorts (Table S10 and Figure S1). Several of these loci overlapped results in African Americans (Table 2, Figure 3), including associations with LPL, ABO, VWF, CTEP, and LDLR. In addition to these 19 loci, we also identified 15 additional secondary signals in European Americans, defined as genome-wide significant SNPs (the SNP with the lowest P – value chosen if multiple significant SNPs were identified for a given locus) in very low linkage disequilibrium (LD) (r2<0.05) with the most significant SNP and within the same 1,000-kb region (Table S11). To verify the independent contributions of these additional loci, we performed a conditional analysis using the most significant SNP at each significant locus as a covariate. Thirteen of these signals remained significant, including one APOC1 variant, after adjusting for the primary signals.
The strongest signal for both European American and African American participants was located on chromosome 9 in the ABO gene (P<1.0×10−300 and P = 6.1×10−75, respectively). These signals overlap earlier findings between factor VIII and von Willebrand factor with ABO [20]. Nine additional loci in European Americans and eight loci in African Americans demonstrated effects limited to one metabolic syndrome trait domain that have already been reported in the GWA literature and are therefore not considered further: ABCA1, APOB, CD36, CELSR2, CETP, CRP, F7, LDLR, LIPC, PVRL2, TRIB1, VWF, and ZNF259.
Six loci were associated with at least two trait domains in European Americans: GCKR, ABCB11, LPL, HNF1A, FTO, and SUGP1, results which overlap published associations identified through GWA studies for individual trait components. For example, several GWA studies have identified associations between GCKR and elevated plasma glucose [21], atherogenic dyslipidemia [22], and vascular inflammation [23]–[25]. GCKR is a plausible unifying mechanism for the clustering of metabolic domains, as the protein inhibits glucokinase, the predominant glucose phosphorylating enzyme [26]. HNF1A, which encodes the transcription factor hepatocyte nuclear factor (HNF)-1a, also suggests a common pathogenic background, as previous GWA studies have identified associations with atherogenic dyslipidemia [27], vascular inflammation [28], and type 2 diabetes [29]. Of note, FTO was the only previously identified and consistently replicated obesity locus we identified.
The strongest new pleiotropic signal in European Americans was for rs4420638 (P 1.7×10−57), located approximately 0.32 kilobases (kb) downstream of APOC1 and associated with elevated plasma glucose (P = 8.7×10−4), atherogenic dyslipidemia (1×10−31), vascular inflammation (P = 5×10−12), and central obesity (P = 1.2×10−6). Although associations between APOC1 with atherogenic dyslipidemia [22], [30], [31] and vascular inflammation [32], [33] have been reported and replicated in the GWA study literature, we consider it a novel locus due to the strong and previously unreported associations with elevated plasma glucose and central obesity. Localizing this signal is challenging, as the region contains a 48-kb gene cluster that also includes the APOE and pseudo-APOC’ genes [34]. However, the modest levels of linkage disequilibrium (Figure 4), the presence of a second signal (Table S11), studies which demonstrate that mice overexpressing human APOC1 show a marked reduction in the update of fatty acids into adipocytes [35], and the fact the physiological role of APOC1 is less well established than APOE, APOB, and APOA1 [36] all support further evaluation and fine mapping of APOC1.
The second new locus was rs11065987 (P = 2.9×10−10), located approximately 9.9 kb upstream of BRAP and associated with atherogenic dyslipidemia (3.1×10−3), vascular dysfunction (2.2×10−4), and central obesity (9.7×10−3). Initial reports suggested that the BRAP protein binds the breast cancer suppressor protein BRCA1 [37]. BRAP is also known to modulate mitogen activated protein kinase signaling [38], an established cell survival, growth, differentiation, transformation, and proinflammatory pathway [39].
The GWA study literature provides few clues that link BRAP with metabolic trait domains, as associations have only been identified for alanine aminotransferase [24] and esophageal cancer [40], both in populations of Japanese descent. However, the recombination rate (cM/Mb) is low from approximately 110.3 Mb to 111.5 Mb (Figure 4) and this extended region includes loci associated with type 1 diabetes [41], [42], vascular dysfunction [43], and waist-hip ratio [44]. The ATXN2 gene, located 27 kb from the index SNP, is an intriguing candidate gene. Expansion of a CAG repeat in the ataxin-2 protein causes the neurodegenerative disease spinocerebellar ataxia type 2. However, instead of a neurodegenerative phenotype, ATXN2-deficient rodents exhibited phenotypes characterized by abdominal obesity, insulin resistance, and marked hepatosteatosis (i.e. lipid accumulation in the liver) [45]. Linkage studies of obesity in humans have also associated this region with BMI and total fat percentage [46].
A third genome-wide significant signal was identified for rs753381 (P = 4.3×10−8), a missense mutation in PLCG1 that results in a change from an isoleucine to a threonine. PLCG1 encodes a protein that catalyzes the formation of inositol 1,4,5-trisphosphate and diacylglycerol from phosphatidylinositol 4,5-bisphosphate and plays an important role in the intracellular transduction of receptor-mediated tyrosine kinase activators [47]. Few epidemiologic studies of PLCG1 or neighboring genes have been published. However, mice nullizygous for PLCG1 stop growing mid-gestation and show no evidence of vasculogenesis [48]. Vasculogenesis has been associated with insulin resistance [49], plasminogen activator inhibitor-1(PAI-1) concentration [50], hyperglycemia, and adiponectin levels [51]. This suggests that PLCG1 may contribute to the clustering of metabolic domains in a more subtle manner, such as through small alterations in the structure of the PLCG1 protein. Thus, the missense mutation we identified would serve as a highly intriguing candidate SNP for further study.
In this study composed of approximately 20,000 European American and 6,200 African American participants, we identified three new loci associated with multiple metabolic trait domains: APOC1, BRAP, and PLCG1. These loci were in or near genes previously associated with atherogenic dyslipidemia, vascular inflammation, type I diabetes, vascular dysfunction, and central adiposity. No previous genome-wide or gene-centric studies examining evidence for pleiotropy in metabolic domains has detected these loci at genome-wide significant levels.
The pathogenesis of the clustering of metabolic phenotypes remains poorly understood, although it is likely that a sedentary lifestyle, combined with dietary patterns and genetic susceptibility factors, contribute. Candidate genes associated with metabolic syndrome phenotypes largely reflect current knowledge of established pathways regulating obesity, free fatty acid metabolism, insulin sensitivity, lipid metabolism, and inflammation. Although candidate gene and GWA studies have successfully identified loci influencing variation in these pathways, studies examining genetic factors influencing the co-occurrence of metabolic phenotypes are limited. Additionally, those that examine the clustering of syndromic components using the pre-defined clinical cutpoints are largely inconsistent or inconclusive. This general lack of success may reflect ongoing controversy over metabolic syndrome definitions, leading to phenotypic heterogeneity and inconsistent genetic findings across studies [52]. The utility of studying the syndrome as a binary entity as opposed to a series of component traits is also debated [12], especially since the dichotomization of interval scale traits will discard information.
Methods for examining evidence of pleiotropy remain uncommon in the GWA literature and most likely reflect the lack of methodologies and software that are scalable to GWA studies. In this paper, we present a statistically efficient and computational feasible approach to testing for pleiotropy on a genome-wide scale. Our method is applicable to population-based and family studies and identified several associations that would not have been identified through typical univariate analyses. The approach presented herein is also not limited to metabolic phenotypes. Instead, our method could be applied to any set of interval scale traits that are heritable and exhibit evidence of phenotypic clustering.
Although alternative analytic approaches were available, for example estimating principal components for all traits simultaneously, we focused on the phenotype clusters presented in Figure 1. First, evaluating the nineteen phenotypes of interest as six domains of interest is biologically plausible given evidence of phenotypic clustering. It was also easier to interpret principal components that were derived in separate phenotype domains rather than components estimated simultaneously. Additionally, estimating principal components within each phenotype domains ensured that each domain was sufficiently represented in the analysis.
Challenges to the approach presented herein include careful phenotype curation, made more difficult by the inclusion of 19 traits across multiple cohorts that were not measured with a common protocol. Only the ARIC and CARDIA studies had full phenotype information on all 19 traits and CHS was the only study with all traits measured during a single visit. The use of a multivariate phenotype comprised of 19 variables also limited the number of contributing cohorts and the identification of replication cohorts, as few studies have such comprehensive phenotypic data. Nonetheless, we were able to identify approximately 25,000 participants from studies that used standardized, comparable protocols and many of the associations were consistent across cohorts.
Further challenges that are not unique to large scale genetic studies incorporating a phenomics approach include the consistency of results across populations defined by age, race, sex, or other demographic characteristics. For example, the three new loci identified in the European American population were not detected in the African American population. Given a modest sample size of 6,287 participants it is difficult to determine whether an inability to generalize results to the African American population reflects different patterns of LD, varying environmental contexts, or limited statistical power. Variation in mean age between contributing cohorts, which ranged from 25 years in the CARDIA study to 72 years in the CHS, could introduce additional heterogeneity, as associations between metabolic phenotypes have been shown to diminish with age [53]. Finally, marked variation in the prevalence of the metabolic syndrome by gender, regardless of clinical definition, suggest the possibility of sex-specific metabolic syndrome effects [54]. Analyses that examine modification by sex, age, and other important clinical covariates are therefore warranted.
Our use of the IBC array, which is composed of variants implicated in cardiovascular, inflammatory, hemostasis/coagulation, and metabolic pathways, was beneficial in that it allowed us to leverage the wealth of information on pathways implicated in metabolic disturbances while reducing multiple testing penalties. Admittedly this approach was limited in that it potentially excludes novel pathways not captured by the IBC chip. Although imputation allowed us to increase the number of variants, genome-wide approaches might identify additional pleiotropic loci.
In summary, our results support phenomics as a complementary approach that leverages phenotypic variation for the evaluation of pleiotropy, a clear limitation of existing studies examining the metabolic syndrome using clinical definitions. Our approach, which is applicable to studies of heritable, clustered interval scale outcomes, also takes advantage of the wealth of phenotype data available in longitudinal cohort studies as well as emerging analytical and bioinformatics approaches. Ultimately, these results support the presence of genetic variants with pleiotropic effects on adiposity, inflammation, glucose regulation, dyslipidemia, vascular dysfunction and thrombosis. Such loci may help characterize metabolic dysregulation and identify targets for intervention.
This study arose from a collaboration between investigators from two National Institute of Health funded consortia examining the genetic basis of common complex diseases: the Population Architecture using Genomics and Epidemiology (PAGE) study, a National Human Genome Research Institute funded effort examining the epidemiologic architecture of common genetic variation that have been reproducibly associated with human diseases and traits [55] and the CARe Consortium [56], a National Heart, Lung, and Blood Institute-supported resource for genetic analyses examining cardiovascular phenotypes. Briefly, PAGE investigators participating in the phenomics working group wanted to extend existing efforts examining evidence for pleiotropy in approximately 300 replicated genetic variants [57] to include a more comprehensive evaluation of common SNPs. A collaboration between PAGE and CARe investigators was therefore initiated, and used data from five CARe studies of European American and African American with adequate phenotype data: ARIC, CARDIA, CHS, FHS, and MESA. All participating institutions and CARe sites obtained Institutional Review Board approval for this study. Additional information on the participating CARe studies is provided in Text S1.
The Institute for the Translational Medicine and Therapeutics (ITMAT)-Broad-CARe (IBC) genotyping array [58] was used to evaluate approximately 2,100 genes related to cardiovascular, inflammatory, hemostasis/coagulation, and metabolic phenotypes and pathways. The IBC array tagging approach was designed to capture maximal genetic information for both common and lower frequency SNPs (<5% minor allele frequency (MAF)) in HapMap as well as European American and African American populations. The array included 49,320 SNPs, 15,000 of which were gene variants not present in HapMap. Additional details of the SNP selection and tagging approach are given in Text S1.
Imputation of untyped and missing SNP genotypes was performed using MACH 1.0.16. [59] For the European samples, phased haplotypes from the CEU founders of HapMap 2 were used as reference. For African American populations, a combined CEU+YRI reference panel was created that includes SNPs segregating in both CEU and YRI, as well as SNPs segregating in one panel and monomorphic and non-missing in the other. Imputation for the IBC array was performed in two steps. First, individuals with pedigree relatedness or cryptic relatedness were filtered. A subset of individuals was randomly extracted from each panel and used to generate recombination and error rate estimates for the corresponding sample. Second, these rates were used to impute all sample individuals across the entire reference panel. Before cleaning, there were an average of 246,740 (range: 245,816, 247,505) and 227,224 (range: 225,111, 229,061) imputed SNPs in the European American and African American study populations, respectively. Imputation results were then filtered at an imputation quality limit of 0.30 and a MAF threshold of 0.01, yielding 235,077 (95.3% of total) and 227,222 (96.2% of total) SNPs for analysis in European American and African American participants, respectively.
The clustered risk factors of interest were characterized as a six-domain phenotype: atherogenic dyslipidemia, vascular dysfunction, vascular inflammation, pro-thrombotic state, elevated plasma glucose, and central obesity (Figure 1). These domains were constructed a priori based on a review of literature examining clustering in metabolic phenotypes, placing specific emphasis on the National Cholesterol Education Program’s Adult Treatment Panel III report [4], [60]. Nineteen variables were then selected to represent one of the six domains with preference for variables measured in at least four of the contributing cohort studies or variables that were highly correlated with available measures. Measurement protocols for each variable by study are provided in Table S21. We assessed normality, and transformations were used when variables exhibited excessive skewness or kurtosis as determined by numerical summary information and visual inspection of histograms and normal probability plots. Dimension reduction using principal components analysis was then performed for each phenotype domain separately in each race/ethnic and study population. For example, principal components for the vascular inflammation domain were calculated using the following traits: albumin, C reactive protein, fibrinogen, uric acid, and white blood cell count. Principal components were chosen so that>55% of the variance for each domain was explained (Tables S12, S13, S14, S15, S16, S17, S18, S19, S20). This threshold was chosen because all of the first (waist circumference, pro-thrombotic state, elevated plasma glucose, and vascular dysfunction) and the sum of first and second (vascular inflammation and atherogenic dyslipidemia) principal components exceeded 55% across all studies and racial/ethnic groups.
For each phenotype, we fit a linear regression model relating the phenotype to the SNP genotype under the additive mode of inheritance; the model includes environmental variables (i.e., age, sex and study center) as well as the first ten principal components from EIGENSTRAT to adjust for population substructure [61]. Ten population substructure components were included because each component was associated with at least one of the eight phenotypes of interest in at least one study. If the SNP genotype is not associated with any phenotype domain, then the regression coefficients for the SNP genotype are zero in all eight linear models. We tested this global null hypothesis by constructing a multivariate test statistic based on the joint distribution of the score statistics from the eight linear models, which accounted for the correlation between the eight phenotypes. We chose the score statistic because it is computationally efficient and numerically stable. The test statistic is referred to the chi-squared distribution with eight degrees of freedom. The genome-wide significance level was set as P<2.13×10−7 (i.e. 0.05/235,077). Q-Q plots by race are not presented, as our use of a gene-centric array highly enriched for metabolic loci complicated the identification of markers with low prior probabilities of association (i.e. “null markers”) for all phenotypes of interest. The data from each cohort were analyzed separately and the results were combined via meta-analysis as described in Text S2. All analyses were stratified by race and were performed in SAS 9.1 and C++. Further details are given in the Text S2.
|
10.1371/journal.ppat.1006284 | Arabidopsis leucine-rich repeat receptor–like kinase NILR1 is required for induction of innate immunity to parasitic nematodes | Plant-parasitic nematodes are destructive pests causing losses of billions of dollars annually. An effective plant defence against pathogens relies on the recognition of pathogen-associated molecular patterns (PAMPs) by surface-localised receptors leading to the activation of PAMP-triggered immunity (PTI). Extensive studies have been conducted to characterise the role of PTI in various models of plant-pathogen interactions. However, far less is known about the role of PTI in roots in general and in plant-nematode interactions in particular. Here we show that nematode-derived proteinaceous elicitor/s is/are capable of inducing PTI in Arabidopsis in a manner dependent on the common immune co-receptor BAK1. Consistent with the role played by BAK1, we identified a leucine-rich repeat receptor-like kinase, termed NILR1 that is specifically regulated upon infection by nematodes. We show that NILR1 is essential for PTI responses initiated by nematodes and nilr1 loss-of-function mutants are hypersusceptible to a broad category of nematodes. To our knowledge, NILR1 is the first example of an immune receptor that is involved in induction of basal immunity (PTI) in plants or in animals in response to nematodes. Manipulation of NILR1 will provide new options for nematode control in crop plants in future.
| Host perception of pathogens via receptors leads to the activation of antimicrobial defence responses in all multicellular organisms, including plants. Plant-parasitic nematodes cause significant yield losses in agriculture; therefore resistance is an important trait in crop breeding. However, not much is known about the perception of nematodes in plants. Here we identified an Arabidopsis leucine-rich repeat receptor-like kinase, NILR1 that is specifically activated upon nematode infection. We show that NILR1 is required for the induction of immune responses initiated by nematodes and nilr1 loss-of-function mutants are hypersusceptible to a broad category of nematodes. Manipulation of NILR1 will provide new options for nematode control in crop plants in the future.
| Plant-parasitic nematodes attack the majority of economically significant crops, as shown by international surveys indicating an overall yield loss of 12%. In some crops, such as banana, a loss of up to 30% has been reported. Losses amount to $100 billion annually worldwide [1]. The economically most important nematodes belong to the group of sedentary endoparasitic nematodes that includes root-knot nematodes (Meloidogyne spp.) and cyst nematodes (Globodera spp. and Heterodera spp.). Most chemical pesticides used for control of plant-parasitic nematodes are environmentally unfriendly, expensive and ineffective in the long term. Therefore, an increased demand for novel crop cultivars with durable nematode resistance is inevitable [2, 3]. In this context, it is important to identify and characterize the different natural means by which plants defend themselves against nematodes.
The infection cycle for root-knot and cyst nematodes begins when second-stage juveniles (J2) hatch from eggs. J2, the only infective stage, search for roots guided by root exudates. They invade the roots by piercing the epidermal root cells using a hollow spear-like stylet. After entering the roots, they migrate through different cell layers until they reach the vascular cylinder. There, root-knot nematodes induce the formation of several coenocytic giant cells, whereas cyst nematodes induce the formation of a syncytium. Because established juveniles become immobile, the hypermetabolic and hypertrophic feeding sites serve as their sole source of nutrients for the rest of their lives. In a compatible plant-nematode interaction, plant defence responses are either down-regulated or overcome by the nematodes [4–6]. A cocktail of secreted molecules including effectors that are synthesized in the oesophageal glands of the nematodes is purportedly responsible for modulating the plant defences as well as the induction and development of the syncytium [7–10]. Whereas most root-knot nematodes reproduce parthenogenically, cyst nematodes reproduce sexually. Although the mechanism of sex determination in cyst nematodes is not clear, studies have shown that the majority of juveniles develop into females under favourable nutritional conditions. When juveniles are exposed to adverse growth conditions, as it is the case with resistant plants, the number of male nematodes increases considerably [11].
Numerous studies have shown that plants sense microbes through the perception of pathogen/microbe-associated molecular patterns (PAMPs or MAMPs) via surface-localised pattern recognition receptors (PRRs), leading to the activation of PAMP-triggered immunity (PTI). The activation of PTI is accompanied by the induction of an array of downstream immune responses including bursts of calcium and reactive oxygen species (ROS), cell-wall reinforcement, activation of mitogen-associated and calcium-dependent protein kinases (MAPKs and CDPKs), and massive reprogramming of the host transcriptome [12–15]. Together, these downstream responses can fend off the pathogen’s infection. PAMPs are typically evolutionary conserved across a class of pathogens and perform an important function in the pathogen life cycle [16].
Plant PRRs are either plasma membrane-localised receptor-like kinases (RLKs) or receptor-like proteins (RLPs) [14]. Both RLKs and RLPs consist of an extracellular receptor domain (ECD) for ligand perception, a single membrane-spanning domain, but only RLKS have a cytoplasmic kinase domain. The major classes of RLKs are leucine-rich repeat (LRR)-RLKs, lysine-motif (LysM)-RLKs, crinkly4 (CR4)-RLKs, wall-associated kinases (WAKs), pathogenesis-related protein 5 (PR5)-RLKs and lectin-RLKs (LeCRKs). Nevertheless, it is becoming increasingly clear that PRRs do not act alone but are part of multiprotein complexes at the plasma membrane [13]. For example, the LRR-RLK BRASSINOSTEROID INSENSITIVE-1 (BRI1)-ASSOCIATED KINASE 1 (BAK 1) forms receptor complexes with various LRR-containing PRRs to positively regulate PTI [14–15, 17]. In addition to PAMPs, plant PRRs can also perceive endogenous molecules, so-called damage-associated molecular patterns (DAMPs) that are released upon cell damage or pathogenic attack [16].
Although extensive studies have been conducted to characterise the role of PTI response in various models of plant-pathogen interactions, relatively less information is available pertaining to nematode-induced PTI responses in plants. To date, no PRR that recognises a nematode-associated molecular pattern (NAMP) has been identified [18]. However, some recent work suggests that nematode infection triggers PTI responses in host through surface-localised receptors. For example, silencing of the orthologues of BAK1 in tomato (Solanum lycopersicum, Sl) (SlSERK3A or SlSERK3B) has been shown to increase the susceptibility of these plants to nematodes due to defects in activation of basal defence [19]. In a more recent publication, it was shown that nematode infection triggers PTI responses in Arabidopsis in a BAK1-dependent and BAK1-independent manners. These authors showed that several PTI-compromised mutants including bak1-5 were significantly more susceptible to root-knot nematodes as compared to control [20]. However, the identity of ligands and/or receptors involved in BAK1-mediated response remains unknown. As far as NAMP identification is concerned, ascarosides, which are conserved nematode-secreted molecules, have been shown to elicit plant defence responses that lead to reduced susceptibility against various pathogens [21].
In comparison to PTI, Effector-triggered immunity (ETI) during plant-nematode interaction is relatively well studied. A number of host resistance genes (R-genes) against nematodes have been described and their mode of action is relatively well investigated [22]. Notably, a host cell-surface immune receptor Cf-2 has been shown to provide dual resistance against a parasitic nematode Globodera rostochiensis and a fungus Cladosporium fulvum through sensing perturbations of the host-derived protease RCR3 by the venom allergen-like protein of Globodera rostochiensis [23]. In the present study, we provide evidence that nematodes induce PTI-like responses in Arabidopsis that rely on the perception of elicitors by membrane-localised LRR-RLKs.
To reveal changes in gene expression in response to nematodes at and around the infected area, GeneChip analysis was performed. Small root segments (approx. 0.5 cm) containing nematodes that were still in their migratory stage (defined as continuous stylet movement), were cut and compared with corresponding root segments from plants that were not infected. Total RNA was extracted, labelled, and amplified to hybridize with the GeneChip Arabidopsis ATH1 Genome (Affymetrix UK Ltd). The ATH1 Genome Array contains more than 22,500 probe sets representing approximately 24,000 genes. Subsequent analysis of the data showed that approximately 2,110 genes were differentially expressed (FDR < 0.05; Fold change > 1.5). Among them, 1,139 were upregulated, whereas 971 were downregulated (S1 Data). To explore regulation of the biological processes, molecular functions, and their distribution across different cellular components, a gene ontology enrichment analysis was performed on significantly upregulated genes. Those categories which were particularly over-represented in the differentially upregulated genes included the immune system response, response to stimulus, death, and the regulation of the biological processes (Fig A in S1 Text). We have previously published a subset of 62 genes representing selected jasmonic acid (JA), ethylene (ET) and salicylic acid marker (SA), signalling and biosynthesis genes from this GeneChip data, which were also validated by qRT-PCR [24]. In general, transcript levels of genes involved in JA/ET signalling and biosynthesis were increased. However, in comparison to JA/ET, changes in SA-related genes were relatively less pronounced. Nevertheless, a slight increase in a SA biosynthesis (PAL1) and few SA signalling genes (NPR1, NPR3) was also observed (S2 Data). A detailed look at the transcriptomic data indicate that nematode infection triggered the induction of genes previously shown to be induced during PTI (Fig 1A) [25–27].
Our transcriptome data showed the induction of PTI-like responses upon nematode infection, however, it was unclear whether this induction was due to the recognition of nematodes by plant receptors or whether it was the result of wounding due to continuous nematode movement. To clarify this, we established a PTI screening assay involving the measurement of ROS burst, one of the hallmark responses of PTI. For this purpose, we incubated the pre-infective J2 of H. schachtii in H2O for 24 hours at RT. The water obtained after removing the nematodes was termed as NemaWater (Heterodera schachtii NemaWater, HsNemaWater; Meloidogyne incognita NemaWater, MiNemaWater) and was used to treat Arabidopsis roots (see Methods for details). After treatment, ROS burst was measured using a root-based procedure adapted from a previous work [27]. Flg22 and H2O treatments were used as positive and negative controls, respectively. Treatment with flg22 as well as with HsNemaWater induced a strong and consistent ROS burst in roots (Fig 1B). The ROS burst with HsNemaWater was, however, slightly delayed as compared to flg22; the ROS burst to flg22 occurs within 10 to 40 min, while that to HsNemaWater occurred after 20 to 120 min. Although HsNemaWater induced a consistent ROS burst in Arabidopsis roots, it was not clear whether this is due to the presence of a NAMP in HsNemaWater or whether it is due to the production of an eliciting-molecule by plants (upon NemaWater treatment), which in turn induced production of ROS burst in roots. Such an eliciting-molecule could be called as DAMP or a NIMP (nematode-induced molecular pattern). One way to address the question of NAMP, or DAMP/NIMP was to dilute the HsNemaWater with H2O and analysed the production of ROS burst in roots. We hypothesised that if ROS burst is due to production of a DAMP or NIMP, diluting the NemaWater would not only reduce the magnitude of the ROS burst but may also slow its kinetics. However, our data showed that although magnitude of ROS burst was reduced strongly upon dilution, there was no delay in production of ROS between different dilutions (Fig 1C). Next, we incubated the HsNemaWater with Arabidopsis roots for 60 min and then used this HsNemaWater for production of ROS burst on fresh roots. The data showed that prior incubation of HsNemaWater with roots did not cause any significant change in magnitude as well as kinetics of ROS Burst (Fig 1D). Regardless of the nature or origin of elicitor, activation of ROS burst upon HsNemaWater treatment confirmed our observations from transcriptomic studies indicating that PTI-like responses are induced upon nematode detection.
To confirm whether NemaWater from different species of nematodes elicit a similar response, we produced NemaWater from the root-knot nematode species, Meloidogyne incognita (MiNemaWater) and performed ROS burst assays. We observed a strong and consistent ROS burst (Fig 1E) similar to that of H. schachtii (Fig 1B). A prolonged treatment of young Arabidopsis seedlings with flg22 activated defense responses and leads to growth inhibition [28]. Although the mechanism underlying this growth inhibition is unclear, it is commonly accepted that activation of defense responses may take the resources away from growth. Importantly, this assay has frequently been used to analyse the eliciting capacity of PTI components [28, 29]. We tested whether NemaWater also caused seedling growth inhibition, and found that both flg22 and HsNemaWater inhibited seedling growth and reduced the root weight to a similar extent (Fig 1F, Fig B in S1 Text). Our results suggest that NemaWater contains potential elicitor/s that is/are recognized by an immune receptor in plants leading to the activation of PTI-like responses. To test this hypothesis, we incubated 12-day-old Arabidopsis seedlings in HsNemaWater for one hour: ddH2O alone was used as a control. RNA was extracted from the roots of both the non-treated control and NemaWater-treated seedlings. They were subsequently labelled, amplified, and hybridized with a GeneChip, as described above. The data analysis showed that 2,520 genes were differentially expressed, of which, 1,422 were upregulated and 1,098 were downregulated (FDR < 0.05; Fold change > 1.5; S3 Data). A gene ontology enrichment analysis for differentially upregulated genes showed the over-representation of categories such as immune system response, response to stimulus, death, signaling and the regulation of the biological processes (Fig C in S1 Text). A look at the expression of hormonal response gene upon HsNemaWater treatment showed the same tendency for upregulation of JA/ET-related genes as observed upon nematode infection as described above (S2 Data). Moreover, a significant increase in the expression of genes characteristics for PTI was detected (Fig 2A). This upregulation in expression of PTI marker genes was very similar to that observed upon infection with nematodes (Fig 2B). Interestingly, expression of camalexin biosynthesis genes (PAD3/CYP71B15, CYP71A12) was upregulated only in nematode-infected plants but was not regulated upon HsNemaWater treatment (Fig 2B). This was further confirmed by analyzing a reporter line (pCYP71A12:GUS) [30] on treatment either with nematodes or with HsNemaWater. We found a strong GUS expression upon nematode infection, whereas such an expression was absent in seedlings treated with HsNemaWater (Fig 2C–2E). We validated the microarray data by measuring the expression of 13 genes via qRT-PCR upon treatment with HsNemaWater. Our analysis showed a similar trend for expression of selected genes as shown by microarray data (Table 1). Together, these results suggest that both nematode infection and NemaWater treatment induce PTI responses including a significant activation of JA pathways. The data analysis also showed that the changes in gene expression triggered upon treatment of seedlings with HsNemaWater were to an extent similar to those that were observed upon nematode infection (Fig 2F and S4 Data). Even so, both treatments induced expression of a distinct set of genes, which may reflect differences in both treatments such as number and concentration of elicitors, duration of treatments, physical damage, etc.
On the basis of our finding that NemaWater triggers PTI responses, we asked whether pre-treatment with NemaWater effects plant responses to nematodes and other pathogens. To test this, plants were pre-treated with HsNemaWater 24 hours prior to inoculation and were then infected with juveniles of H. schachtii or M. incognita or the virulent bacterial pathogen Pseudomonas syringae pv. tomato (see Methods for details). We found a strong decrease in number of nematodes in HsNemaWater-treated plants compared with Col-0 (Fig 3A and 3B, Fig D in S1 Text). Similarly, the growth of virulent P. syringae was also reduced strongly upon HsNemaWater treatment (Fig 3C and 3D).
Induction of PTI by NemaWater indicated the presence of putative elicitor(s) in NemaWater. To test whether these elicitors is/are of proteinaceous nature, we added Proteinase K to HsNemaWater and performed a ROS production assay. Duration and intensity of NemaWater-induced ROS burst varied in different experimental batches, which may be due to differences in the concentration of elicitors in different preparations of NemaWater and the possibility that NemaWater may contain more than one elicitor. Therefore, we used total photon count as a more reliable parameter for quantification of ROS burst activation in this study. We observed that the treatment of HsNemaWater with Proteinase K or heat strongly reduced the induction of ROS burst (Fig 4A). These results were further confirmed by seedling growth inhibition assays (Fig 4B). BAK1 has been shown to act as a co-receptor for LRR-RLKs and LRR-RLPs, which typically detect proteinaceous ligands [14, 15]. Considering the data from Proteinase K treatment (Fig 4A and 4B) and recently published data on root-knot nematodes [20], we hypothesized that bak1 mutants would be more susceptible to cyst nematodes. A nematode infection assay was performed on bak1-5 and the double mutant bak1-5 bkk1-1 (BKK1 being the closest homolog of BAK1) [31]. Both mutants were significantly more susceptible to nematodes compared with Col-0, as they allowed more females to develop (Fig 4C). We also investigated whether BAK1 is required for PTI-responses upon HsNemaWater treatment and found that the nematode-derived ROS burst was strongly reduced in bak1-5 mutants (Fig 4D). Similar results were obtained in seedling growth inhibition assays (Fig 4E and Fig E in S1 Text).
Within the group of 593 commonly upregulated genes between two microarray experiments, 52 genes encoded RLKs (including 11 LRR-RLKs, 7 LeCRKs and 1 LysM-RK) and 2 encoded RLPs (S4 and S5 Data). Out of 52 candidate RLKs, we selected homozygous loss-of-function T-DNA mutants for ten genes (from five different RLK families), including those coding for three LRR-RLKs and one LeCRK. Confirmed loss-of-function mutants were then screened for infection against H. schachtii. Of particular interest, we found one LRR-RLK mutant, termed NILR1 (NEMATODE-INDUCED LRR-RLK 1; NILR1, At1g74360), which showed a consistent increase in the number of female nematodes as compared with Col-0 (Fig 5A and Fig F and G in S1 Text). In comparison to nilr1-1, the loss-of-function mutant for NILR2 (AT1G53430) did not show any change in susceptibility to nematodes (Fig 5A). Based on our data with Proteinase K and BAK1, we hypothesized that NILR1 may be a PRR involved in the perception of nematodes. Therefore, this study focused on the characterization of NILR1 and NILR2, while other candidate genes will be described elsewhere.
To test NILR1’s involvement in nematode perception other than H. schachtii, we analysed nilr1-1 mutants for infection with root-knot nematode M. incognita. Our data showed that nilr1-1 was significantly more susceptible to M. incognita than Col-0. In comparison, there was no change in susceptibility of nilr2-1 to M. incognita (Fig 5B). To investigate whether enhanced susceptibility of nilr1-1 to nematodes is due to impairment in PTI responses, we performed ROS burst assays on root segments from Col-0 and nilr1-1 upon treatment with NemaWater from two different nematode species (H. schachtii and M. incognita). Notably, the NemaWater-induced ROS burst was strongly reduced in nilr1-1 (Fig 5C and Fig H in S1 Text). Similar results were obtained in seedling growth inhibition assays (Fig 5D and Fig I in S1 Text). We also tested nilr2-1 for seedling growth inhibition and ROS burst induction upon treatment with NemaWater. We found that even though ROS production was reduced in nilr2-1 upon HsNemaWater treatment, the growth of these plants was inhibited to the same extent as Col-0 (Fig 5E and 5F and Fig I in S1 Text). Next, we isolated an additional homozygous knock-out T-DNA line for NILR1 (nilr1-2) and analysed it for infection by H. schachtii and production of ROS burst upon HsNemaWater treatment (Fig J-L in S1 Text). We observed that nilr1-2 plants were impaired in ROS production and were also significantly more susceptible to H. schachtii as compared to Col-0 (Fig K-L in S1 Text). Together our results show that NILR1 is an important component of host immune responses that are activated upon nematode infection.
NILR1 is closely related to LRR-RLK BRI1, belonging to the subfamily X of LRR-RLKs [32]. NILR1 encodes a serine/threonine kinase with 1,106 amino acid residues (predicted molecular weight 121.8 kDa) and shows all of the characteristics of an LRR-RLK. NILR1 has been suggested to have an extracellular domain with 22 tandem copies of LRRs, which are interrupted by a 76-amino acid island located between LRR17 and LRR18. The island domain of NILR1 is longer than those of BRI1 and contains a cysteine cluster with the pattern of Cx25Cx16C, which is followed by a transmembrane domain and a cytoplasmic kinase domain (Fig M-N in S1 Text) [31]. Moreover, a pair of cysteines at the amino terminal flanks NILR1’s LRR domain with the characteristic spacing formerly observed in several plant LRR-RLKs [33]. Previous analysis has shown that NILR1 is presumably localised to the cell membrane, and that homologs are conserved among ten different species of flowering plants [32]. To gain further insights into molecular functions of NILR1, we determined its subcellular localization by confocal microscopy transiently expressing 35S::NILR1-GFP in the epidermis of Nicotianna benthamiana. We detected a strong GFP signal at the plasma membrane (PM) (Fig 6A). The PM localization of NILR1 was confirmed by co-localization with PM marker (see Methods for details). To investigate the conservation of NILR1, we conducted a BLAST search using ECD’s amino acid sequence of NILR1 against non-redundant protein sequences of all land plants. We detected homologues of NILR1 among different species of the Brassicaceae family. Additionally, orthologues of NILR1 were found to be widely conserved in the genome of various dicotyledonous as well as monocotyledonous plant species. (Fig O in S1 Text). To further determine whether NILR1 is conserved across the plant kingdom and to test for effects of NemaWater, we measured the ROS burst upon HsNemaWater treatment in the dicotyledonous tomato, sugar beet (Beta vulgaris) and tobacco (Nicotianna benthamiana), as well as in monocotyledonous rice (Oryza sativa). We detected a strong ROS burst in sugar beet and tomato (Fig 6B and 6C), the magnitude of ROS burst was delayed and reduced in N. benthamiana (Fig 6D). In comparison to dicotyledonous, experiments with monocotyledonous rice showed that NemaWater induce a ROS burst, which was above the water control (Fig 6E). However, this burst was strongly delayed and was not consistent across several experiments.
A further exploration of publicly available Arabidopsis expression data through the eFP browser [33] revealed that NILR1 is only moderately expressed in sepals and in senescent leaves under controlled growth conditions. However, NILR1 expression is upregulated in response to biotic stresses such as Botrytis cinerea, Phytophthora infestans and non-adapted Pseudomonas syringae strains (Fig P and Q in S1 Text). Also NILR1 shows a low basal expression in various root tissues but displays a relatively high expression in endodermis, pericycle and stele [34]. The overall structure of NILR1 and its similarity to BRI1 supports its role as a surface-localised receptor that is involved in the perception of extracellular signals.
In comparison to other pathosystems, not much is known about the importance of PTI in host defense against nematodes. In fact, no PRR involved in nematode perception has thus far been characterized. Additionally, so far only ascarosides have been recently shown to act as NAMPs. On the other hand, a number of nematode resistance genes (R-genes) either at the cell surface or inside cells have been characterised [22, 23]. In the present study, we provide insights into the molecular events associated with the basal resistance of plants to nematodes. We demonstrate that PTI-like responses are activated upon nematode infection and that they contribute significantly to basal resistance against nematodes.
The observation that cyst nematode infection induces the activation of a number of JA biosynthesis and signalling genes during migratory stages is supported with biochemical measurements showing an elevated amount of JA in Arabidopsis roots 24 hours after nematode infection [24]. In contrast to JA there was no strong activation of SA signalling in our transcriptome data during migratory stages. Nevertheless, a slight increase in some SA biosynthesis and signalling genes was observed. Intriguingly, plants that are deficient in different aspects of SA-signalling and biosynthesis have been shown to be more susceptible to cyst nematode infection [35]. These observations raise the question as to whether JA activation in roots upon nematode infection is only because of wounding during migration. Remarkably, we observed the same pattern of JA activation in roots upon treatment with HsNemaWater indicating that JA activation is an important component of defense responses that are activated upon nematode recognition and is not only correlated to wounding. This hypothesis contradicts the general view that SA plays a more prominent role against biotrophs while JA/ET appears to be more important in resistance against necrotrophic pathogens and herbivorous insects [36–38]. This view, however, is mainly based on observations with leaf pathogens, whereas only limited information is available on the role of plant hormones in defense against root pathogens [39]. It may be that JA plays a more dominant role in the plant-pathogen interactions in roots. This hypothesis is supported by experiments on rice plants that indicated a key role for JA during interaction with root-knot nematodes [40]. Unlike the migratory phase, a number of studies addressing changes in gene expression during the sedentary phase of cyst and root-knot nematodes infection revealed a strong suppression of host defence responses [4–6]. Based on data from the current study and previous literature, we concluded that nematode invasion activates PTI responses, which are suppressed during later stages of nutrient acquisition and feeding site development. Indeed, an increasing number of nematode effectors involved in suppression of PTI have been characterised during last few years [8, 10, 18, 22, 23].
We observed that NemaWater treatment triggers responses, including ROS burst, immune gene expression and seedling growth inhibition that are characteristic of PTI. In addition, plants treated with NemaWater were more resistant to nematodes compared with water-treated control plants. On the basis of these data we propose that NemaWater contains elicitor/s that is/are perceived by plant surface-localised receptors leading to activation of PTI.
The fact that NemaWater derived from two different nematode species induces similar responses suggests that the elicitor component/s is/are conserved among different nematode species. Although the identity of the elicitor in NemaWater remains unknown, it is likely to be a heat-sensitive protein since treatment with heat as well as with Proteinase K strongly reduced its activity. Nevertheless, the residual growth inhibition in spite of addition of Proteinase K in NemaWater hints towards the possibility of an additional non-proteinaceous NAMP in NemaWater. However, it is also plausible that the residual growth inhibition is caused by Proteinase K itself. This view is supported by our data (Fig 4A) and some previous studies where a slight ROS burst was observed upon Proteinase K treatment alone [27].
NemaWater-induced responses are dependent on BAK1, which has been shown to act as a co-receptor for LRR-type PRRs, which typically detect proteinaceous ligands [12, 15, 17]. Even though we hypothesise that the NemaWater-derived elicitor/s is/are perceived by a surface-localized receptor, the possibility remains that such elicitor/s may not come into contact with host plants during infection. However, the fact that NemaWater was produced by incubating the nematodes without any further treatment strongly supports the idea that the elicitor is naturally secreted into the environment. It is also possible that the treatment of seedlings with NemaWater leads to the release of plant endogenous elicitors (DAMPs), which are again sensed by plants leading to the activation of PTI responses. However, since diluting NemaWater reduced only the magnitude but did not slow down the kinetics of ROS burst and thus makes it unlikely that a NemaWater induced DAMP is responsible for activation of PTI responses. Regardless of the origin of elicitor, it is clear that induction of PTI responses involves a component of NemaWater (therefore a NAMP) and is not only due to direct mechanical wounding by nematodes.
Loss of NILR1 expression enhances the susceptibility of plants to nematodes suggesting that it is involved in the recognition of nematode-associated patterns. We propose that NILR1 is a PRR (or a component of a PRR complex) that recognises a NAMP leading to the activation of PTI responses. This hypothesis is supported by experiments showing that nilr1-1 is defective in the ROS burst as well as in seedling growth inhibition upon NemaWater treatment compared with Col-0. Notably, nilr1-1 and nilr1-2 did not respond differently to flg22 as compared with Col-0. On the other hand, bak1-5 was defective in PTI activation in response to both flg22 and NemaWater indicating a BAK1-mediated role for NILR1 in nematode recognition. In comparison to nilr1 (nilr1-1, nilr1-2), nilr2-1 did not show any change in susceptibility to neither cyst nor to root-knot nematodes compared to Col-0. Similarly, there was no change in seedling growth inhibition as compared with Col-0. Nevertheless, activation of ROS burst upon NemaWater treatment was decreased in nilr2-1 as compared with Col-0. This seemingly contradictory observation raises the question as to whether NILR2 also plays a role in perception of nematodes. A possible explanation could be that knocking out NILR2 may alter receptor complex formation and function, which selectively influence downstream signalling pathways without substantially influencing plant susceptibility to nematodes. This hypothesis also predicts that distinct signalling pathways that are activated during nematode perception may lead to diverse signalling outputs independently from each other. In fact, a recent study suggests activation of BAK1-dependent and BAK1-independent PTI pathways in response to RKN infection [19].
In conclusion, the identification of NILR1 as an LRR-RLK required for NemaWater-induced immune responses and basal resistance to nematodes is a major step forward in understanding of the molecular mechanisms underlying plant-nematode interactions. Moreover, the wide distribution of NILR1 among monocot and dicot plants is different from the majority of currently known PRRs and provides a unique opportunity for manipulation. However, sequence similarity does not necessarily indicate similar functions. It is therefore plausible that some of these homologues represent BRI1 or similar receptors and appeared in our analysis due to close similarity between NILR1 and BRI1. In fact, absence of a consistent ROS burst in rice plants upon NemaWater treatment hints that rice plants may not encode a functional NILR1. However, it is also possible that production of ROS burst upon treatment with NemaWater in some plant species such as rice requires further optimisation. A more detailed study would be needed to investigate this aspect.
Future work will focus on the purification and identification of elicitor/s present in NemaWater that are recognised in an NILR1-dependent manner. Further, conservation and function of NILR1 in various crop plants will be investigated. This will not only help in increasing our understanding of induced immune responses, but also provide potential opportunities to breed or engineer durable resistance against nematodes.
Arabidopsis thaliana seeds were sterilized with 0.6% sodium hypochlorite and grown in Petri dishes containing agar medium supplemented with modified Knop’s nutrient medium under the previously described conditions [41, 42]. The infection assays with cyst nematodes were performed as previously described [41]. Briefly, 60–70 J2s of H. schachtii were inoculated to the surface of an agar Knop medium containing 12-days-old plants under sterile conditions. For each experiment, 15–20 plants were used per genotype and experiments were repeated at least three times independently. The number of females per plant was counted at 14 days after inoculation (dai). For each experiment, 15–20 plants were used per genotype, and experiments were repeated at least three times independently.
For infection assays with root-knot nematodes, approximately 100 J2s of M. incognita were inoculated to the surface of agar MS-Gelrite medium containing 12-day-old plants and number of galls was counted at 21 dpi. M. incognita was propagated on greenhouse cultures of tomato (Solanum lycopersicum cv. Moneymaker) plants. Galls on roots of tomato were cut into smaller pieces of approximately 1 cm, crushed, and incubated for 3 min in 1.5% NaOCl2. Subsequently, the suspension was passed through a series of sieves to separate nematode eggs from root pieces. Eggs were collected in a 25 μm sieve. For surface sterilisation, eggs were incubated in a 10% NaOCl2 for 3 minutes and washed with abundant sterile water. The clean egg suspension was further washed with 150 μL Nystatin (10,000 U/ mL) and 2mL gentamycin sulphate (22.5 mg/mL) in a total volume of 30 mL. The suspension was stored at RT in darkness. Freshly hatched J2s were rinsed in water, incubated for 20 minutes in 0.5% (w/v) streptomycin-penicillin and 0.1% (w/v) ampicillin-gentamycin solution and for 3 minutes in 0.1% (v/v) chlorhexidine and washed three times with liberal amounts of sterile autoclaved water. For each experiment, 15–20 plants were used per genotype, and experiments were repeated at least three times independently.
Ten hours after inoculation with H. schachtii, small root segments containing nematodes with moving stylets were marked under the binocular. Movement of stylet indicates the migration phase of nematodes. The infected area around nematode head was then dissected. Corresponding root segments from uninfected plants were used as a control. RNA was extracted using a Nucleospin RNA extraction kit (Macherey-Nagel, Durren, Germany) according to the manufacturer’s instructions. The quality and quantity of RNA was analysed using an Agilent Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA) and a Nanodrop (Thermo Fisher Scientific, Waltham, MA, USA) respectively. The cDNA synthesis was performed with NuGEN’s Applause 3’Amp System (NuGEN, San Carlos, CA, USA) according to the manufacturers’ instructions. NuGEN’s Encore Biotin Module (NuGEN) was used to fragment cDNA. Hybridization, washing and scanning were performed according to the Affymetrix 30 GeneChip Expression Analysis Technical Manual (Affymetrix, Santa Clara, CA, USA). Three chips each were hybridized with control and infected samples, with each microarray representing an independent biological replicate. The primary data analysis was performed with the Affymetrix Expression Console v1 software using the MAS5 algorithm.
Approximately 300 brown cysts were collected from nematode stock culture, which was maintained on mustard roots under sterile conditions. These cysts were incubated in 3 mM ZnCl2 in funnels (hatching chambers) to induce hatching. Before collection of J2s, the hatching chamber was checked for microbial contamination. After seven days, J2s were collected in a falcon tube containing double distilled autoclave water. The mixture of nematode in ZnCl2 was spinned at 800 rpm for 3 min and supernatant was discarded. Afterwards, 1 ml of 0.05% HgCl2 was added and nematodes were incubated in it for 3 min to surface-sterilize them. HgCl2 was then removed and autoclaved double distilled water was added in excess (approximately 30 ml). The J2s were left in water for three min to wash them and remove HgCl2. After 3 min, nematodes were spinned down at 800 rpm for 3min and the entire washing step was repeated three times.
Approximately 40,000 sterile J2s of H. schachtii were incubated in 2 ml dd H2O for 24 hours at room temperature with continuous shaking. Afterwards, the nematode-water mixture was briefly centrifuged at 800 rpm for 2 minutes. The supernatant was removed to a new Eppendorf tube and was labelled as NemaWater. All steps of NemaWater production were performed under sterile conditions. Twelve-days-old Arabidopsis plants grown in Knop medium, as described above, were removed from agar plates and incubated in NemaWater for one hour each. Whole roots from 10 plants were cut and frozen in liquid nitrogen. Arabidopsis roots treated only with dd H2O were used as a control. Three biological replicates were performed. RNA was extracted, amplified and hybridised to perform a microarray analysis, as described above. Three chips for each were hybridised for a control and for NemaWater treated samples, with each microarray representing an independent biological replicate.
Affymetrix.CDF and.CEL files were loaded into the Windows GUI program RMAExpress (http://rmaexpress.bmbolstad.com/) for background correction, normalisation (quantile) and summarisation (median polish). After normalisation, the computed robust multichip average (RMA) expression values were exported as a log scale to a text file. Probe set annotations were performed by downloading Affymetrix mapping files matching array element identifiers to AGI loci from ARBC (http://www.arabidopsis.org). All genes that were more than 1.5 fold differentially regulated (t-test; P < 0.05) were pre-selected for further analysis using False discover rate at 5%.
To validate the microarray expression data, 11 up- and two down-regulated genes were randomly selected. The samples were collected in the same manner as the microarrays analysis for NemaWater. RNA was extracted using a Nucleospin RNA Xs (Macherey- Nagel, Germany) kit according to the manufacturer’s instructions. cDNA was synthesized using a High Capacity cDNA Reverse Transcription Kit (Life technologies cat.no. 4368814), according to the manufacturer’s instructions. The transcript abundance of targeted genes was analysed using the Stepone Plus Real-Time PCR System (Applied Biosystems, USA). Each sample contained 10 μL of Fast SYBR Green qPCR Master Mix with uracil-DNA, glycosylase, and 6-carboxy-x-rhodamine (Invitrogen), 2 mM MgCl2, 0.5 μL of forward and 0.5 μL of reverse primers (10 μM), 2 μL of complementary DNA (cDNA) and water in 20 μL of total reaction volume. Samples were analysed in three technical replicates. To serve as an internal control, 18S genes were used. Relative expression was calculated as described previously [43], by which the expression of the target gene was normalized to 18S to calculate fold change. All primer sequences are listed in S6 Data.
Single T-DNA inserted knockout mutants for selected genes (AT1G74360: nilr1-1, SAIL_859_H01, nilr1-2, GK-179E06; AT1G53430: nilr2-1, SALK129312C) were ordered from relevant stock centre. The homozygosity of mutants was confirmed via PCR using primers given in S6 Data. The homozygous mutants were confirmed to be completely absent from expression through RT-PCR with primers given in S6 Data.
The production of an ROS burst was evaluated using a modified protocol adapted from previous work [27]. Small root segments (approx. 0.5 cm) were cut from 12-days-old plants and floated in ddH2O for 12 hours. Afterwards, the root segments were transferred to a well in a 96-well plate containing 15 μl of 20 μg/ml horseradish peroxidase and 35 μl of 0.01M 8-Amino-5-chloro-2,3-dihydro-7-phenyl-pyrido[3,4-d] pyridazine sodium salt (L-012, Wako Chemicals). Next, 50 μl of either 1 μM flg22 or NemaWater was added to the individual wells. The experiments were performed in four technical replicates, and ddH2O was used as a negative control. Light emission was measured as relative light units in a 96-well luminometer (Mithras LB 940; Berthold Technologies) over 120 minutes and analysed using instrument software and Microsoft Office Excel. For experiments with Proteinase K, 100 μl of Proteinase K was added to 1 ml of NemaWater or flg22, and the mixture was incubated at 37°C for 4 hours. For heat treatment, samples were incubated at 90°C for 30 min. ddH2O was used as a negative control. The experiments were performed in three technical replicates and independently repeated multiple times as indicated in figure legends.
Arabidopsis plants were grown in Knop medium, as described above. Five-days-old plants were transferred to a well in a 6-well plate containing a liquid MS medium supplemented with either 1 ml of 1 μM flg22 or NemaWater. ddH2O was used as a negative control. Fresh weight and length of the roots were measured 7 days after they were transferred to MS medium. The experiments were performed in three technical replicates and independently repeated multiple times as indicated in figure legends.
The amino acid sequence for ECD of NILR1 was used to blast against all land plants sequences resulting in 318 hits across kingdom. Representative sequences from 44 unique species were used to generate a multiple alignment file. A Gblock function was used to refine alignment, and a maximum-likelihood analysis was performed with the PHYML software [44]. A nonparametric approximate likelihood ratio test was used for branch support as an alternative to usual bootstrapping procedure [45].
ECD sequence of NILR1 was used to search the SWISS-MODEL template library (SMTL version 2016-03-23, PDB release 2016-03-18) with Blast and HHBlits for evolutionary related matching structures matching [46–48]. NILR1 match best with BRASSINOSTEROID INSENSITIVE 1 (BRI1) and the PDB file from SWISS-MODEL was used to view 3-dimensional structures with NCBI Cn3D [49].
Coding region of NILR1 was amplified without stop codon using gateway forward and reverse primers as given in S6 Data. The amplified fragment was cloned into pDONR207 using BP clonase (Invitrogen) according to manufacturer’s instructions. The resultant pENTRY vector (pENTRY/NILR1) was then used to clone NILR1 into the destination vector pMDC83:CGFP [50] using LR clonase (Invitrogen) according to manufacturer’s instructions. The expression vector (35S:NILR1-GFP) was transformed into Agrobacterium strain GV3101 and co-infiltrated together with a plasma membrane mCherry marker 35S:PIP2A-mCherry [51] into epidermis of 6-week old Nicotianna benthamiana leaves [52]. The GFP and mCherry signal was detected using a confocal microscope (Zeiss CLSM 710).
|
10.1371/journal.pntd.0001395 | Ebola Virus Glycoprotein Needs an Additional Trigger, beyond Proteolytic Priming for Membrane Fusion | Ebolavirus belongs to the family filoviridae and causes severe hemorrhagic fever in humans with 50–90% lethality. Detailed understanding of how the viruses attach to and enter new host cells is critical to development of medical interventions. The virus displays a trimeric glycoprotein (GP1,2) on its surface that is solely responsible for membrane attachment, virus internalization and fusion. GP1,2 is expressed as a single peptide and is cleaved by furin in the host cells to yield two disulphide-linked fragments termed GP1 and GP2 that remain associated in a GP1,2 trimeric, viral surface spike. After entry into host endosomes, GP1,2 is enzymatically cleaved by endosomal cathepsins B and L, a necessary step in infection. However, the functional effects of the cleavage on the glycoprotein are unknown.
We demonstrate by antibody binding and Hydrogen-Deuterium Exchange Mass Spectrometry (DXMS) of glycoproteins from two different ebolaviruses that although enzymatic priming of GP1,2 is required for fusion, the priming itself does not initiate the required conformational changes in the ectodomain of GP1,2. Further, ELISA binding data of primed GP1,2 to conformational antibody KZ52 suggests that the low pH inside the endosomes also does not trigger dissociation of GP1 from GP2 to effect membrane fusion.
The results reveal that the ebolavirus GP1,2 ectodomain remains in the prefusion conformation upon enzymatic cleavage in low pH and removal of the glycan cap. The results also suggest that an additional endosomal trigger is necessary to induce the conformational changes in GP1,2 and effect fusion. Identification of this trigger will provide further mechanistic insights into ebolavirus infection.
| Ebolavirus causes often fatal hemorrhagic fever in humans and nonhuman primates. During infection, the virus is internalized into the low pH endosomes prior to the delivery of viral RNA to the infected cell. Cleavage by endosomal cathepsins of the heavily glycosylated mucin-like domain and glycan cap from the ebolavirus surface glycoprotein GP1,2 is an essential step in infection. The effect of cleavage and the low pH of the endosomes on the conformation of GP1,2 is as yet unknown. To investigate the effect of priming, we cleaved the mucin-like domain and glycan cap of Zaire ebolavirus (ZEBOV) GP1,2 with thermolysin and engineered a mutant of Sudan ebolavirus (SEBOV) GP1,2 that is cleaved with furin. We demonstrate by DXMS and antibody binding studies that cleavage of the mucin domain and glycan cap and incubation at low pH are insufficient to trigger the conformational changes of GP1,2 that effect fusion. Unraveling the trigger that leads to the conformational change of GP1,2 to its fusogenic form will enhance the understanding of ebolavirus infection and pinpoint key sites for therapeutic intervention.
| Ebolaviruses cause severe hemorrhagic fever in humans and non-human primates with 50–90% lethality. No specific vaccines or treatments for ebolavirus infection have yet approved for human use [1]–[5]. Among the five different members of the ebolavirus genus, Zaire ebolavirus (ZEBOV) and Sudan ebolavirus (SEBOV) are the most lethal and are the most commonly associated with outbreaks among humans [6]. The virus displays a trimeric glycoprotein (a class I fusion protein) on its surface, termed GP1,2, which is solely responsible for attachment and internalization of the virus [7], [8]. The glycoprotein is initially expressed as a single polypeptide that is then cleaved by furin in the producer cell to yield two disulphide-linked subunits termed GP1 and GP2 [9]. Of these, GP1 attaches to target cells while GP2 drives the fusion of viral and host cell membranes for the delivery of viral RNA into the host cells. The GP1,2 trimer is extended by heavily glycosylated mucin-like and “glycan cap” regions that are attached to the top of GP1 by a single polypeptide and reach upwards and outwards toward the target cell. The extensive glycocalyx provided by these domains and other glycans on GP1 and GP2 may shield the complex from immune surveillance and/or play an additional role in the natural host reservoir [10].
Numerous studies have revealed that the 450 kDa trimeric GP1,2 is further proteolytically cleaved, after entry into target cells, by the endosomal cathepsins L and B [11]–[14]. This trimming operates on the loop formed by residues 190–213 in GP1 and yields a ∼39 kDa fragment containing the N-terminal portion of GP1 (prior to the cleavage site) and all of GP2. Cleavage is thought to expose the receptor-binding region (RBR) on the remaining GP1 core and enhance fusion of viral and cellular membranes [11], [13], [15]. Cathepsins L and B cleave at slightly different sites. Cathepsin L cleaves at residue 201 [15], [16] and is sufficient to remove the mucin-like domain and glycan cap. Cathepsin B deletes additional residues N-terminal to the site of cathepsin L cleavage, removing an additional ∼1 kDa of mass from GP1 [11], [15].
Cleavage by the combination of cathepsins L and B can be functionally mimicked by thermolysin [12], [15]. Thermolysin cleaves GP around residue 190, leaving residues 33–190 of GP1 and all of GP2 [15] that together assemble a ∼39 kDa GP1,2 core. Although it is known that these cleavage events are usually required for infection, the structural manifestation of enzymatic cleavage is as yet unclear.
Here we demonstrate by antibody binding and peptide amide hydrogen-deuterium exchange mass spectrometry (DXMS) that priming of the GP1,2 ectodomain and endosomal pH themselves are insufficient for triggering the conformational changes necessary for fusion, and that an additional trigger must be required in the infected cell.
The design of a construct amenable to high-level expression of Zaire ebolavirus GP1,2 (ZEBOV-GP1,2) and KZ52 is described previously [17], [18]. Briefly, ZEBOV-GP1,2 has an N-terminal HA tag and comprises residues 33–632 with the mucin and transmembrane regions (residues 313–465 and 633–676) deleted, and a T230V mutation that significantly improves expression yields. The protein was transiently expressed in HEK293T cells and purified using an anti-HA column followed by size exclusion chromatography. ZEBOV-GP1,2 is cleaved by thermolysin (75 µg/ml) overnight and the cleaved protein (ZEBOV-GP1,2CL) is separated from released fragments by size exclusion chromatography using a Superdex-200 10/300 GL column equilibrated in PBS buffer. To obtain a complex between ZEBOV-GP1,2CL and KZ52, the proteins were mixed in a ratio of 1∶1.5 by weight respectively and incubated on ice for 2 h. The complex was separated from excess KZ52 by size exclusion chromatography (Suppl. Figure S1A).
A construct for Sudan ebolavirus GP1,2 (SEBOV-GP1,2; strain Gulu) was cloned into the pDISPLAY vector and contains a N-terminal HA tag for purification. The construct comprises residues 33–637 with the mucin and transmembrane regions (residues 314–472 and 638–676) deleted. SEBOV-GP1,2 is more sensitive to cathepsin and thermolysin cleavage and it is difficult to obtain sufficient yields of cleaved, wild-type SEBOV-GP1,2 for DXMS studies. Hence, for production of homogeneous SEBOV-GP1,2CL, a furin cleavage site was introduced with the mutation 205RKKR208 into the 190–213 loop upon which cathepsins and thermolysin act (gift of Paul Bates, University of Pennsylvania). This protein is thus “born” cleaved twice by furin in the producer cell: one furin-cleavage site separates GP1 from GP2, while the second removes the glycan cap from GP1. This “born-cleaved” protein was transiently expressed in HEK293T cells supplemented with a plasmid encoding excess furin (in a glycoprotein-to-furin plasmid ratio of 4∶1 by weight). The expressed protein was purified using an anti-HA column. Further size exclusion chromatography illustrates it is the same size as cleaved wild-type GP1,2.
Binding of KZ52 to ZEBOV-GP1,2CL under various pH conditions was determined by ELISA. Briefly, aliquots of ZEBOV-GP1,2CL were buffer exchanged to a range of pH (pH 4.2, 4.6, 5.0, 5.4, 6.0, 6.6 and 7.0, each obtained from a 1 M stock solution pH screen (Hampton) using Millipore Ultramax 10 K cutoff centrifugal concentrators [17]), and incubated in 10 mM buffer and 150 mM NaCl at room temperature for 2 h. The proteins were subsequently neutralized in 100 mM Tris-HCl buffer pH 7.5 (note that any conformational changes in GP2 leading to fusion are irreversible). The changes in pH were monitored by litmus paper.
0.25 µg of the protein per well was adsorbed (∼50 µl/well) overnight at 4°C in a 96-well microtiter plate. The wells were washed with PBS containing 0.05% Tween (PBS-Tween) and blocked with 100 µl/well of 3% bovine serum albumin (BSA) in PBS for 1 h at room temperature. The wells were washed (with PBS-Tween) and incubated with 50 µl of the antibody between 0.1 and 1 µg/ml for 1 h, washed again, and incubated with HRP conjugate (1∶2000 dilution in 1% BSA and PBS). The wells were washed thoroughly with PBS-Tween, and bound HRP was detected by incubation with 50 µl/well of 1∶1 solution of TMB substrate (Pierce). The reaction was terminated by addition of 50 µl/well of 2 N H2SO4. The resulting solutions were analyzed by a UV/Vis spectrophotometer microtiter plate reader at 450 nm.
Prior to the exchange experiments, disulfide reduction and protein denaturation conditions were identified that produced an optimal pepsin fragmentation pattern under exchange-quench conditions. It was found that incubation with quench buffer with TCEP and denaturant, followed by dilution of denaturant prior to exposure to pepsin gave the best peptide probe coverage map. For ZEBOV, 1 µl of ZEBOV-GP1,2 or ZEBOV-GP1,2CL stock solution (at 5.3 mg/ml and 5.5 mg/ml respectively) was diluted into 1 µl of non-deuterated buffer (8.3 mM Tris, pH 7.5, 150 mM NaCl, in H2O) at 0°C. Next, 3 µl of 1 M TCEP, 6.4 M GuHCl, pH 2.85 (quench buffer) was added and the sample was incubated for 5 min on ice. Then 15 µl of 0.8% formic acid (FA), 16.6% glycerol (quench diluent) was added to the denatured sample prior to protease digestion.
For SEBOV, 1 µl of SEBOV-GP1,2 stock solution (6.3 mg/ml) was diluted with 7 µl of non-deuterated buffer (8.3 mM Tris, 150 mM NaCl, in H2O, pH 7.15) at 0°C and then quenched with 12 µl of quench buffer containing 3.2 M GuHCl, 15 mM TCEP, 0.8% formic acid, 16.6% glycerol (pH 2.4). For SEBOV-GP1,2CL sample, 2 µl of protein (5.1 mg/ml) stock solution was diluted into 3 µl of quench buffer (1 M TCEP, 6.4 M GuHCl, pH 2.85) and the sample was incubated for 5 min on ice. Then 15 µl of quench diluent (0.8% FA, 16.6% glycerol) was added. The samples were then frozen on dry ice and stored at −80°C for further analysis.
In order to demonstrate the ability of our methods to detect differences in deuteration of these proteins, GP1 and GP2 of SEBOV-GP1,2CL were separated from each other by adding 10 mM DTT and heating at 37°C for 30 min. The disulfide-reduced sample was prepared and analyzed in a similar fashion as SEBOV-GP1,2CL for DXMS experiments (Figures S4A–S4C in supplementary information).
All samples were subsequently thawed at 4°C and passed over an AL-20-pepsin column (16 µl bed volume, 30 mg/ml porcine pepsin (Sigma)), at a flow rate of 20 µl/min. The resulting peptides were collected on a C18 trap (Michrom MAGIC C18AQ 0.2×2) and separated using a C18 reversed phase column (Michrom MAGIC C18AQ 0.2×50 3 µm 200 Å) running a linear gradient of 0.046% (v/v) trifluoroacetic acid, 6.4% (v/v) acetonitrile to 0.03% (v/v) trifluoroacetic acid, 38.4% (v/v) acetonitrile over 30 min with column effluent directed into an LCQ mass spectrometer (Thermo-Finnigan LCQ Classic). Data were acquired in both data-dependent MS1∶MS2 mode and MS1 profile mode. SEQUEST software (Thermo Finnigan Inc.) was used to identify the sequence of the peptide ions. DXMS Explorer (Sierra Analytics Inc., Modesto, CA) was used for the analysis of the mass spectra as described previously [19].
Functional deuteration of ZEBOV glycoproteins were performed by diluting 1 µl of ZEBOV-GP1,2CL or ZEBOV-GP1,2 stock solution into 3 µl of D2O buffer (8.3 mM Tris, 150 mM NaCl, in D2O, pDREAD 7.2) at 0°C. At 10 sec and 1000 sec, 6 µl of ice-cold quench buffer (6.4 M GuHCl, 1 M TCEP, pH 2.85) was added, samples incubated on ice for 5 min, 10 µl quench diluent (0.8% FA, 16.6% glycerol) was added, and samples frozen on dry ice. The equilibrium-deuterated back-exchange control samples were prepared by diluting 1 µl of ZEBOV-GP1,2CL or ZEBOV-GP1,2 into 3 µl of 1% formic acid in 99.9% D2O with incubation overnight at room temperature, cooled to 0°C and mixed with 6 µl quench buffer, incubated for 5 min, supplemented with 10 µl quench diluent, and frozen and further processed as above.
Functional deuteration of SEBOV-GP1,2 was performed by first diluting 2 µl of SEBOV-GP1,2 with 2 µl of non-deuterated buffer (8.3 mM Tris, 150 mM NaCl, in H2O, pH 7.15). At 0°C, 4 µl of D2O buffer (8.3 mM Tris, 150 mM NaCl, in D2O, pDREAD 7.2) was added. At 10 sec and 1000 sec, 12 µl of ice-cold quench buffer (3.2 M GuHCl, 15 mM TCEP, 0.8% formic acid, 16.6% glycerol, pH 2.4) was added. For the SEBOV-GP1,2CL sample, 2 µl of SEBOV-GPCL (5.95 mg/ml) stock solution was diluted into 2 µl of D2O buffer (8.3 mM Tris, 150 mM NaCl, in D2O, pDREAD 7.2) at 0°C. At 10 sec and 1000 sec, 6 µl of ice-cold quench buffer (6.4 M GuHCl, 1 M TCEP, pH 2.85) was added, samples incubated on ice for 5 min, 15 µl quench diluent (0.8% FA, 16.6% glycerol) was added.
The centroids of the isotopic envelopes of nondeuterated, functionally deuterated, and fully deuterated peptides were measured using DXMS Explorer and then converted to corresponding deuteration levels with corrections for back-exchange [20].
All figures in the manuscript were generated using Pymol and Adobe photoshop.
The crystal structure of Zaire ebolavirus GP1,2 (ZEBOV-GP1,2) has been previously determined in its prefusion form, in complex with an antibody termed KZ52 that was derived from a human survivor of Zaire ebolavirus infection [17]. The structure reveals that antibody KZ52 binds a conformational epitope on GP1,2 where the GP1 and GP2 subunits meet, and illustrates the specificity of KZ52 for the prefusion state of ZEBOV-GP1,2. In addition, we have also determined the crystal structure of the prefusion form of Sudan ebolavirus GP1,2 (SEBOV-GP1,2) bound to a novel neutralizing antibody termed 16F6 [21]. 16F6, like KZ52, recognizes a conformational epitope on the prefusion form of SEBOV-GP1,2 that bridges the two subunits together. The loop of GP1,2 upon which both endosomal cathepsins and thermolysin act (residues 190–213) is mobile and disordered in both structures, and is located at the outer (“side”) surface of the trimeric GP1,2.
In these prefusion GP1,2 structures, each GP2 in the trimer wraps around its corresponding GP1 in a metastable state. The GP1s form a “clamp” on the metastable prefusion conformation of GP2 and prevent it from springing irreversibly into its fusion-active and more stable six-helix bundle conformation, which has also been crystallized [22], [23]. In order for fusion to occur, GP2 must unwrap from each GP1 in the prefusion complex and rearrange to yield the six-helix bundle conformation of GP2 alone.
To investigate the effects of enzymatic cleavage of GP1,2, we digested ZEBOV-GP1,2 with thermolysin and purified the resulting protein (ZEBOV-GP1,2CL) by size exclusion chromatography. Note that the ZEBOV-GP1,2 used for the studies has a deletion of the mucin-like domain to enhance protein expression. Thermolysin functionally mimics cleavage by endosomal cathepsins, but operates at physiological pH, while cathepsins require low pH (∼5.5) for cleavage. Use of thermolysin allows us to decouple the effects of enzymatic cleavage from the effects of low pH. The elution profile of thermolysin-cleaved ZEBOV-GP1,2CL suggests that ZEBOV-GP1,2 remains trimeric after cleavage (Suppl. Figure S1A).
SEBOV-GP1,2 is quite sensitive to proteolysis and cleavage by either cathepsins L/B or thermolysin results in significant degradation, making it difficult to purify sufficient material for DXMS studies. In order to analyze the effects of enzymatic cleavage of SEBOV-GP1,2, we engineered a version of SEBOV-GP1,2 with an additional furin site at residues 206–210 in the loop of GP1 that is operated on by cathepsins and thermolysin (the wild-type furin-cleavage site that separates GP1 and GP2 also remains in this construct). The furin-engineered SEBOV-GP1,2CL, is stable, is released from the expression host already deleted of its glycan cap and mucin-like domain and structurally mimics enzymatically cleaved SEBOV-GP1,2 (Suppl. Figure S1B). Like ZEBOV-GP1,2CL, SEBOV-GP1,2CL also exists as a trimer.
DXMS is able to measure the ability of deuterium in solvent water to exchange with hydrogens that are covalently bound to peptide amide nitrogen atoms in a protein [24]. To determine if cleavage of GP1,2 results in conformational change, we analyzed uncleaved ZEBOV/SEBOV-GP1,2 and cleaved ZEBOV/SEBOV-GP1,2CL proteins by DXMS, obtaining 49%/68% and 68%/83% coverage, respectively. Peptide regions throughout the glycoproteins were analyzed. Peptides covalently linked to glycans of heterogeneous mass are not accessible for measurement with currently employed DXMS methods, and therefore result in small gaps in the sequence coverage. Note that enzymatic cleavage removes nearly all N-linked glycans from GP1 and hence greater sequence coverage is obtained for cleaved GP (68%/83%) than uncleaved GP (49%/68%).
For ZEBOV-GP1,2 the peptide fragmentation analysis confirms that thermolysin cleaves after the aromatic residues in the disordered loop 190KKDFFSS196 and deletes the glycan cap (as observed by Dube et. al. [15]). The deuteration levels of ZEBOV-GP1,2 and SEBOV-GP1,2 (measured over a time of 10–1000 sec) are consistent with the respective crystal structures (PDB code 3CSY and 3S88) in that peptide fragments that are buried or that have amide hydrogen atoms involved in hydrogen bonding show low levels of deuteration.
Comparison of uncleaved GP1,2 with cleaved GP1,2CL, for both ZEBOV and SEBOV, reveals that no significant changes in deuteration occur upon cleavage for any measured peptide spanning the whole of GP1,2, observed over a 10–1000 sec time scale (Figure 1 and S2A–S3C in supplementary information). The deuteration pattern of the residues in the fusion loop (residues 520–540 of GP2) is also identical between the cleaved and uncleaved proteins (Figure 2). The unwinding of the fusion loop is a required early step in the springing of GP1,2CL to the post-fusion form. Alteration in the deuteration pattern of the fusion loop could indicate the beginning of unwinding of GP1,2CL, but is not observed after enzymatic cleavage.
In ZEBOV, residues R64, F88, K95, K114, K115, and K140 are critical for binding and are thought to comprise a part of the receptor-binding region [15], [25]. Peptides containing the residues R64, F88, K95, K114, and K115 are equally accessible by solvent before and after cleavage and removal of the glycan cap (Figure 3), indicating that no major conformational changes occur at these sites upon cleavage. Note that peptides containing residue K140 are not detected in DXMS of ZEBOV/SEBOV GP1,2 or GP1,2CL, and so it is not possible to compare deuteration levels of K140. Of this set of residues, F88, K114, K115 and K140 are inward of the glycan cap and solvent-accessible, and could potentially interact with the receptor. Residues R64 (K64 in SEBOV) and K95 are buried and probably have auxillary roles in receptor binding. Alternately, a conformational change could occur upon receptor binding that brings R64 and K95 into direct contact with the receptor.
DXMS thus suggests that no residues in ZEBOV/SEBOV-GP1,2CL dramatically change conformation when the mucin-like domain and glycan cap are released. These results also suggest that the glycan cap, itself, does not significantly occlude access by solvent to this site. The glycan cap, however, could block steric access by a protein molecule and receptor access would likely be enhanced by enzymatic removal of the glycan cap.
The epitope of the human neutralizing antibody KZ52 bridges GP1 and GP2 and KZ52 only binds when the subunits are assembled in their prefusion conformation. Conformational changes in GP2, such as those required for fusion, would likely abrogate KZ52 binding. Hence, binding studies of KZ52 to ZEBOV-GP1,2CL provide additional insights into the structure of ZEBOV- GP1,2CL. By both size exclusion chromatography and ELISA, we find that KZ52 binds well to cleaved, trimeric ZEBOV-GP1,2CL (See Figure 4 and Suppl. Figure S1A). An equivalent, GP1/GP2-bridging, prefusion-specific antibody, termed 16F6 [21], that recognizes SEBOV-GP1,2 also forms a stable complex with SEBOV-GP1,2CL. The binding of the prefusion-specific conformational antibodies KZ52 and 16F6 indicates that the GP1,2CL ectodomain from both viruses remains in its pre-fusion state upon thermolysin or furin cleavage of the 190–213 loop. In addition, successful binding of KZ52 to cathepsin L-cleaved ZEBOV-GP1,2CL with a KD of 1.5 nM has been recently reported by Hood et al. using surface plasmon resonance [16]. By contrast, Shedlock et al. [26] find that KZ52 does not neutralize cathepsin L-cleaved ZEBOV-GP1,2 that has been pseudotyped onto a viral surface (binding not directly measured in these studies). Hence, it seems that some as-yet-undetermined differences exist between ectodomain GP and viral-surface GP upon cathepsin L cleavage that do not occur upon thermolysin cleavage.
The endosomal compartments have an acidic pH ranging from ∼5.9–6.0 in the early endosome to ∼5.0–5.5 in the late endosome and the role of this low pH in triggering irreversible conformational changes leading to fusion has been speculated. To investigate the effect of pH on the conformation of ZEBOV-GP1,2CL, we monitored binding of KZ52 to acid pH-treated ZEBOV-GP1,2CL by ELISA (Figure 4). Bovine serum albumin (BSA), and reduced and denatured GP1,2 were used as negative controls. Indeed, KZ52 binding is unaffected by incubation of ZEBOV-GP1,2CL in endosomal pH, suggesting that pH alone does not cause rearrangement of ZEBOV-GP1,2CL from the pre-fusion state.
The DXMS and ELISA binding studies together suggest that the priming of GP1,2 of ebolaviruses and the low pH in which priming occurs, are themselves, insufficient for triggering the conformational changes required for fusion. An additional trigger such as binding of the receptor to cleaved GP or the action of another cellular factor thus appear to be essential for fusion (Figure 5).
The requirement of enzymatic cleavage for ebolavirus GPs may instead serve different, non-exclusive purposes. Cleavage might simply expose the receptor-binding site for binding to an endosomal receptor. Further, removal of the glycan cap and the heavily glycosylated mucin domain (∼75 kDa of protein and carbohydrate) could facilitate membrane fusion by reducing steric barriers to GP2 rearrangement and membrane association. Alternatively, cleavage of the residue 190–213 loop that covers the outside of the fusion loop may remove a flexible tether that anchors the fusion loop in place on the outside of the prefusion trimer. The enzymatic cleavage step of ebolavirus GP1,2 could indeed be required for one or all of these reasons, but the specific trigger of ebolavirus fusion remains to be identified.
|
10.1371/journal.pntd.0001411 | Targeting the Wolbachia Cell Division Protein FtsZ as a New Approach for Antifilarial Therapy | The use of antibiotics targeting the obligate bacterial endosymbiont Wolbachia of filarial parasites has been validated as an approach for controlling filarial infection in animals and humans. Availability of genomic sequences for the Wolbachia (wBm) present in the human filarial parasite Brugia malayi has enabled genome-wide searching for new potential drug targets. In the present study, we investigated the cell division machinery of wBm and determined that it possesses the essential cell division gene ftsZ which was expressed in all developmental stages of B. malayi examined. FtsZ is a GTPase thereby making the protein an attractive Wolbachia drug target. We described the molecular characterization and catalytic properties of Wolbachia FtsZ. We also demonstrated that the GTPase activity was inhibited by the natural product, berberine, and small molecule inhibitors identified from a high-throughput screen. Furthermore, berberine was also effective in reducing motility and reproduction in B. malayi parasites in vitro. Our results should facilitate the discovery of selective inhibitors of FtsZ as a novel anti-symbiotic approach for controlling filarial infection.
The nucleotide sequences reported in this paper are available in GenBank™ Data Bank under the accession number wAlB-FtsZ (JN616286).
| Filarial nematode parasites are responsible for a number of devastating diseases in humans and animals. These include lymphatic filariasis and onchocerciasis that afflict 150 million people in the tropics and threaten the health of over one billion. The parasites possess intracellular bacteria, Wolbachia, which are needed for worm survival. Clearance of these bacteria with certain antibiotics leads to parasite death. These findings have pioneered the approach of using antibiotics to treat and control filarial infections. In the present study, we have investigated the cell division process in Wolbachia for new drug target discovery. We have identified the essential cell division protein FtsZ, which has a GTPase activity, as an attractive Wolbachia drug target. We describe the molecular characterization and catalytic properties of the enzyme and demonstrate that the GTPase activity is inhibited by the natural product, berberine, and small molecule inhibitors identified from a high-throughput screen. We also found that berberine was effective in reducing motility and reproduction in B. malayi parasites in vitro. Our results should facilitate the discovery of selective inhibitors of FtsZ as a novel antibiotic approach for controlling filarial infection.
| Filarial nematode parasites are responsible for a number of devastating diseases in humans and animals. These include lymphatic filariasis and onchocerciasis that afflict 150 million people in the tropics and threaten the health of over one billion. Unlike other nematodes, the majority of filarial species are infected with an intracellular bacterium, Wolbachia [1]. In the human filarial nematode Brugia malayi, these obligate α-proteobacterial endosymbionts have been detected in all developmental stages [2]–[4]. Moreover, their presence is essential for the worm, as tetracycline-mediated clearance of bacteria from Brugia spp. leads to developmental arrest in immature stages and reduction in adult worm fertility and viability [5]–[10]. These findings have pioneered the approach of using antibiotics to treat and control filarial infections. However, in humans, tetracycline therapy is not ideally suited for widespread use because several weeks of treatment are required and the drug has contra-indications for certain individuals. Therefore, there is considerable interest in identifying new endosymbiont drug targets and other classes of compounds with anti-Wolbachia activity. Importantly, the completed genome sequence of the Wolbachia endosymbiont of B. malayi (wBm) [11] now enables genome-wide mining for new drug targets [11]–[14] and a foundation for rational drug design. These approaches should lead to the discovery of new classes of compounds with potent anti-Wolbachia/antifilarial activities targeting essential processes that are absent or substantially different in the mammalian host.
Bacterial cytokinesis has emerged as a major target for the design of novel antibacterial drugs [15]–[17] since several of the components that are essential for multiplication and viability are absent from mammals. The bacteria-specific “filamenting temperature sensitive” protein, FtsZ, plays a central role during bacterial cytokinesis. In Escherichia coli, temperature sensitive mutations in the ftsZ gene cause blockage in cell division with limited cell growth and the generation of long filaments. FtsZ assembles into the contractile Z-ring and coordinates more than a dozen other cell division proteins at the midcell site of the closing septum [18]–[21]. Formation of the septal Z-ring requires two important functional properties of FtsZ, namely, polymerization of the FtsZ monomers into protofilaments and GTPase activity. Since inhibition of either function is lethal to bacteria, both GTP-dependent polymerization [22]–[27] and enzymatic [27]–[28] activities of FtsZ have been targeted for the identification of new antibacterial agents. Several inhibitors have been discovered including synthetic compounds [17], [29] and natural products [17], [30]–[33].
In the present study, we identify the cell division machinery present in wBm and characterize the FtsZ protein (wBm-FtsZ). Using quantitative real time RT-PCR, Wolbachia ftsZ was found to be expressed throughout the life cycle, but up-regulated in fourth stage larvae and adult female worms. Recombinant wBm-FtsZ was shown to possess a robust GTPase activity, which was inhibited by the natural plant product berberine. Berberine was also effective in reducing motility and reproduction in B. malayi parasites in vitro. A library of small molecules was also examined for its inhibitory activity against the wBm and E. coli FtsZ proteins. Several compounds were identified as potent inhibitors, and structure-activity relationship studies revealed a derivative with selectivity for wBm-FtsZ. Thus, our results support the development of wBm-FtsZ as a promising new drug target in an anti-symbiotic approach for controlling filarial infection.
Living B. malayi adult female worms were purchased from TRS Laboratories, Athens GA. Genomic DNA and RNA were isolated following the protocols developed by Dr. Steven A. Williams (http://www.filariasiscenter.org/molecular-resources/protocols).
To clone full-length wBm-ftsZ for expression studies, forward 5′(GAGAGCTAGCATGTCAATTGACCTTAGTTTGCCAG)3′ (NheI site underlined) and reverse 5′(GAGACTCGAGTTACTTCTTTCTTCTTAAATAAGCTGG) 3′ (XhoI site underlined) primers were designed according to the wBm-ftsZ sequence (accession number: YP_198432) in order to amplify the gene from B. malayi genomic DNA. The PCR product was then cloned into the NheI and XhoI sites of pET28a(+) (Novagen) to generate a fusion protein with a His6 tag at the N terminus. The authenticity of the insert was verified by sequencing.
Total RNA supplied by the Filariasis Research Resource Center (FR3) was treated with RNase-free Dnase (New England Biolabs, Cat# M0303S) and purified using the RNeasy Kit from Qiagen. cDNA was obtained using random primers and the ProtoScript® AMV First Strand cDNA Synthesis Kit (New England Biolabs, Cat# E6550S). Forward primer 5′ (AACAAGAGAGGCAAGAGCTGGAGT) and reverse primer 5′(CGCACACCTTCAAAGCCAAATGGT) were utilized to amplify a 102 bp Wolbachia ftsZ amplicon. Wolbachia 16S rRNA amplified with forward primer 5′ (TGAGATGTTGGGTTAAGTCCCGCA) and reverse primer 5′(ATTGTAGCACGTGTGTAGCCCACT) was utilized for bacterial total RNA quantification. B. malayi 18S rRNA amplified with forward primer 5′ (ACTGGAGGAATCAGCGTGCTGTAA) and reverse primer 5′(TGTGTACAAAGGGCAGGGACGTAA) was utilized as a total worm RNA control. Quantitative PCR was performed using the DyNAmo™ HS SYBR® Green qPCR Kit (Thermo Fisher) and a CFX-96 Real Time PCR instrument (Bio-rad, Hercules, CA). Relative levels of ftsZ expression (ratio of ftsZ to 16S rRNA), and abundance of Wolbachia in B. malayi (ratio of Wolbachia 16S to B. malayi 18S rRNA) were calculated for each RNA sample. Experiments were performed twice with triplicate samples. Controls consisting of samples processed in the absence of reverse transcriptase were included in qPCR and no DNA contamination was detected.
To determine the sequence of the ftsZ gene from the Wolbachia endosymbiont wAlB present in the insect cell line Aa23 [34], multilocus sequence typing (MLST) ftsZ forward 5′ (TGTAAAACGACGGCCAGTATYATGGARCATATAAARGATAG) and reverse 5′ (CAGGAAACAGCTATGACCTCRAGYAATGGATTRGATAT) [35] primers were utilized to obtain a PCR fragment. Using BLAST analysis, the sequence of the PCR product was compared to the corresponding region of known full-length ftsZ sequences and their conserved downstream and upstream sequences and 6 additional primers 5′(TCTATTTTTAATTCTTTTAGAGAAGCATT), 5′(CGTTCGGTTTTGAAGGTGTGC), 5′ (ACCGTTGTGGGAGTGGGTGGT), 5′ (TTATTTTTTTCTTCTTAAATAAGCTGGTATATC), 5′ (GGAATGACAATAAGTGTATCTACGTA), and 5′(TGCATTTGCAGTTGCTCATCC) were designed to obtain a complete wAa-ftsZ sequence. Phusion® High-Fidelity DNA Polymerase (New England Biolabs, M0530) was utilized for all PCR reactions according to manufacturer's instructions.
wBm-ftsZ and E. coli ftsZ (Ec-ftsZ) were amplified using genomic DNA isolated from B. malayi and E. coli wild-type strain MG1655 respectively, and were then cloned into the pET28a plasmid to generate fusion proteins with a N-terminal His tag. Each protein was expressed in the Escherichia coli strain C2566 (New England Biolabs). Optimum conditions for production of soluble recombinant wBm-FtsZ involved co-transformation with the pRIL plasmid isolated from BL21-CodonPlus (DE3) cells (Stratagene) together with the pET28a-ftsZ plasmid. Cultures were grown at 37°C till the OD600 reached 0.6, before induction with 0.1 mM IPTG overnight at 16°C. Both Ec-FtsZ and wBm-FtsZ were purified using a similar method. The cells expressing the recombinant proteins were suspended in lysis buffer (20 mM NaPO4, 500 mM NaCl, 10 mM imidazole, pH 7.4) plus 1 mg/mL lysozyme and protease inhibitor cocktail (Roche) and incubated on ice for 30 min, followed by sonication. The lysate was then cleared by centrifugation at 12,500 rpm, 4 °C for 30 min. The His-tagged proteins were purified on a 5 mL HiTrap chelating HP column (GE Healthcare) using an AKTA FPLC following manufacturer's instructions. After application of the sample, the column was washed with 5 column volumes of buffer A (20 mM NaPO4, 500 mM NaCl, 10 mM imidazole, pH 7.4) followed by 10 column volumes of 92% buffer A:8% buffer B (20 mM NaPO4, 500 mM NaCl, 400 mM imidazole, pH 7.4). Protein was then eluted using a linear gradient (8–100%) of buffer B equivalent to 40–400 mM imidazole.
Fractions containing wBm-FtsZ or Ec-FtsZ were pooled, dialyzed against dialysis buffer (40 mM Tris-HCl, 200 mM NaCl and 50% glycerol, pH 7.5) and stored at −20°C prior to use. Purity of the proteins was estimated by 4–20% SDS-PAGE and the protein concentration was determined using the Bradford assay.
GTPase activity was measured using an enzyme-coupled assay [36]. Activity was determined by measuring the consumption of NADH, which is monitored by absorbance at 340 nm. The amount of NADH oxidized to NAD corresponds to the amount of GDP produced in the reaction. Reactions were optimized for a 96-well format to enable compound screening. The 100 µL reaction mixture containing 50 mM MOPS (4-morpholinepropanesulfonic acid) pH 6.5, 50 mM KCl, 5 mM MgCl2,1 mM PEP, 500 mM NADH, 0.1% Tween-20, 20 units/mL of L-lactate dehydrogenase (Sigma L2518) and pyruvate kinase (Sigma P7768), 1 mM GTP and 5 mM FtsZ was distributed into 96-well plates. The plate was incubated at 30 °C for 45 min with data collected at 20 second intervals using a SpectraMax® Plus 384 (Molecular Devices) spectrophotometer. Control assays without FtsZ were performed to provide a baseline and with GDP to ensure the function of the coupling enzymes.
For inhibitor screening, 100 µL of reaction mixture was added to each well of a 96-well plate and 1 µL of compound dissolved in DMSO, or berberine sulfate (MP Biomedicals) in water, in varying concentrations were added. The reaction was initiated at 30 °C by adding 1 mM GTP. Experiments were performed in triplicate.
Living B. malayi adult female and male worms were washed extensively with RPMI1640 medium supplemented with 2 mM glutamine, 10% Fetal Calf Serum (Gibco) and 100 U/mL streptomycin, 100 mg/mL penicillin, 0.25 mg/mL amphotericin B (Sigma). Three worms of either gender were distributed into each well of a 6-well plate and incubated at 37 °C, 5% CO2. After overnight recovery, motility and microfilaria production were recorded. Worms were then transferred to a new well containing varying amounts of berberine sulfate dissolved in water, namely 40 µM, 20 µM, 10 µM and 5 µM. Control wells containing either no drug or 10 µM doxycycline, were also included. Culture media were replaced with fresh medium containing drug daily. Adult worm and microfilaria motility production were recorded daily as described [37]. Motility was scored as described [38] and expressed as % of motility relative to motility scored on day 0 of the experiment. Microfilaria production was counted in 10 µL of either diluted or concentrated culture medium using a hemocytometer. The results were presented as the number of microfilaria released in 1 mL of medium from each well on the indicated day. Each treatment was performed in triplicate and the experiment was repeated several times.
Berberine sulfate (MP Biomedicals) was added at a final concentration of 0–400 µM to growth medium containing E. coli ER1613 (acrA13 Δ(top-cysB)204 gyrB225 IN(rmD-rmE) mcrA) (New England Biolabs) and growth determined during 5 h or 20 h of incubation. For the 5 h evaluation, an overnight culture of E. coli ER1613 (acrA13 Δ(top-cysB)204 gyrB225 IN(rmD-rmE) mcrA) (New England Biolabs) was diluted 100-fold and 1 mL volumes were dispensed into a 48-well deep well plate (Axygen Scientific) containing various concentrations (0–400 µM) of berberine sulfate (10 µL of serial diluted berberine sulfate in water). The plate was then incubated at 30 °C with shaking. After 90 min of initial growth, bacterial growth was determined every 30 min for 5 h by monitoring absorption at 600 nm using a microtiter plate reader (Spectramax M5, Molecular Devices). Alternatively, an overnight culture of E. coli was diluted 1∶1000 fold and incubated with varying amounts of berberine sulfate for 20 h before growth was determined. All experiments were performed at least twice. Viability of berberine sulfate-treated (24 h) cells was evaluated by spotting 3 µL serial dilutions (10−2–10−7) of bacteria on a petri dish and incubation overnight at 30 °C.
Bacterial morphology was visualized using a Zeiss AxioVert 200 microscope and images were obtained using a 20× objective.
Reactions were carried out under a nitrogen atmosphere with dry, freshly distilled solvents under anhydrous conditions, unless otherwise noted. Yields refer to chromatographically and spectroscopically homogenous materials, unless otherwise stated. Reactions were monitored by thin-layer chromatography (TLC) carried out on 0.25-mm EMD silica gel plates (60F-254) using UV-light (254 nm). Flash chromatography separations were performed on Silicycle silica gel (40–63 mesh). Purity analyses were performed using HPLC (254 nm).
A stirring solution of aldehyde (1.0 equiv) in MeOH at 25°C was treated with carboxylic acid (2.0 equiv), amine (2.0 equiv) and isonitrile (2.0 equiv). The solution was heated to reflux, and stirred for 24 h. The solution was then cooled to 25°C and concentrated in vacuo. The crude residue was purified via flash column chromatography (10–50% EtOAc in hexanes) to afford the purified product. For characterization data, see references [39]–[40].
The bacterial cell-division pathway has been extensively studied in E. coli and several essential proteins have been identified [17], [19]. Many of the genes encoding putative orthologs of these proteins are also present in wBm (Table 1). A total of 18 major cell division genes were identified in wBm genome (Table 1), including ftsZ, ftsA, ftsI, ftsK, ftsQ and ftsW, which are known to be essential for cell division [17]. These wBm genes were mapped and found to be more scattered throughout the genome, in comparison with their E. coli homologs. In E. coli the majority of genes were found in one major operon, with the remaining 5 genes distributed randomly. Of these, FtsZ was one of the most highly conserved essential proteins possessing 43% identity to Ec-FtsZ (Table 1). Wolbachia ftsA, ftsI, ftsK, ftsQ and ftsW were less related (13–34%) to the E. coli homologs. Some previously described essential cell division genes in E. coli (including ftsB, ftsL, ftsN and ZipA) were not found in wBm, indicating that there are differences in the cell division machinery present in free living E. coli and intracellular Wolbachia.
wBm-ftsZ exists as a single gene on the chromosome and is 1182 bp in length. It encodes a 394-amino acid protein with a predicted molecular mass of 42 kDa containing four distinct domains characteristic of FtsZ proteins. These comprise the variable N-terminal domain, a highly conserved core region, variable spacer, and a C-terminal conserved domain. The core region contains the highly conserved catalytic aspartate residue [41]–[42] and the GGGTGTGA motif (8 residues see [41], [43]), which are responsible for GTP hydrolysis and required for polymerization of the protein. The C-terminal region is not required for assembly, but is essential for interactions with the cell division proteins FtsA, FtsW and ZipA [17]. A similar organization was also found in the insect Wolbachia, wMel-FtsZ (NP_966481) and wAlB-FtsZ (JN616286). The FtsZ proteins of Wolbachia from different hosts share 89–91% identity and 43% identity to E. coli FtsZ proteins, with a substantially lower level at the carboxyl-terminal region (17.2% identity).
Wolbachia have been identified in all developmental stages of B. malayi, from studies on individual worms and isolates from regions endemic for lymphatic filariasis [2]–[4]. To determine the relative expression of wBm-FtsZ throughout the parasite life cycle and validate its suitability as a drug target, wBm-ftsZ mRNA expression was analyzed by quantitative real-time reverse transcription polymerase chain reaction (qRT-PCR). Relative levels of ftsZ expression (ratio of Wolbachia ftsZ to 16S rRNA) and abundance of Wolbachia in B. malayi (ratio of Wolbachia 16S to B. malayi 18S rRNA) were calculated for each RNA sample.
wBm-ftsZ was found to be expressed throughout all stages examined (adult female and male worms, microfilariae, third- and fourth-stage larvae). Moreover, wBm-ftsZ/16S ratios were found to be increased substantially following infection of the mammalian host since levels were significantly higher (p value<0.001) in fourth-stage larvae and adult female worms compared to the vector-derived infective third-stage larvae. The wBm-ftsZ/16S ratio was also higher in microfilariae compared with the vector-derived third-stage larvae, but was significantly lower than the ratios obtained for fourth-stage and adult female worms. Of the various developmental stages examined, the lowest level of wBm-ftsZ expression was found in male worms (Figure 1A). No DNA contamination was detected in controls consisting of samples processed in the absence of reverse transcriptase. Wolbachia 16S rRNA/B. malayi 18S rRNA ratios were also determined to measure the relative abundance of bacteria in different stages of B. malayi (Figure 1B). Wolbachia was found to be most abundant in fourth stage larvae and adult female worms and least abundant in infective third stage larvae, indicating a massive multiplication of Wolbachia soon after infection of the mammalian host. Taken together, these data indicate that while wBm-ftsZ is expressed in all stages, gene activity and bacterial multiplication is most pronounced in fourth-stage larvae and adult females.
Recombinant wBm-FtsZ was expressed in E. coli with a His-tag at the C-terminus and purified by nickel-affinity chromatography (Figure 2A). Optimum conditions for production of soluble recombinant wBm-FtsZ involved growth of cultures at 37°C until the OD600 reached 0.6, followed by induction with 0.1 mM IPTG overnight at 16°C. Purified protein was eluted with 100 mM imidazole. The apparent molecular weight of 43 kDa (Figure 2A) was consistent with the predicted molecular size of wBm-FtsZ with an N-terminal His-tag. For comparative studies, E. coli FtsZ (41 kDa) was also expressed and purified in a similar manner (Figure 2B).
GTPase activity was measured using an enzyme-coupled assay involving pyruvate kinase and lactate dehydrogenase [36]. GTP hydrolysis was determined by measuring the decrease in fluorescence emission following oxidation of nicotinamide adenine dinucleotide (NADH) to NAD (Figure 3A). As Figure 3B shows, recombinant wBm-FtsZ was found to possess GTPase activity. Moreover, the specific activities for wBm-FtsZ and Ec-FtsZ were comparable (0.18±0.012 µmolµmin−1mg−1 and 0.22±0.015 µmol min−1mg−1, respectively).
Berberine, an alkaloid natural product, is a known inhibitor of the GTPase activity of FtsZ in E. coli [33], [44]. Thus, we were interested in examining the generality of berberine's GTPase inhibitory activity against wBm-FtsZ. As Figure 4 shows, dose-dependent inhibition (25–1000 µM) was found with an IC50 value of 320 µM. E. coli FtsZ [33], [44] was included for comparison, and an IC50 value of 240 µM was observed (Figure 4). Since wBm-FtsZ possesses all but one of the key residues proposed in the binding of E. coli FtsZ to berberine (lysine instead of glycine at position 183 of Ec-FtsZ), this may account for the higher concentration of berberine required to inhibit 50% of wBm-Ftsz's GTPase activity.
Since filarial Wolbachia remain unculturable, we were unable to evaluate the direct effect of berberine on the endosymbiont. Therefore, we examined the indirect effect of the drug on adult female worm. As Figure 5A shows, berberine (10–40 µM) had adverse effects on the motility of adult female B. malayi worms, as well as microfilariae production (Figure 5B) when compared to untreated controls. Two days after treatment with berberine (40 µM), female worms showed almost no movement and the production of microfilaria had virtually ceased. Berberine at 20 µM was comparable to 10 µM of doxycycline in terms of effect on female worm motility. Reduction in adult female motility coincided with a decrease in microfilariae production. Similarly, motility of the freshly released microfilaria was decreased when berberine was present, with some effect observed at the lowest concentration (5 µM) tested (Figure 5C). On the other hand, male worms were more resistant to the effects of the drug with limited reduction in motility observed following treatment with berberine (5–40 µM) for 6 days (Figure 5D). However, treatment with 100 µM berberine for 24 h did completely paralyze male worms (data not shown). Doxycycline (10 µM) had a comparable affect on the motility of male and female worms.
To demonstrate that berberine's in vitro GTPase inhibitory activity and anti-parasitic activity correlates with its known antibacterial activity, studies were performed on E. coli strain ER1613. Berberine is known to act as a substrate for the multi-drug resistance efflux pumps and ER1613 contains a mutation in the acrA gene, which inactivates the multidrug efflux pump [45]. Overnight incubation of ER1613 with 0–100 µM berberine showed a dose-dependent effect with complete inhibition of bacterial growth observed at 60 µM (Figure 6A). Similarly, no growth was evident when experiments were initiated with greater bacterial densities and the cells were treated with 50 µM berberine for up to 5 h (Figure 6B). Treatment with berberine resulted in the filamentous phenotype (Figure 5C) typically observed in ftsZ mutant strains [46], indicating that berberine was inhibiting cell division. Moreover, the presence of elongated bacteria also correlated with decreased growth and viability. Viability was also evaluated by ability to form colonies on an agar plate. Berberine sulfate-treated (24 hours) cells produced substantially fewer colonies (Figure 6D), compared to untreated controls. Untreated bacteria had approximately 4×105 - fold growth in 24 h, whereas bacteria treated with 40 µM berberine had 4×102 - fold growth. At concentrations of 80 µM and higher, the treated bacteria failed to produce viable colonies (Figure 6D), demonstrating that without active replication E. coli die.
To initiate a campaign to identify molecularly unique inhibitors of wBm-FtsZ GTPase activity, a library of small molecules based on naphthalene, quinoline and biphenyl core scaffolds were examined [39]–[40] (Figure 7A). The library was constructed using Ugi multicomponent reaction chemistry, and each compound consists of a flat aromatic scaffold for enhanced π-stacking interactions decorated with varying diversity elements (R1–R4 in Figure 7A). Importantly, these scaffold motifs are also found in berberine (Figure 7B) and known FtsZ inhibitors [17], [29]–[33]. The ∼500-member library was screened using the wBm-FtsZ GTPase assay, and 13 compounds with greater than 30% inhibition at 100 µM were identified. From these screening efforts, compounds AV-C6 and N938 (Figure 7C) emerged as leading hits, and each showed dose-dependent inhibition of wBm-FtsZ (Figure 8A). AV-C6 and N938 were also examined for inhibition of the E. coli FtsZ enzyme (Figure 8A). As shown in Figure 8A, both compounds inhibited Ec-FtsZ activity although each was slightly less potent compared to the inhibitory activity against wBm-FtsZ.
Structure-activity relationship (SAR) studies were then performed on N938 as this compound showed the most potential in dose response experiments. In addition to identifying compounds with enhanced potency, we were also interested in exploring the possibility of tuning down any inhibitory activity against Ec-FtsZ in order to obtain a more specific Wolbachia FtsZ inhibitor. A series of analogues were synthesized with varying aromatic side chains (R3 in Figure 7A). As shown in Figure 8B, both goals were met: N982 with an ortho-chloro substituent (Figure 7D) showed enhanced potency in the wBm-FtsZ assay and N983 with a para-cyano substituent (Figure 7D) showed some specificity for wBm-FtsZ over that from E. coli. Future SAR studies should enable the discovery of compounds with both enhanced inhibitory properties and specificity. Finally, as the solubility of these compounds is poor, 100% inhibition of FtsZ with this scaffold was not possible and true IC50 values could not be obtained. Scaffold modification and/or hopping strategies will be investigated in the future to afford enhanced solubility.
The use of antibiotics targeting the Wolbachia endosymbionts of filarial parasites has been validated as an approach for controlling filarial infection in animals and humans. As a result, there is considerable interest in identifying new compounds that specifically target the obligate bacterial endosymbiont. In the present study, we investigated the cell division pathway in wBm to identify new drug targets that may be exploited for the development of new antifilarial therapies. Filamenting temperature sensitive (fts) genes produce many of the proteins essential for cell division in E. coli [17]. In wBm, we identified the majority of core genes that are indispensable to cytokinesis including ftsA, ftsI, ftsK, ftsQ, ftsW and ftsZ.
Interestingly, ftsB, ftsL, ftsN and ZipA were not found in wBm. ZipA is a bitopic membrane protein with a large cytoplasmic domain that binds and bundles FtsZ protofilaments in vitro and helps to stabilize the Z ring in vivo. FtsN is a core component of the divisome that accumulates at the septal ring at the initiation of the constriction process. The C-terminal SPOR domain specifically recognizes a transient form of septal murein, which helps trigger and sustain the constriction process. However, in E. coli, it has been found that alterations in FtsA can compensate for the absence of ZipA, FtsK [47] and FtsN [48] and a gain-of-function FtsA variant, FtsA*(R286W), efficiently stimulates cell division in the complete absence of ZipA [47]. Thus, Wolbachia FtsA may function like the mutant FtsA, as an alanine residue is present in the same position.
ftsB, ftsL, ftsN and ZipA are also absent in some important bacterial pathogens including certain Gram-negative (Neisseria spp., Bordetella pertussis, Helicobacter pylori, Chlamydia spp.) and Gram-positive (Mycobacterium tuberculosis) bacteria and cell wall-lacking (Mycoplasma pneumoniae) organisms [17]. It is likely that this reflects the reduced genome size present in these intracellular bacteria.
FtsZ is the most highly conserved essential bacterial cell division protein and is present in all bacteria except Chlamydia spp [17]. We determined that wBm-FtsZ shares substantial similarity (43% identity) to the highly characterized E. coli FtsZ protein and is highly similar (∼90% identity) to insect Wolbachia FtsZ proteins. While the majority of wBm genes are expressed in a stage-specific manner [49], wBm-ftsZ was found to be expressed in both male and female worms as well as in all larval stages examined. It was not surprising to find wBm-ftsZ expressed throughout the entire lifecycle of the parasite since the bacterial Z-ring is known to exist in a state of dynamic equilibrium in order to fulfill its many roles in the cell. Using fluorescence recovery after photo bleaching (FRAP), the E. coli Z-ring was found to continually remodel itself with a halftime of 30 seconds with only 30% of cellular FtsZ present in the ring with continuous and rapid exchange of subunits within a cytoplasmic pool [17]. E. coli ftsZ transcription analysis has revealed that the rate of ftsZ expression is constant with a sudden doubling at a specific cell age, suggesting that ftsZ expression is regulated [50]. Similarly, we observed up-regulation of wBm-ftsZ gene expression in fourth-stage larvae and adult female worms with microfilariae likely contributing to the increased expression in the latter case. While the lowest levels of gene expression were evident in adult males, FtsZ protein was easily detected in proteomic analyses of male worms [49]. In general, the gene expression pattern of ftsZ correlated with bacterial multiplication. The increased bacterial multiplication in the worm during early infection of the mammalian host and embryogenesis is in agreement with an earlier study [4]. These data are consistent with the third- and fourth-stage larval stages, and embryogenesis being particularly sensitive to the effects of antibiotic treatment [4], [51]. This result indicates that ftsZ gene expression could be used as a marker to monitor Wolbachia multiplication in the filarial parasite much like the ftsZ gene in the intracellular bacterium Candidatus Glomeribacter gigasporarum that resides in the mycorrhizal fungus Gigaspora margarita [43].
Molecular studies have established the importance of conserved amino acids in the FtsZ protein that when changed results in ftsZ mutants blocked at different stages of cell division [42], [46], [52]–[55]. wBm-FtsZ possesses the key residues and conserved GTP-binding pocket required for GTPase activity. Our functional analysis revealed that the GTPase activities of recombinant wBm-FtsZ and Ec-FtsZ are similar, and both proteins are sensitive to the plant alkaloid berberine. Most of the residues in Ec-FtsZ that are thought to bind berberine and inhibit FtsZ GTPase activity are also present in wBm-FtsZ. An earlier detailed study in E. coli determined that the target of this commonly used compound is FtsZ [33]. Plants containing berberine have been used in traditional Chinese and Native American medicine to treat many infectious diseases and the sulfate, hydrochloride and chloride forms are used in Western pharmaceutical medicine as antibacterial agents [56]. It is active against a number of Gram-positive and Gram-negative pathogenic bacteria, including drug resistant Mycobacterium tuberculosis [57] and Staphylococcus aureus [58].
Our experiments in E. coli demonstrate that berberine has both bacteriostatic and bacteriocidal effects. Since filarial Wolbachia remain unculturable, we were unable to evaluate the direct effect of berberine on the endosymbiont. However, following berberine treatment, we did observe reductions in adult female worm and microfilariae motility and microfilariae production. On the other hand, we did not see any effect on male worms, which had the lowest level of wBm-ftsZ gene expression. We examined berberine- and doxycycline-treated worms for Wolbachia load by qPCR analysis and did not observe a significant difference between control and treated parasites. A similar result was also found in a study evaluating the effects of globomycin and doxycycline on filarial Wolbachia, and the authors [59] suggested several possibilities which can also apply to our study, namely: the Wolbachia qPCR assay may not have sufficient sensitivity to detect effects on Wolbachia load over this time frame in nematodes, inhibition of FtsZ is sufficient to affect nematode motility and viability independent of or prior to any effect on Wolbachia load, and/or a direct effect of berberine on nematode motility and viability and alternative mechanisms of action. Nonetheless, our results suggest that FtsZ inhibitors that operate via inhibition of enzyme activity including natural products [28], [30]–[33], [53] and synthetic molecules [29], [60] may have also activity against wBm-FtsZ.
To complement the berberine studies, a library of naphthalene-, quinoline- and biphenyl-based compounds constructed using Ugi multicomponent reaction chemistry was examined for the discovery of new and ultimately highly specific antagonists of either E. coli or Wolbachia FtsZ. Of interest, compounds based on similar scaffolds have already been demonstrated as potent FtsZ inhibitors [17], [29]–[33]. From our screening efforts, the (6-{butylcarbamoyl-[(aryl)-(butylcarbonyl)-amino]-methyl})-naphthen-2-ol scaffold (Figure 7A, C) emerged as an antagonist of both E. coli and Wolbachia FtsZ. Interestingly, from basic SAR studies it appears that modification of the aryl substituent on the scaffold may afford selectivity for Wolbachia FtsZ, a key element of our initial goal. Additional compounds are currently being prepared to examine this possibility. Although not discussed here, compounds based on our lead scaffold had no effect on growth or viability in E. coli. Based on these findings and their potency in the in vitro assays, it is plausible that penetrability or metabolism issues are to blame for their attenuated activity. Finally, the solubility of these compounds is also poor precluding measurement of true IC50 values. Further iterations of chemical synthesis will be necessary to address these potential liabilities.
While we have focused on assaying the GTPase activity of wBm-FtsZ using a medium- to high-throughput coupled enzyme assay for the discovery of inhibitors that target cell division in Wolbachia, it is also possible to screen for compounds that would target wBm-FtsZ via other mechanisms of action. FtsZ is considered a distant functional relative of the mammalian cytoskeletal protein β-tubulin [61]–[63]. Microtubule formation is a major target in cancer chemotherapy and the anticancer drug Taxol binds to β-tubulin and blocks cell division by interfering with microtubule formation. Interestingly, the FtsZ inhibitor PC190723 [60] operates by a similar mechanism and more recently, novel inhibitors of B. subtilis cell division have been identified in an in vitro FtsZ protofilaments polymerization assay [64]. Importantly, significant differences exist in the active sites in tubulin and FtsZ polymers, and several small molecule inhibitors of FtsZ have been identified [65] that do not inhibit tubulin [66]–[67]. Tubulin is also the target of the broadly anti-parasitic benzimidazole drugs [68]–[69], which have been used extensively to control soil-transmitted nematodes [70]–[71].
FtsZ is also responsible for recruiting and coordinating more than a dozen other cell division proteins at the midcell site of the closing septum [18]–[19], [21], [72]. Many of these interactions are essential and it has been suggested that they might also be useful targets, particularly in light of developments in the discovery of small molecule inhibitors of protein-protein interactions [17], [73]–[74]. Therefore, it might be feasible to screen for inhibitors of the interactions between wBm-FtsZ and its various binding partners that modulate its polymerization. Another Wolbachia cell division protein worth considering for drug discovery is FtsA, as this protein also possesses enzymatic activity and contains an ATP-binding site that might be targeted with drug-like molecules. Moreover, this protein is essential in E. coli [75] and Streptococcus pneumoniae [76].
In summary, we have investigated the cell division pathway in wBm and determined that it possesses a FtsZ protein with GTPase activity. We demonstrated that the activity is inhibited by berberine and identified small molecule inhibitors in a high-throughput screen. Furthermore, berberine was found to have adverse affects on B. malayi adult worm and microfilariae motility, and reproduction. Our results support the discovery of selective inhibitors of Wolbachia FtsZ as a new therapeutic approach for filariasis.
|
10.1371/journal.pgen.1000727 | The Rts1 Regulatory Subunit of Protein Phosphatase 2A Is Required for Control of G1 Cyclin Transcription and Nutrient Modulation of Cell Size | The key molecular event that marks entry into the cell cycle is transcription of G1 cyclins, which bind and activate cyclin-dependent kinases. In yeast cells, initiation of G1 cyclin transcription is linked to achievement of a critical cell size, which contributes to cell-size homeostasis. The critical cell size is modulated by nutrients, such that cells growing in poor nutrients are smaller than cells growing in rich nutrients. Nutrient modulation of cell size does not work through known critical regulators of G1 cyclin transcription and is therefore thought to work through a distinct pathway. Here, we report that Rts1, a highly conserved regulatory subunit of protein phosphatase 2A (PP2A), is required for normal control of G1 cyclin transcription. Loss of Rts1 caused delayed initiation of bud growth and delayed and reduced accumulation of G1 cyclins. Expression of the G1 cyclin CLN2 from an inducible promoter rescued the delayed bud growth in rts1Δ cells, indicating that Rts1 acts at the level of transcription. Moreover, loss of Rts1 caused altered regulation of Swi6, a key component of the SBF transcription factor that controls G1 cyclin transcription. Epistasis analysis revealed that Rts1 does not work solely through several known critical upstream regulators of G1 cyclin transcription. Cells lacking Rts1 failed to undergo nutrient modulation of cell size. Together, these observations demonstrate that Rts1 is a key player in pathways that link nutrient availability, cell size, and G1 cyclin transcription. Since Rts1 is highly conserved, it may function in similar pathways in vertebrates.
| A critical point in the cell cycle occurs in G1 phase, when cells must decide whether to enter a new round of cell division. At this time, cells assess nutrient availability to ensure that they have sufficient resources to complete cell growth and division. Vertebrate cells also assess growth factors that control cell growth and determine when and where cell division occurs in the context of a multi-cellular organism. A cell-size checkpoint acts during G1 to delay entry into the cell cycle if the cell is below a critical size. When the appropriate signals have been received, cells commit to entry into the cell cycle by initiating transcription of G1 cyclins. The mechanisms that integrate external signals, cell growth, cell size, and entry into the cell cycle are poorly understood and represent a fundamental unsolved problem in cell biology. We discovered that a specific form of protein phosphatase 2A (PP2ARts1) functions in the pathways that integrate nutrient availability, cell size, and entry into the cell cycle. PP2ARts1 is highly conserved and may therefore carry out similar functions in all eukaryotic cells.
| Entry into the cell cycle is initiated by G1 cyclins, which bind and activate cyclin-dependent kinases [1]. There are two cyclin-dependent kinases in budding yeast that function during G1, called Cdk1 and Pho85, which are activated by numerous different G1 cyclins [1]. Cdk1 is activated by the cyclins Cln1, Cln2, and Cln3, while Pho85 is activated by Pcl1 and Pcl2, as well as by additional cyclins that do not appear to directly regulate G1 events. The G1 cyclins are redundant: cells lacking any two of the cyclins Cln1, Cln2 or Cln3 are viable, but loss of all three cyclins is lethal [2],[3]. Similarly, cells lacking Cln1 and Cln2 or Pcl1 and Pcl2 are viable, but loss of all four cyclins is lethal [4],[5]. The cyclin Cln3 plays a role in triggering transcription of a suite of genes required for initiation of G1 events, including the genes for Cln1, Cln2, and Pcl1, which are often referred to as late G1 cyclins [5]–[9]. Transcription of the late G1 cyclins is generally considered to be the key molecular event that marks entry into the cell cycle [10],[11]. The late G1 cyclins initiate growth of a new daughter bud and are also required for polar growth after bud emergence [4],[12].
Production of late G1 cyclins is tightly regulated. Cyclin mRNA and protein undergo rapid turnover, so mechanisms that act at the level of transcription play an important role [13]–[15]. Transcription of G1-specific genes, including the late G1 cyclin genes, is dependent upon the SBF and MBF transcription factors. SBF and MBF each include a distinct DNA binding subunit, called Swi4 and Mbp1, respectively, and a shared subunit called Swi6. SBF and MBF are kept inactive early in the cell cycle by a repressor called Whi5 [16],[17]. Loss of Whi5 causes transcription of late G1 cyclins to occur before the mother cell has completed growth, leading to premature bud emergence and a reduced cell size. Cdk1/Cln3 triggers transcription of late G1 cyclins by phosphorylating and inactivating Whi5. Transcription of late G1 cyclins can also be triggered by a redundant Cln3-independent pathway that is dependent upon the Bck2 protein [18]–[21]. The late G1 cyclin Cln2 promotes its own transcription via a positive feedback loop, which ensures that initiation of G1 events occurs in a coordinated, switch-like manner [6],[7],[22].
Mechanisms that control G1 cyclin transcription play an important role in control of cell size. A cell size checkpoint links initiation of G1 cyclin transcription to cell size. Thus, transcription of late G1 cyclins is only initiated when the mother cell has reached a critical size, which contributes to cell size homeostasis. An interesting property of cell size control in yeast is that the critical cell size is modulated by external nutrients, such that cells growing in poor nutrients are significantly smaller than cells growing in rich nutrients [23],[24]. It is thought that nutrients modulate cell size by rapidly changing the critical cell size for initiation of G1 cyclin transcription [11].
The mechanisms that link initiation of G1 cyclin transcription to cell size and nutrient availability are unknown. Interestingly, cln3Δ bck2Δ whi5Δ triple mutants, which lack all upstream regulators known to play an important role in the control of G1 cyclin transcription, undergo normal nutrient modulation of cell size [25]. Thus, the signals that control cell size by linking G1 cyclin transcription to nutrient availability must act by a different mechanism. The mechanisms that link G1 cyclin transcription to cell size and nutrient availability are likely to be a key to understanding cell size control.
Here, we report that a specific form of protein phosphatase 2A (PP2A) is required for control of G1 cyclin transcription and nutrient modulation of cell size. PP2A is a trimeric complex that consists of a catalytic “C” subunit, a scaffolding “A” subunit, and a regulatory “B” subunit [26]. Binding of different B-type regulatory subunits is thought to direct PP2A activity toward different substrates. Thus, the key to understanding PP2A is to understand the function and regulation of specific regulatory subunits. In budding yeast, two B subunits called Cdc55 and Rts1 bind to PP2A in a mutually exclusive manner, forming two distinct PP2A complexes: PP2ACdc55 and PP2ARts1 [27],[28].
We discovered a role for Rts1 in controlling G1 cyclin levels while characterizing a genetic interaction between RTS1 and the septin CDC12. The septins are a conserved family of proteins that localize to the site of bud emergence in early G1 and to the bud neck during bud growth and cytokinesis [29]. The septins have been proposed to form a diffusion barrier between the mother and daughter cell, to serve as a signaling scaffold for activation of kinases, or to carry out functions in the secretory pathway. Temperature sensitive alleles of the septins cause cells to undergo a prolonged delay at G2/M while undergoing continuous polarized growth, leading to the formation of highly elongated cells [30]. The G2/M arrest is mediated by Swe1, the budding yeast Wee1 homolog, which phosphorylates and inhibits Cdk1 to delay entry into mitosis [31]. The G2/M delay and the elongated cell phenotype are eliminated by swe1Δ. A number of kinases have been identified that regulate septin function and localization, and may in turn be regulated by the septins [31]–[36]. These kinases include Elm1, Gin4, Cla4, and Hsl1. Loss of these kinases can cause a Swe1-dependent G2/M delay and an elongated cell phenotype similar to septin mutants.
Previous work found that rts1Δ increased the restrictive temperature of the cdc12-6 allele [37]. Loss of RTS1 also caused altered septin ring dynamics; however, it remained unclear whether the observed changes in septin ring dynamics were sufficient to explain the rescue of the cdc12-6 temperature sensitive phenotype. For example, it was possible that in addition to regulating septin ring dynamics, Rts1 played additional roles in pathways that promote polar growth or Swe1-dependent G2/M delays. We therefore further investigated the role of Rts1 in polar cell growth and cell cycle progression.
Since rts1Δ suppressed the temperature sensitivity of cdc12-6, we tested whether rts1Δ also suppressed the elongated cell phenotype of these cells. We found that rts1Δ cdc12-6 cells showed reduced elongation compared with cdc12-6 cells when shifted to 30°C (Figure 1A). In addition, rts1Δ significantly reduced the elongated cell phenotype caused by loss of GIN4, ELM1, and CLA4 (Figure 1B).
We next considered the possibility that rts1Δ rescued the elongated cell phenotype of these mutants by eliminating the Swe1-dependent G2/M delay. To test this, levels of the mitotic cyclin Clb2 were assayed by Western blotting during a synchronized cell cycle in wild type, elm1Δ, and elm1Δ rts1Δ cells (Figure 1C). As previously shown, elm1Δ cells underwent a prolonged G2/M delay with elevated Clb2 levels when compared to wild type cells [35]. The prolonged G2/M delay was not eliminated by rts1Δ. Thus, although rts1Δ reduced the elongated cell phenotype caused by loss of CDC12, GIN4, CLA4, and ELM1, it did not appear to do so by reducing the Swe1-dependent G2/M delay.
Since rts1Δ did not rescue the G2/M delay in elm1Δ cells, we considered the possibility that Rts1 plays a direct role in promoting polar growth. To test this, we utilized cells that over express SWE1 from the GAL1 promoter, which arrest at G2/M with high levels of G1 cyclins and undergo constitutive polar growth [12],[38]. Wild type or rts1Δ cells carrying GAL1-SWE1 were released from a G1 arrest in the presence of galactose to induce expression of SWE1. We then measured bud lengths at time intervals after induction of GAL1-SWE1 to determine the rate of polar bud growth. We found that polar bud growth in GAL1-SWE1 rts1Δ cells occurred at a slower rate than GAL1-SWE1 control cells (Figure 2A and 2B). Controls showed that wild type and rts1Δ cells expressed similar levels of Swe1 protein (Figure 2C). Similar results were also obtained by measuring rates of bud elongation after induction of GAL1-SWE1 in log phase populations of cells (not shown).
We next determined whether Rts1 plays a role in initiation of polar bud growth. To do this, we assayed initiation of bud growth in synchronized populations of wild type and rts1Δ cells. Cells were released from an early G1 arrest and the timing of bud emergence was determined (Figure 2D). Cells lacking Rts1 showed a delay in bud emergence of 22 minutes ± 0.15 minutes compared with wild type cells. Together, these results demonstrate that Rts1 is required for both the timely initiation and the normal rate of polar bud growth.
Since G1 cyclins are required for initiation and maintenance of polar bud growth, it seemed likely that Rts1 is required for functions mediated by G1 cyclins. To test this, we determined whether rts1Δ showed genetic interactions with the G1 cyclins. We found that cln2Δ significantly reduced the rate of proliferation of rts1Δ cells (Figure 3A). Moreover, we failed to recover rts1Δ cln1Δ cln2Δ spores when rts1Δ was crossed to a cln1Δ cln2Δ strain, which suggested that rts1Δ is synthetically lethal with cln1Δ cln2Δ. To further test this, we created a GAL1-CLN2 cln1Δ rts1Δ strain, in which the expression of CLN2 could be repressed by switching from galactose-containing media to dextrose-containing media. This strain grew well on galactose, but was inviable on dextrose, which confirmed that rts1Δ is lethal in cln1Δ cln2Δ cells (Figure 3B). To characterize the defects caused by loss of Rts1, Cln1, and Cln2, we turned off CLN2 expression in the GAL1-CLN2 cln1Δ rts1Δ cells by shifting the cells to media containing dextrose for 8 hours (Figure 3C). The GAL1-CLN2 cln1Δ rts1Δ cells arrested primarily as large, unbudded cells with a small percentage of budded cells (6.5%). Control cells carrying rts1Δ or cln1Δ GAL1-CLN2 had 35% and 15% budded cells respectively. The GAL1-CLN2 cln1Δ rts1Δ cells also became abnormally large, which is commonly observed in cells that fail to undergo bud emergence (Figure 3C) [39],[40].
We next tested whether rts1Δ showed a genetic interaction with cln3Δ. When rts1Δ cells were crossed to cln3Δ cells, a small proportion of the expected progeny was recovered (2/80 spores were recovered rather than the predicted 20/80), while the other progeny segregated according to the expected Mendelian ratios (rts1Δ: 21/80, cln3Δ: 20/80, wild type: 18/80). The few recovered rts1Δ cln3Δ cells formed colonies at the same rate as rts1Δ cells, but their low rate of recovery from the cross suggested that they could contain suppressor mutations. To analyze the viability of cells lacking RTS1 and CLN3 in a context unlikely to select for suppression, we utilized a GAL1-CLN3 rts1Δ strain. When switched to dextrose-containing medium we found that GAL1-CLN3 rts1Δ cells were viable, although they formed colonies at a slightly slower rate than rts1Δ cells (not shown).
It was previously reported that rts1Δ is synthetically lethal with the cyclin-dependent kinase Pho85 in the S288C strain background [41]. We were able to isolate a few rts1Δ pho85Δ spores in the W303 strain background, although they were poorly viable and grew significantly slower than either rts1Δ or pho85Δ cells (Figure 4A). We were also able to recover pcl1Δ pcl2Δ rts1Δ cells from crosses. These grew more slowly than rts1Δ cells but were more robust than pho85Δ rts1Δ cells (Figure 4B). Thus, the poor viability of pho85Δ rts1Δ cells is not strictly due to lack of Pho85 activity associated with the Pcl1/2 cyclins, and may indicate that additional Pho85/Pcl complexes are important for normal growth in rts1Δ strains. In contrast to cells lacking CLN1, CLN2 and RTS1, the pcl1Δ pcl2Δ rts1Δ cells did not show severe defects in bud formation, although they did become larger than the rts1Δ or pcl1Δ pcl2Δ cells. (Figure 4C).
In summary, rts1Δ showed genetic interactions with multiple G1 cyclins and cyclin-dependent kinases. Because the late G1 cyclins show extensive redundancy, mutations that cause reduced function of G1 cyclins should show synthetic interactions with mutations that cause a further reduction in cyclin levels. The fact that rts1Δ showed lethality when combined with cln1Δ cln2Δ, and reduced viability when combined with pcl1Δ pcl2Δ, suggests that Rts1 is required for the normal function of both pairs of cyclins, rather than mediating the functions of specific cyclins. The fact that rts1Δ also showed genetic interactions with cln3Δ, which is upstream of the late G1 cyclins, indicates that Rts1 does not act solely in a Cln3-dependent pathway that promotes transcription of the late G1 cyclins.
We next determined whether accumulation of the G1 cyclin Cln2 occurred normally in synchronized rts1Δ cells. For these experiments, we utilized a 3XHA-tagged version of CLN2 expressed from the CLN2 promoter and quantitative Western blotting to assay Cln2 protein levels. These experiments revealed that the peak of Cln2 accumulation was delayed by 10–15 minutes in rts1Δ cells and that Cln2 failed to reach normal levels (Figure 5A and 5C). The effects of rts1Δ on Cln2 accumulation were more severe at 34°C and 37°C (Figure 5D). Accumulation of the mitotic cyclin Clb2 was correspondingly delayed and cells appeared to delay in G2/M, as revealed by sustained levels of Clb2 relative to the wild type control (Figure 5A).
Since the cells used in these experiments were synchronized with mating pheromone, it was possible that the delayed accumulation of Cln2 was due to delayed release from mating pheromone arrest. To determine whether rts1Δ cells underwent normal release from mating pheromone arrest, we assayed the phosphorylation state of Cdc24, which is the guanine nucleotide exchange factor for Cdc42 [42]. In previous work, it was found that Cdc24 becomes hyperphosphorylated during mating pheromone arrest and undergoes dephosphorylation upon release from the arrest [12],[43]. Cdc24 then undergoes hyperphosphorylation during G1 that is dependent upon the Cla4 kinase and Cdk1 [12],[43],[44]. We found that Cdc24 underwent normal dephosphorylation in rts1Δ cells after release from mating pheromone arrest, which suggested that rts1Δ does not cause delayed release from mating pheromone arrest (Figure S1A). Cdc24 showed delayed phosphorylation in rts1Δ cells, consistent with delayed initiation of G1 events.
To further rule out the possibility that the G1 delay was due to mating pheromone-induced arrest, we used an alternative method for cell synchronization. Cells can be arrested in mitosis by depletion of Cdc20, which is required for proteolytic destruction of the mitotic cyclins [45]–[47]. CDC20 was placed under the control of the GAL1 promoter in wild type and rts1Δ cells. Synchronization in metaphase was achieved by shifting cells to media lacking galactose for 4 hours, followed by releasing cells into galactose-containing media to initiate synchronous exit from mitosis. Cells lacking Rts1 showed a 30–40 minute delay in Cln2 accumulation and reduced Cln2 levels under these conditions (Figure 5B).
We also tested whether the effects of rts1Δ on Cln2 accumulation were dependent upon the strain background. Several commonly used laboratory yeast strains contain different alleles of the SSD1 gene, which can cause significant differences in phenotypes [48]. However, we found that rts1Δ caused similar defects in Cln2 accumulation in both the W303 (ssd1-d2) strain background and the S288C (SSD1-v1) strain background (Figure 5 and Figure S1B, respectively).
Our finding that Cln2 accumulation was delayed and reduced in rts1Δ cells suggested an explanation for the reduced polar growth caused by rts1Δ in mutants that undergo excessive polar growth (Figure 1A and 1B and Figure 2A). We hypothesized that rts1Δ leads to reduced and delayed Cln2 accumulation in these cells, thereby causing reduced polar growth. We tested this directly by assaying Cln2 accumulation in synchronized gin4Δ and rts1Δ gin4Δ cells. As expected, rts1Δ caused reduced and delayed accumulation of Cln2, and a corresponding delay in Clb2 accumulation (Figure S2).
We next tested whether rts1Δ affected levels of CLN2 mRNA or mRNAs encoding additional G1 cyclins. Northern blotting revealed that accumulation of CLN2, CLN1 and PCL1 mRNA was reduced and delayed in rts1Δ cells (Figure 5E and 5F and Figure S1C). Transcription of the late G1 cyclins is controlled by the SBF transcription factor. To test whether rts1Δ caused delayed transcription of MBF targets as well, we assayed RNR1 mRNA expression. RNR1 mRNA accumulation was reduced and delayed to a similar extent as CLN2 mRNA in rts1Δ cells (Figure 5G and 5H). Together these results show rts1Δ causes reduced and delayed accumulation of both SBF and MBF-regulated transcripts.
Since rts1Δ caused reduced and delayed transcription of G1 cyclins, the delayed bud emergence observed in rts1Δ cells could be due solely to a role for Rts1 in promoting G1 cyclin transcription. Alternatively, Rts1 could play diverse roles in regulating events required for bud emergence. To distinguish these possibilities, we tested whether expression of CLN2 from the GAL1 promoter could rescue the delayed bud emergence of rts1Δ cells. Wild type and rts1Δ cells carrying GAL1-CLN2 or an empty vector were released from a G1 arrest under conditions that induce expression of CLN2, and the timing of bud emergence was assayed. We found that expression of CLN2 from the GAL1 promoter dramatically advanced the timing of bud emergence in rts1Δ cells, providing nearly complete rescue of the delay in bud emergence (Figure 6). This observation, combined with our previous observations, established that Rts1 functions in mechanisms directly involved in controlling transcription of the G1 cyclins. Expression of GAL1-CLN2 did not rescue the temperature sensitivity of rts1Δ cells, which indicated that the temperature sensitivity of rts1Δ cells must be due, at least in part, to additional functions of Rts1 (not shown).
Recent work found that Cln2 acts in a positive feedback loop to stimulate its own transcription [22]. Thus, the delay in accumulation of CLN2 mRNA could be due to a failure in mechanisms required for normal accumulation of the Cln2 protein, which would disrupt the feedback loop. Overexpression of CLN2 from the GAL1 promoter might be expected to rescue this kind of defect. Therefore, we tested whether Rts1 functions in several mechanisms known to regulate accumulation of CLN2 protein.
Cln2 is a highly unstable protein and proteolysis plays an important role in regulation of Cln2 protein levels. Proteolysis of Cln2 is controlled by the SCF ubiquitin ligase complex, which recognizes phosphorylated sites at the C-terminus of Cln2 and targets Cln2 for destruction [13],[49]. Cdc34 is the E2 ubiquitin conjugating enzyme component of the SCF complex. We hypothesized that the reduced Cln2 protein levels observed in rts1Δ cells could be caused by increased SCF-dependent proteolysis of Cln2 protein. Reduced protein levels, in turn, would disrupt the positive feedback loop that promotes CLN2 transcription, thereby causing reduced and delayed transcription of CLN2 mRNA. To test this possibility, we created an rts1Δ strain that also contained a temperature sensitive allele of CDC34 (cdc34-2). Cells were released from a synchronized G1 arrest into 37°C media, and Cln2 protein expression was followed by Western blotting (Figure 7A). As previously reported, inactivation of Cdc34 in the control cells caused stabilization of Cln2 and a dramatic increase in Cln2 protein levels [13]. In the cdc34-2 rts1Δ cells, Cln2 protein levels were still reduced and did not accumulate to the high levels observed in cdc34-2 cells. This showed that the failure of rts1Δ cells to accumulate normal levels of Cln2 is due to a failure to produce Cln2, rather than to increased SCF-dependent destruction of Cln2.
We also considered the possibility that Rts1 regulates Cln2 stability via SCF-independent mechanisms. The Cln2 protein has a short half-life of 8–10 minutes [13],[49]. To determine whether rts1Δ decreased the half-life of the Cln2 protein, we expressed a burst of CLN2 from the GAL1 promoter and then measured the rate of destruction of Cln2 protein after shutting off the promoter. In wild type control cells, the half-life of Cln2 was 9.6±2.0 minutes, similar to previous reports. In rts1Δ cells, the half-life of Cln2 was 10.8±2.4 minutes, which showed that there is not a significant decrease in the stability of Cln2 protein (Figure 7B).
To further define the function of Rts1, we tested whether it acts in pathways known to control G1 cyclin transcription. The Whi5 transcriptional repressor delays G1 progression by inhibiting transcription of G1 cyclins. Whi5 acts by binding and inhibiting the SBF and MBF transcription factors, which are required for transcription of the G1 cyclins [16],[17]. The Cdk1/Cln3 complex relieves this inhibition by phosphorylating Whi5, which triggers export of Whi5 from the nucleus. Thus, it was possible that Rts1 played a role in the inactivation of Whi5. If this were true, whi5Δ should rescue the delayed expression of Cln2 observed in rts1Δ cells. However, we found that bud emergence and accumulation of Cln2 protein were still delayed in rts1Δ whi5Δ cells compared to whi5Δ cells (Figure 8A, and data not shown). In addition, whi5Δ did not rescue the temperature sensitivity of rts1Δ cells (Figure 8B). Loss of Whi5 advanced the production of Cln2 protein in rts1Δ cells, although not to the same extent observed in whi5Δ cells, which indicated that Whi5-dependent regulation of transcription occurs normally in rts1Δ cells. Thus, the delayed Cln2 expression in rts1Δ cells is not due to a failure to inactivate Whi5.
Bck2 acts in a redundant pathway that works in parallel to Cln3 to promote transcription of G1 cyclins [18]–[20]. To test whether Rts1 acts in this Bck2-dependent pathway, we crossed rts1Δ to bck2Δ to create rts1Δ bck2Δ cells. If Rts1 functioned solely in the Bck2-dependent pathway, we expected to see no additive phenotypic effects in the double mutant.
All of the expected rts1Δ bck2Δ progeny were recovered from the cross. We found that bck2Δ increased the temperature sensitivity of rts1Δ (Figure 9A). To test whether deletion of Bck2 altered the timing of G1 events in rts1Δ cells, we compared the timing of bud emergence and Cln2 accumulation in wild type, rts1Δ, bck2Δ, and rts1Δ bck2Δ cells. Cells lacking BCK2 delayed bud emergence to a similar extent as rts1Δ cells. Bud emergence was severely delayed in rts1Δ bck2Δ cells when compared to either single mutant, and a subset of cells failed to bud by 2 hours after release from G1 arrest (Figure 9B). Cln2 accumulation peaked later in rts1Δ bck2Δ cells than in rts1Δ or bck2Δ cells, and accumulated to lower levels (not shown). These observations are consistent with previous reports that CLN2 mRNA is reduced in bck2Δ cells [18],[19]. When tested at 34°C, we saw results that were similar, although more severe (not shown). We examined the phenotype of rts1Δ bck2Δ cells in log phase cultures that were grown at room temperature, and found that rts1Δ bck2Δ cells appeared much larger than either rts1Δ or bck2Δ cells (Figure 9C). The strong additive effects of rts1Δ and bck2Δ rule out the possibility that Rts1 functions solely in the Bck2-dependent pathway that controls G1 cyclin transcription, although it remains possible that Rts1 contributes to both Bck2-dependent and independent pathways.
We next determined whether we could detect a role for Rts1 in regulating SBF or MBF. The components of SBF and MBF are Swi6, Swi4 and Mbp1. The Stb1 protein also associates with SBF and MBF and regulates their activity [50]. We therefore generated 3XHA-tagged versions of these proteins and determined whether they showed Rts1-dependent changes in modification state or protein levels. Stb1-3XHA and Swi6-3XHA showed multiple forms on Western blots due to phosphorylation, as previously reported [50]–[52]. Loss of Rts1 caused no detectable changes in the levels of Stb1 modification during a synchronized cell cycle (not shown). In contrast, Swi6-3XHA showed reduced phosphorylation in rts1Δ cells at 20 to 30 minutes after release from a mating pheromone arrest (Figure 10A). Notably, the defect in Swi6 phosphorylation occurred at the time that cells would normally be initiating G1 cyclin transcription. We detected no changes in the protein levels of Swi4 or Mbp1 in rts1Δ cells during the cell cycle. Swi4 and Mbp1 both migrated as a single band, so electrophoretic mobility shifts could not be used to assay their modification states.
We also tested for genetic interactions between rts1Δ and swi6Δ, swi4Δ or mbp1Δ. Previous studies found that swi6Δ is recovered poorly from genetic crosses (27% of expected swi6Δ progeny are recovered) [53]. We obtained similar results, and were unable to obtain swi6Δ rts1Δ progeny, suggesting that swi6Δ is synthetically lethal with rts1Δ. We found that mbp1Δ rts1Δ cells grew more poorly than either single mutant (Figure 10B). We detected no genetic interaction with swi4Δ (Figure 10C). The synthetic lethal interaction with swi6Δ must be treated with caution because swi6Δ shows synthetic lethality with a surprisingly broad range of genes, including genes that do not appear to have G1-specific functions (BioGRID database). Thus, the synthetic lethality may be due to functions of Rts1 that are not related to G1 functions.
Our analysis of the role of Rts1 in control of G1 cyclins suggested that Rts1 does not function solely in the known critical pathways for control of G1 cyclin transcription that work through Cln3, Whi5, or Bck2. This was an intriguing discovery, because previous work found that the mechanisms responsible for nutrient modulation of cell size do not control G1 cyclin transcription via Cln3, Whi5 or Bck2 [11],[25]. We therefore hypothesized that Rts1 controls G1 cyclin transcription in a distinct pathway that mediates nutrient modulation of cell size. To test this, we determined whether Rts1 is required for nutrient modulation of cell size. We grew wild type and rts1Δ cells in rich or poor carbon sources and measured cell size with a Coulter counter. We found that Rts1 is required for nutrient modulation of size (Figure 11A). The slight shift in size observed for rts1Δ cells shifted to poor carbon sources is similar to the slight shift observed for sch9Δ and sfp1Δ, which have also been found to be required for nutrient modulation of cell size [25]. Since rts1Δ cells are abnormally large, we were concerned that they may already be above the critical size for initiation of G1 cyclin transcription, and therefore not subject to G1 size control. To test this, we used mih1Δ cells, which are abnormally large because they undergo extra growth during G2/M [54],[55]. The mih1Δ cells showed normal nutrient modulation of cell size control (Figure 11B). Furthermore, others have found that cln3Δ cells, which are also abnormally large, undergo normal nutrient modulation of cell size [25].
Loss of Rts1 caused reduced and delayed expression of multiple G1 cyclins, and a corresponding delay in bud emergence. Overexpression of CLN2 from a heterologous promoter largely rescued the delayed bud emergence in rts1Δ cells. We found no evidence that Rts1 regulates Cln2 protein stability. Together, these observations demonstrate that Rts1 functions in a pathway that regulates G1 cyclins at the level of transcription.
The results of genetic analysis further supported the conclusion that Rts1 works in a pathway that controls G1 cyclin levels. The genetic interactions that we observed for rts1Δ are summarized in Table 1. In general, rts1Δ caused slow growth or lethality when combined with deletions of G1 cyclin genes or pho85Δ. Due to the redundancy of the G1 cyclins, a general reduction in levels of G1 cyclin expression caused by rts1Δ would be expected to cause additive effects when combined with gene deletions that further reduce levels of G1 cyclins.
The finding that Rts1 is required for normal levels of Cln2 protein provides an explanation for why rts1Δ caused reduced polar growth in mutants that fail to properly inactivate Swe1. Failure to inactivate Swe1 causes cells to arrest at G2/M with high levels of Cln2 protein [12]. Since Cln2 promotes polar growth, a reduction in Cln2 levels would be expected to cause reduced polar growth. Accordingly, we found that rts1Δ caused reduced and delayed accumulation of Cln2 in gin4Δ cells, which fail to inactivate Swe1.
Previous work found that G1 cyclin transcription is regulated by another PP2A-like phosphatase called Sit4. Loss of Sit4 caused decreased transcription of late G1 cyclins and defects in bud emergence, similar to rts1Δ [56]. However, there are a number of significant differences in the G1 phenotypes caused by sit4Δ and rts1Δ. First, in contrast to rts1Δ, defects in bud emergence caused by sit4Δ are not rescued by expression of CLN2 from a heterologous promoter, which demonstrates that Sit4 carries out functions required for bud emergence beyond controlling G1 cyclin transcription [56]. Second, the phenotype of sit4Δ cells is strongly enhanced by the ssd1-d2 allele, whereas the phenotype of rts1Δ is not affected by the status of SSD1 [56]. Finally, sit4Δ cln3Δ cells are barely viable, whereas loss of RTS1 caused relatively mild effects in cln3Δ cells [56]. These phenotypic differences suggest that Rts1 and Sit4 may function in independent pathways that regulate late G1 cyclin levels in response to different signals, but do not rule out the possibility that they have overlapping functions.
We used epistasis analysis to determine whether Rts1 regulates G1 cyclin transcription via known mechanisms. This analysis demonstrated that Rts1 does not function solely in a Bck2-dependent pathway, and ruled out a role for Rts1 in a pathway known to regulate G1 cyclin transcription via Cln3-dependent inactivation of the Whi5 transcriptional repressor. Cln3 may also promote G1 cyclin transcription in a Whi5-independent manner. Overexpression of CLN3-1 makes whi5Δ cells smaller, which suggests that Cln3 can drive G1 cyclin transcription by an alternative mechanism [16]. In genetic crosses, we found that most cln3Δ rts1Δ spores failed to germinate, and cells lacking both Rts1 and Cln3 showed slow growth when compared to either single deletion. Thus, Rts1 does not appear to function solely in a Cln3-dependent pathway that promotes G1 cyclin transcription. Together, these observations suggest that Rts1 does not function solely in one of the known pathways that play an important role in promoting G1 cyclin transcription. Our results do not rule out the possibility that Rts1 contributes to multiple pathways.
We found that rts1Δ caused reduced and delayed expression of both SBF and MBF-dependent transcripts, which demonstrated that Rts1 acts in a pathway upstream of both transcription factors. We further found that rts1Δ caused a reduction in Swi6 phosphorylation at the time that cells initiate G1 cyclin transcription. Since Swi6 is the one shared component of SBF and MBF, this suggests that Rts1 regulates both transcription factors via Swi6. In support of this, the pattern of genetic interactions observed for rts1Δ is similar to the pattern previously reported for swi6Δ. Both swi6Δ and rts1Δ cause slow growth in cln2Δ cells, lethality in cln1Δ cln2Δ cells, and slow growth or lethality in bck2Δ cells [57],[58]. In addition, neither rts1Δ nor swi6Δ caused synthetic lethality in combination with cln3Δ cells [58]. In contrast, loss of SWI4 is lethal in combination with cln3Δ [58]. The fact that Swi6 undergoes reduced phosphorylation in rts1Δ cells, rather than hyperphosphorylation, indicates that it is unlikely to be a direct target of PP2ARts1.
Phosphorylation of Swi6 that can be detected by an electrophoretic mobility shift is dependent upon the MAP kinase Slt2, and activation of Slt2 coincides with initiation of polar growth [51],[59]. The Slt2 pathway is activated by Pkc1, and overexpression of Pkc1 suppresses swi4 mutants, which demonstrates a role in controlling G1 cyclin transcription [60]. In previous work, slt2Δ was not found to cause reduction in the levels of CLN1 or CLN2 transcripts, but did cause a reduction in levels of PCL1 and PCL2 transcripts in cells grown at 37°C. These studies did not utilize synchronized cells and may therefore have missed effects of slt2Δ on levels of CLN1 and CLN2 transcripts. In addition, rts1Δ could lead to misregulation of Slt2 that causes effects on CLN1 and CLN2 transcripts that are more significant than the effects caused by slt2Δ. Further work will be necessary to test for possible roles of Rts1 in Slt2-dependent regulation of G1 cyclin transcription.
Previous work found that nutrient modulation of cell size does not require Cln3, Whi5, or Bck2, which suggests that it works via a distinct mechanism. The discovery that Rts1 does not function solely in one of the pathways known to play a critical role in controlling G1 cyclin transcription therefore prompted us to test whether Rts1 is required for nutrient modulation of cell size. We found that Rts1 is required for nutrient modulated control of cell size, which identifies Rts1 as a new component of the pathways that integrate nutrient availability, cell size, and entry into the cell cycle.
Little is known about the pathways responsible for nutrient modulation of cell size. Two conserved pathways play broad roles in controlling nutrient sensing, nutrient utilization, cell growth and cell cycle entry [61]. In one pathway, the TOR kinases respond to the availability of nitrogen and trigger activation of pathways that control cell growth and nitrogen utilization. A second pathway responds to carbon source availability and regulates growth via activation of the Ras GTPase and protein kinase A (PKA) [61]. The Ras/PKA pathway is required for control of cell size: increased activity of the pathway leads to increased cell size, while decreased activity leads to reduced cell size [62]–[64]. However, the mechanisms by which Ras/PKA control cell size are unknown.
A key target of both pathways is ribosome biogenesis. Two key regulators of ribosome biogenesis that are thought to regulate cell growth are Sch9 and Sfp1. Sch9 is a member of the AGC kinase family and is closely related to vertebrate Akt kinase, while Sfp1 is related to transcription factors [65]. Sch9 and Sfp1 carry out overlapping functions in controlling transcription of genes required for ribosome biogenesis [25],[66],[67]. Loss of either protein causes reduced ribosome biogenesis, while loss of both is lethal. The TOR pathway appears to work through Sch9, but the mechanisms by which the Ras/PKA pathway controls ribosome biogenesis are unclear [65],[68].
It is thought that nutrient modulation of cell size is achieved by changing the critical cell size required for initiation of G1 cyclin transcription [11]. Thus far, only three proteins have been found to be required for nutrient modulation of cell size: Sch9, Sfp1 and PKA [25],[64]. Notably, all three converge on control of ribosome biogenesis. Moreover, mutants that cause reduced rates of ribosome biogenesis cause cells to enter the cell cycle at a reduced cell size, leading to an overall reduction in cell size [67]. Together, these observations suggest a model in which the rate of ribosome biogenesis sets the critical cell size [11],[67]. According to this model, cells growing in rich nutrients have a high rate of ribosome biogenesis, which sends a signal that inhibits G1 cyclin transcription to allow the cell to reach a larger critical size.
The mechanisms that link initiation of G1 cyclin transcription to ribosome biogenesis and nutrient availability are unknown. The identification of Rts1 as a new upstream regulator of G1 cyclin transcription that is also required for nutrient modulation of cell size is therefore a significant step towards understanding these pathways. Rts1 could function between ribosome biogenesis and G1 cyclin transcription. Alternatively, Rts1 could promote ribosome biogenesis, in which case the effects of rts1Δ could be due to decreased rates of ribosome biogenesis. Inactivation of factors that promote ribosome biogenesis cause reduced cell size, whereas rts1Δ causes increased cell size. However, we do not yet know what causes rts1Δ cells to have an increased cell size, and it could be due to a G2/M delay that occurs later in the cell cycle [28].
Since Rts1 is highly conserved, it may also play a role in mechanisms that integrate external signals, cell size, and G1 cyclin transcription in vertebrate cells. The signaling pathways that control G1 cyclin transcription in vertebrate cells are of considerable interest, since deregulation of these pathways contributes to cancer [69].
The strains used for this study are listed in Table 2. Cells were grown in yeast extract-peptone-dextrose (YEPD) media supplemented with 40 mg/liter adenine, unless otherwise noted. Full length CLN2 was expressed from plasmid pSL201-5[GAL1-CLN2-3XHA URA3] [49]. 3XHA-tagging of CLN2 was carried out by digesting plasmid pMT104 with PvuII and integrating at the Cln2 locus using standard yeast transformation techniques [9],[70]. 3XHA-tagging of other genes was carried out as previously described [71].
Cells were grown overnight at room temperature on a shaking platform. Cells at an OD600 of 0.6 were arrested in G1 by the addition of 0.5 µg/ml α factor for 3.5 hours. Cells were released into a synchronous cell cycle by washing 3× with fresh YEPD pre-warmed to 30°C, and time courses were carried out at 30°C unless otherwise noted. To prevent cells from entering a second cell cycle, α factor was added back at 65 minutes. For metaphase arrest, strains containing GAL1-CDC20 were first grown overnight in YEP +2% raffinose +2% galactose and then washed into media containing 2% raffinose and allowed to arrest for four hours. Cells were released from metaphase arrest by adding 2% galactose and were then shifted to 30°C. At each time interval, 1.6 ml samples were collected in screw-top tubes. The cells were pelleted, the supernatant was removed, and 250 µl of glass beads were added before flash freezing. To lyse cells, 100 µl of 1× sample buffer (65 mM Tris-HCl (pH 6.8), 3% SDS, 10% glycerol, 50 mM NaF, 50 mM β-glycerolphosphate, 5% 2-mercaptoethanol, bromophenol blue) was added. 2 mM PMSF was added to the sample buffer immediately before use from a 100 mM stock made in 100% isopropanol. Cells were lysed by shaking in a Biospec Multibeater-8 at top speed for 2 minutes. The tubes were immediately removed, centrifuged for 1 minute in a microfuge and placed in a boiling water bath for 5 minutes. After boiling, the tubes were again centrifuged for 1 minute and 10 µl was loaded on a gel (5 µl when probing for Nap1). For microscopy, 180 µl samples were collected and fixed by adding 20 µl of 37% formaldehyde for 1 hour. Cells were washed twice in 1× PBS, 0.05% Tween-20, 0.02% sodium azide and imaged by differential interference contrast microscopy. Bud emergence was quantified by counting the number of buds per 200 cells for each sample.
Polyacrylamide gel electrophoresis was carried out as previously described [72]. Gels were run at 20 mA on the constant current setting. Western blots were transferred for 1 hour at 1 Amp at 4°C in a Hoeffer transfer tank in a buffer containing 20 mM Tris base, 150 mM glycine, and 20% methanol. Blots were probed overnight at 4°C with affinity purified rabbit polyclonal antibodies raised against Clb2, Swe1, Cdc24, Nap1 or HA peptide. Blots were probed with an HRP-conjugated donkey anti-rabbit secondary antibody (GE Healthcare).
For quantitative western blotting, protein was transferred onto Millipore Immobilon-FL membrane. Before transfer, the membrane was first briefly wetted in 100% methanol. A Cy5-conjugated secondary antibody was used (Affinipure goat anti-rabbit, Jackson ImmunoResearch) and images were collected on a Typhoon 9410 variable mode imager. ImageQuant was used to quantify band intensity. Local background was subtracted from each band. An antibody that recognizes the Nap1 protein was used as a loading control. Since Nap1 migrates below Cln2, the western blots could be cut into two pieces to independently probe Cln2 and Nap1 in the same samples. For each lane, the ratio of Cln2/Nap1 signal was determined to normalize for differences in loading. Graphing was done with GraphPad Prism version 4.00 for Mac [73].
Probes that specifically recognized CLN1, CLN2, PCL1, RNR1 and ACT1 RNA were made using a gel-purified PCR product (CLN1 oligos: AGAATGGTCCTGTAAGAGAAAGT, AGAAACTGATGATGAAGAGGCAT; CLN2 oligos: TGAACCAATGATCAATGATTACGT, TCAAGTTGGATGCAATTTGCAG; PCL1 oligos: ACTAGATTGGTCAGATACACCAA, TGGTTACATCTTTTAGCCTTCTTAGA; RNR1 oligos: ACTTAGGTGTCATCAAGTCATCA, TCTACCACCATGCTTCATGATATCTT; ACT1 oligos: TCATACCTTCTACAACGAATTGAGA, ACACTTCATGATGGAGTTGTAAGT) and the Megaprime DNA labeling kit (GE Healthcare Amersham). Total yeast RNA was extracted as previously described [74] and blotting was carried out using standard methods [75]. Blots were stripped and re-probed for ACT1 as a loading control. Images were collected on a Typhoon 9410 variable mode imager. ImageQuant was used to quantify band intensity. Local background was subtracted from each band. The amount of CLN1, CLN2, PCL1, and RNR1 RNA was normalized to the amount of ACT1 RNA for each lane and used to determine relative expression level. Data from three independent time course experiments was used to determine error bars (standard error of the mean).
Samples of synchronized GAL1-SWE1 or GAL1-SWE1 rts1Δ cells were collected at 30 minute intervals following release from a G1 arrest into galactose-containing media. Cells were imaged using differential interference contrast microscopy and ImageJ 1.37v software was used to measure bud length and bud width for 100 cells for each sample [76]. The extent of polar growth was measured as the ratio of length/width, and was plotted for each time point.
The half-life of the Cln2 protein was determined as previously described with the following modifications [49]. Cells were grown overnight in YEP media containing 2% glycerol/ethanol. GAL1-CLN2 transcription was induced by washing cells into YEP media containing 2% galactose for 1 hour. Expression of CLN2 was then repressed by washing cells into YEPD media. Quantitative Western blotting was carried out to quantify Cln2 levels over time. Non-linear regression curve fitting for one phase exponential decay was carried out using GraphPad Prism version 4.00 for Mac [73].
Cultures of cells were grown in triplicate overnight at 25°C in either YEP +2% dextrose or YEP +2% glycerol and 2% ethanol. A 1 ml sample of log phase (OD600 = 0.60) culture was fixed with 1/10 volume 37% formaldehyde for 1 hour, then washed twice with 1× PBS+0.04% sodium azide +0.02% Tween 20. Cell size was measured using a Channelizer Z2 Coulter counter as previously described [77]. 150 µl of fixed culture were diluted in 20 ml of Isoton II and sonicated for 20 seconds prior to Coulter counter analysis. Each plot is the average of 4 independent experiments in which 3 independent samples were analyzed.
|
10.1371/journal.pcbi.1003966 | Evolution of Integrated Causal Structures in Animats Exposed to Environments of Increasing Complexity | Natural selection favors the evolution of brains that can capture fitness-relevant features of the environment's causal structure. We investigated the evolution of small, adaptive logic-gate networks (“animats”) in task environments where falling blocks of different sizes have to be caught or avoided in a ‘Tetris-like’ game. Solving these tasks requires the integration of sensor inputs and memory. Evolved networks were evaluated using measures of information integration, including the number of evolved concepts and the total amount of integrated conceptual information. The results show that, over the course of the animats' adaptation, i) the number of concepts grows; ii) integrated conceptual information increases; iii) this increase depends on the complexity of the environment, especially on the requirement for sequential memory. These results suggest that the need to capture the causal structure of a rich environment, given limited sensors and internal mechanisms, is an important driving force for organisms to develop highly integrated networks (“brains”) with many concepts, leading to an increase in their internal complexity.
| The capacity to integrate information is a prominent feature of biological brains and has been related to cognitive flexibility as well as consciousness. To investigate how environment complexity affects the capacity for information integration, we simulated the evolution of artificial organisms (“animats”) controlled by small, adaptive neuron-like networks (“brains”). Task environments varied in difficulty due primarily to the requirements for internal memory. By applying measures of information integration, we show that, under constraints on the number of available internal elements, the animats evolved brains that were the more integrated the more internal memory was required to solve a given task. Thus, in complex environments with a premium on context-sensitivity and memory, integrated brain architectures have an evolutionary advantage over modular ones.
| Many studies have sought to elucidate the role of information in evolution [1]–[4], its relation to fitness [5]–[7], and how information about the environment is acquired and inherited by an organism [8], [9]. Common to most current approaches to characterize and quantify information in biology is the notion that biological information has to be physically implemented and should be functional, meaning valuable to the organism and related to the environment [3], [4], [9]. There is also growing interest in how measures of information can shed light on the apparent growth in complexity during evolution [2], [10]–[12].
Artificial adaptive agents (“animats”) have proven useful for investigating how various information and complexity measures change during evolution [6], [7], [13]. Animats consist of small neural networks (“brains”), with sensors, hidden elements, and motors, which are evolved under selection based on task fitness. In recent work we used animats consisting of Hidden Markov elements evolving in a task environment that requires integrating current sensor inputs with memory. We showed that the animats' increasing fitness is associated with an increase in the capacity to integrate information [6], [7].
In this study, we extend these initial results in two ways. First, we evaluate the animats' capacity for integrated information using the comprehensive set of measures recently introduced in the context of integrated information theory (IIT 3.0, see Box 1 and [14], [15], for previous versions see [16] (“IIT 2.0”) and the original formulation for stationary systems [17], [18] (“IIT 1.0”)). Specifically, we ask whether adaptation to an environment leads to an increase in the number of evolved concepts and in the total amount of integrated conceptual information (ΦMax, “Big Phi”). Second, we compare how different task environments influence the evolution of animats and their capacity to integrate information depending on memory requirements and size of the sensory-motor interface. In this way, we aim to elucidate under which conditions integrated brains with high ΦMax become advantageous.
Information-theoretic approaches to assess the evolved complexity of (artificial) organisms are typically based on extrinsic correlational measures, either between the system's genome and its environment [8], [19] or between the system's sensors and motors [20] (sensory-motor information), or between successive system states [6], [7] (predictive information [21]). By contrast, IIT quantifies information from the intrinsic perspective of the system, based on the causal power of its internal mechanisms - the “differences that make a difference” within the system [14]–[16], [18], [22], [23]. In the animats employed here, a mechanism consists of one or more system elements that, at a given time, are in a particular state (on or off). A mechanism in a state specifies a concept if it meets the following conditions (see Methods for details). First, the mechanism must specify which past and future states of the system are possible and which are not (information). The particular way in which it does so constitutes its cause-effect repertoire, the probability distribution of past and future system states given the current state of the mechanism. Second, its cause-effect repertoire must be irreducible to that specified by sub-mechanisms (integration). Irreducibility of a mechanism is assessed by measuring its integrated information φ (“small phi”) - the distance between the cause-effect repertoire of the intact mechanism and that of its minimum partition (MIP), which renders the weakest connection of the mechanism causally ineffective. φ thus quantifies how much causal information is lost due to the MIP. A mechanism can specify only one cause-effect repertoire, the one that is maximally irreducible (exclusion, φMax, see Methods). This constitutes its concept—what the mechanism in a state “does” for the system from the intrinsic perspective of the system itself.
The set of all concepts and associated φMax values generated by a set of elements constitutes a conceptual structure (information, see Methods for details). As for individual concepts, the integration of a conceptual structure can be evaluated by measuring the distance Φ (“big phi”) between the conceptual structure of the intact set and that of its minimum (unidirectional) partition (see Methods). Within some animats, a set of elements may generate a maximally integrated conceptual structure (ΦMax), which constitutes a main complex (MC, exclusion). Other animats may not contain complexes (Φ = 0) because their brains are constituted of functionally segregated modules with feed-forward architecture (containing at most self-loops) [15]. In sum, Φ can be viewed as a measure of complexity, since only systems with many specialized, but integrated mechanisms have high Φ, whereas systems that have only a few different mechanisms and/or are very modular have low or no Φ [15], [16], [22].
From an engineering point of view, modular systems with segregated functions are much simpler to design and understand than integrated systems. However, simplicity of design is not an issue for evolution by natural selection. Instead, important factors are economy of elements/wiring [24], composition of functions [14], degeneracy (multiple ways to achieve the same function) [25], adaptability in the face of change [26], [27], integrated control [14], and robustness to failure [28]. These factors should favor the evolution of organisms with integrated brains in an environment that is complex, changing, and requires sensitivity to context [14], [25], [29]. Based on these considerations, we predict that measures of integrated information should increase with the complexity of the environment. Specifically, i) evolving animats should show an increase in the number of concepts; ii) integrated conceptual structures should become larger and more irreducible; iii) the increase in concepts and integrated conceptual structures should be related to the complexity of the environment and to the requirements for memory. Moreover, to the extent that IIT is correct in claiming that the capacity for information integration underlies consciousness [14], [15], [18], [23], finding an increase in animats' ΦMax values in complex environments would provide a plausible account of why and how consciousness evolved.
In what follows, we test and confirm these predictions by evolving animats solving perceptual categorization tasks [13], [30] in task environments that vary in the amount of sequential memory necessary to solve the task optimally. The results show that, given strict constraints on the number of elements in the animat's brain, integrated network architectures become advantageous over modular or feed-forward architectures when the environment was more complex. Moreover animats with restrictions on the number/fidelity of their sensors or motors evolved more concepts and larger integrated conceptual structures, in line with an increased reliance on memory.
In order to investigate the causal structure of a system from an evolutionary perspective, we simulated the adaptation of simple neural networks (“animats”) [6], [7], [13] in task environments of varying difficulty. For these animats, their concepts and integrated conceptual information Φ can be calculated rigorously across many generations (see Methods). This permits testing the following predictions about the evolution of (integrated) conceptual information during adaptation to specific environments:
Each animat is equipped with a fixed number of sensors, hidden elements, and two motor outputs (to move either left or right, see Fig. 1). All elements are binary Markov variables, whose value is specified by deterministic logic gates. Each animat has a genome, which encodes the wiring diagram of the animat's brain and the logic functions of its elements. More precisely, each gene specifies a hidden Markov gate (HMG) and all HMGs together determine the brain's causal structure (see Methods and [6], [13]). The animats are allowed to evolve over 60,000 generations using a genetic algorithm, starting with an initial population of 100 animats without connections between brain elements (generation zero). To compose the next generation, the genetic algorithm selects a new sample of 100 animats based on an exponential measure of the animats' fitness (roulette wheel selection). The genome of each selected animat is mutated according to three probabilistic mutation mechanisms (point mutations, deletions, and duplications) [13]. The mutated genomes then determine the wiring diagrams and logic functions of the next animat generation, which are tested for fitness in the respective task environment. In sum, adaptation arises through mutation and selection driven by the animat's task performance.
Throughout this study, the animats' task environments are variants of “Active Categorical Perception” (ACP) tasks [13], [30], where moving blocks of different sizes have to be distinguished in a ‘Tetris-like’ game (Fig. 1B). Adaptation is measured as an increase in fitness, where fitness corresponds to the fraction of successfully caught or avoided blocks within a fixed number of trials (128 for each animat at each generation, with one falling block per trial). Blocks move sideways and down at 1 unit per time step either to the right or left starting from one of 16 possible initial positions. If a block moves out on the left it will reappear on the right and vice versa. A block is “caught” if the animat overlaps with at least one of its units when it reaches the bottom (after 36 time steps); otherwise the block is “avoided”. Each animat's size is 3 units, with a space of 1 unit between the two sensors (a “blind spot”). Therefore, only blocks of size ≥3 can activate both sensors at the same time (Fig. 1C,D). Note that the sensors of the animat convey limited information about the environment and only at a single time step, yet solving ACP tasks successfully requires integration of sensor inputs over multiple time steps. Hence, information about past sensor states (memory) has to be stored through the states of internal elements.
At the end of each evolutionary run at generation 60,000, the line of descent (LOD) of one animat is traced back through all generations. Every 512 generations along the LOD, a transition probability matrix (TPM) is generated for all possible states of the animat's brain, which captures how the brain transitions from one state to another. From these TPMs, concepts and integrated conceptual information Φ can be calculated across the LOD. We averaged the causal measures for a particular generation in one LOD across all network states experienced by the animat during the 128 test trials, weighted by their probability of occurrence. For each task condition, 50 independent LODs were obtained, each from a different evolutionary run.
To investigate how the number of concepts and their integration depends on the causal structure of the task environment, we tested the animats in four tasks (Task 1–4) with different block categories and strategic requirements (Fig. 1E). Given the periodic boundary conditions and the fact that the animats can actively explore their environment, predicting the evolutionary difficulty of an ACP design is not straightforward. Nevertheless, if solving a task requires more memory of input sequences, the number of concepts developed by the animats should increase. Since the number of evolvable hidden elements is limited to four, the number of time steps that can be combined without feed-back between elements and thus Φ = 0 (see Methods and [15]) is limited, too. Higher memory requirements should thus bias the animats towards developing brains with more integrated conceptual structures with larger main complexes and higher Φ.
As a first simple task environment (Task 1), the animats have to catch blocks of size 1 and avoid blocks of size 3. In Task 1, the two block conditions can in principle be distinguished based on a momentary sensor state (S1S2 = 11, see Fig. 1C,D). Categorization can thus be achieved in a modular manner (e.g., “if S1S2 = 11 avoid, else follow”). However, memory is still required to identify the direction of the moving blocks, since sensor information of at least two time steps must be combined to infer movement direction. Task 1 will serve as the comparison environment in the following sections.
In Task 2, the blocks to be avoided are smaller (2 units). Consequently, the two block categories cannot be distinguished based on a single sensor state, since neither block can activate both sensors at the same time. Here, memory is required for both categorization of block size and direction.
In Task 3, four instead of just two different block sizes have to be distinguished. The blocks to be caught (size 1 and 4) and avoided (size 2 and 3) cannot be distinguished based on a single threshold (e.g. “≥3”), nor based on a single sensor state. Adaptation to Task 3 is thus expected to be more difficult. However, sensor state S1S2 = 11 allows distinguishing blocks of size 1 and 2 from blocks of size 3 and 4. Whether to catch or avoid a block can then be decided based on a memory of one time step, just as in Task 2. Note also that in Task 3 at least 75% fitness can be achieved with the same categorization strategy as in Task 2 (“≥2”). Therefore, more concepts than in Task 2 are expected only for fitness levels>75%.
Finally, in Task 4, four blocks of sizes ≥3 have to be distinguished. To successfully catch blocks of size 3 and 6 and avoid blocks of size 4 and 5 the animats have to combine memory of at least 3 time steps.
In sum, the evolutionary pressure to develop brains with integrated concepts should be lowest for Task 1, intermediate for Task 2/3, and highest for Task 4, in line with the requirements of sequential memory in Task 1–4. According to IIT, both the average number of concepts and their integration (ΦMax) should therefore be highest in Task 4 and lowest in Task 1.
Throughout the following analysis, the animats are evaluated in two ways: first, all concepts and the sum of their φMax values are calculated for the animat's brain as a whole, including the sensors, motors, and all hidden elements. These measures quantify all causal relations (“IF-THEN”) in the animat's brain. Second, the main complex (MC) within the animat's brain is identified and the number of elements that form the MC (“MC elements”), the number of concepts in the MC (“MC concepts”), and its ΦMax value are calculated according to IIT 3.0 [15]. These measures quantify the amount of integration in the animat's brain. In this way, increases in fitness that rely on integrated structures can be distinguished from those that can be achieved with modular networks with feed-forward architecture (containing at most self-loops). Fig. 2 illustrates all the causal measures of a potential animat brain in one particular state. The maximal possible number of concepts specified by an animat's brain is 15 (24−1, the power-set of all hidden elements excluding the empty set, see Fig. 2B). An animat's main complex can, at most, comprise the 4 hidden elements. Determining upper bounds for ΣφMax and ΦMax is not straightforward (see S1 Text). In the present set of simulations, the overall highest observed values for an animat in a particular state were ΣφMax = 3.11 and ΦMax = 4.125. Note that all the above measures are state-dependent [15]. At a particular generation, these measures are evaluated for every brain state experienced by the animat during the test trials. The resulting state-dependent values are then averaged, weighted by the probability of occurrence of each brain state.
Fig. 3 shows the evolution of all causal measures during adaptation over 60,000 generations in all four task conditions. For each task condition, 50 independent LODs are assessed every 512 generations. In Table 1, the average Spearman rank correlation coefficients across all 50 LODs are listed for all measures and tasks. As previously observed in a different kind of task environment [7], trial-by-trial correlation coefficients with fitness were rather broadly distributed (see histograms in S1 Fig.). While the causal measures are interrelated to some extent and the MC measures in particular tend to correlate, dissociations among them occur for individual LODs (see S2 Fig. for examples).
Task 1 (Fig. 3, 1st column): At generation 59,904 the average fitness across all 50 LODs was 94.2±0.7% (mean ± SEM); in 13 out of 50 evolutionary lines the animats reached perfect fitness. On average, all causal measures were found to increase during the initial steep rise in fitness. The number of concepts and their ΣφMax values measured in the whole animat brain showed significant positive correlation with fitness (p<0.05) in 34/50 LODs. MC measures only correlated positively with fitness in 12/50 LODs (S1 Fig.), reflecting the fact that both modular (functionally segregated) and integrated concepts can lead to an increase in fitness. In other words, not every increase in fitness requires an increase in integration. In the case of Task 1, perfect categorization can be achieved with a purely modular (no MC, ΦMax = 0, 7/13 animats) as well as with an integrated network (ΦMax>0, 6/13 animats, see below, Fig. 4).
Task 2 (Fig. 3, 2nd column): In terms of adaptation, Task 2 was as difficult as Task 1 since the same level of fitness was reached (94.0±1.2%). Perfect fitness was achieved in 22/50 LODs. 16 out of these 22 animats developed integrated brains. Compared to Task 1 (black), with increasing fitness in later generations the animats developed brains with more concepts and higher ΣφMax values in Task 2 (U98 = 695.5/749.0, Z = −3.844/−3.454, p = 0.000/0.001 respectively for #concepts/ΣφMax averaged across the last 3,000 generations). MC measures in Task 2 increased more subtly, but reached higher values than in Task 1 (U98 = 985/966/922, Z = −1.899/−2.035/−2.350, p = 0.058/0.042/0.019 respectively for #MC elements/#MC concepts/ΦMax). The number of LODs with significant positive correlation with fitness (p<0.05) was also higher than in Task 1 for number of concepts and ΣφMax (42/50) and MC measures (24/50).
Task 3 (Fig. 3, 3rd column): The average fitness reached at generation 59,904 was 82.9±1.0%. Perfect fitness was achieved only temporarily in one LOD (with final fitness 98.4%). The average number of concepts and ΦMax evolved to higher values in Task 3 compared to Task 1 (black) (U98 = 854/899, Z = −2.746/−2.530, p = 0.006/0.011 for #concepts/ΦMax), while ΣφMax and the number of MC concepts and MC elements stayed comparable to those of Task 1. To compare the different tasks without confounding effects due to differences in fitness, a subset of LODs with high final fitness was chosen out of the 50 LODs of Task 3, so that the average fitness across the last 5,000 generations matched that of Task 1 (9 fittest LODs, shown in dark red). When compared at the same level of fitness, all causal measures evolved to significantly higher values, except for the number of MC concepts, which still almost reached significance p<0.05 (U57 = 77.5/71.0/136/141/112, Z = −3.143/−3.247/−1.999/−1.886/−2.538, p = 0.002/0.001/0.046/0.059/0.011 respectively for #concepts/ΣφMax/#MC elements/#MC concepts/ΦMax). As predicted, the evolutionary pressure for concepts and integration in Task 3 appeared to be comparable to that of Task 2. Accordingly, trial-by-trial positive correlation with fitness in Task 3 was also similar to Task 2: number of concepts and ΣφMax correlated significantly with fitness in 39/50 LODs; MC measures correlated significantly with fitness in 24/50 LODs. At comparable average fitness levels, the fact that four instead of just two blocks had to be distinguished only led to a marginal increase in the number of concepts and their integration, since the requirement for sequential memory remained comparable between Task 2 and 3. Solving the more difficult Task 3 perfectly, however, might still require significantly more overall concepts and higher ΣφMax values than Task 2, since the perfect solution requires distinguishing the 4 different block sizes under every initial condition (see below, Fig. 5A).
Task 4 (Fig. 3, 4th column): As expected, Task 4 was the most difficult task in terms of adaptation with an average final fitness of 79.5±1.4% at generation 59,904. The highest overall fitness reached across all 50 LODs was 97.7% (125/128 correct trials) in one LOD. Despite the lower fitness reached, the average number of concepts, ΣφMax, and ΦMax were significantly higher in Task 4 than in Task 1 (U98 = 813/862/850, Z = −3.034/−2.675/−2.879, p = 0.002/0.007/0.004 for #concepts/ΣφMax/ΦMax). More evolutionary pressure for sequential memory thus led to causal structures with a higher number of concepts and more integration. This became even more evident when comparing a subset of LODs of Task 4 with equivalent average fitness (fittest 7 LODs) to Task 1 (U55 = 28.5/47/53/52/33, Z = −3.604/−3.112/−3.159/−3.185/−3.677, p = 0.000/0.002/0.002/0.001/0.000 for #concepts/ΣφMax/#MC elements/#MC concepts/ΦMax). In this subset, the evolved ΦMax of Task 4 was significantly higher than in any of the other tasks (U55/55/14 = 33/61/11, Z = −3.677/−2.792/−2.170 compared to Task 1/2/3). Also most other causal measures were significantly higher than in Task 2 (U55 = 62.5/103/66/74, Z = −2.740/−1.751/−2.670/−2.474, p = 0.006/0.080/0.008/0.013 for #concepts/ΣφMax/#MC elements/#MC concepts). Moreover, the number of LODs positively correlated with fitness was highest in Task 4: in 48/50 LODs the number of concepts and ΣφMax correlated significantly with fitness, and the MC measures correlated significantly with fitness in 33/50 LODs.
Taken together, comparing the causal measures across different task environments confirmed the predictions of IIT: the number of concepts that evolved during adaptation and their integration was higher in those tasks that required more memory and that could not be solved based on momentary sensor inputs – lowest for Task 1, intermediate for Task 2/3, and highest in Task 4.
Given the restrictions imposed on the animats' brains (binary elements and at most 4 hidden elements), evolutionary selection based on task fitness provides a driving force for more concepts and their integration proportional to the amount of memory necessary to solve the tasks. This can be illustrated by considering the evolved network structures with high fitness in Task 1–4.
In Task 1 the maximum fitness reached with just one hidden element was 92.2% (118/128 correct trials). Yet, perfect fitness in Task 1 can be achieved in both a modular and integrated manner, i.e., with network structures with either ΦMax = 0 or ΦMax>0 (Fig. 4). Out of the 13 LODs in which animats reached perfect fitness, 7 developed modular networks. An example LOD is shown in red in Fig. 4A. In this example, an initial increase in fitness at generation 9,216 was accompanied by an increase in integration. Subsequently, however, the animat's brain turned modular again at generation 13,824 (Φ = 0), which in this case led to a jump in fitness. The evolved network structure is shown in Fig. 4B for generation 59,904. The two hidden elements have memory in the form of self-loops, which however does not count as integration (Φ = 0, since single units cannot form a MC because they cannot be partitioned). In all of the 7 independent LODs that led to perfect fitness and a modular brain, the final generation of animats had evolved the same functional wiring diagram and similar logic functions with only 2 types of behavior (low degeneracy).
In the remaining 6 LODs in which animats achieved perfect fitness, they evolved an integrated main complex with feedback between elements. An example LOD is shown in blue in Fig. 4A. The initial increase in fitness of that LOD to 87.5% was achieved without a main complex (ΦMax = 0) and just one concept in the whole animat brain (generation 8,704-51,200). The rapid increase to 100% fitness at generation 52,224, however, was preceded by the formation of a main complex (ΦMax>0) and thus integration of concepts at generation 51,712. In Fig. 4C the final evolved wiring diagram at generation 59,904 is shown. This network structure is predominant among the evolved animats that reached perfect fitness in an integrated manner (5 out of 6). Despite this “anatomical” uniformity, the evolved logic functions, and thus the evolved behavior of the animats in the final generation, differed for all 6 LODs (high degeneracy). Analyzing all animats with perfect fitness across all generations and LODs, the animats with ΦMax>0 showed 341 different TPMs, leading to 332 different behavioral patterns, which were implemented by 15 different wiring diagrams. By contrast, animats with ΦMax = 0 had only 60 different TPMs, leading to 44 different behavioral patterns, which were implemented by 11 different wiring diagrams. Moreover, once a solution (perfect fitness) with ΦMax = 0 was encountered, subsequent descendants with ΦMa>0 networks (and vice versa) were rather rare and the variability of TPMs within one LOD was lower for modular networks with ΦMax = 0 than for integrated networks (see Fig. 4A and S3 Fig.). This indicates that, while solutions with ΦMax = 0 were encountered with about equal probability to ΦMax>0 solutions across 50 independent LODs, within a LOD neutral mutations without decrease in fitness happen more frequently given integrated networks. Recurrent networks with Φ>0 are thus more flexible, in the sense that there are other solutions close by on the fitness landscape, which can be reached through neutral mutations. Taken together, perfect adaptation to Task 1 seems to require at least 2 hidden elements, but could be achieved in a recurrent/integrated and feed-forward/modular manner with about equal likelihood. However, animats with perfect fitness and ΦMax>0 showed higher degeneracy and variability in structure and behavior (see also S3 Fig.).
In Task 2 the maximum fitness reached with just one hidden unit was only 75% (96/128 correct trials) compared to 92.2% in Task 1. The fact that the two categories of blocks in Task 2 have to be distinguished based on memory without the possibility to rely on momentary evidence thus appears to increase the evolutionary pressure to develop more hidden elements. Nevertheless, in Task 2 as well, perfect fitness was achieved with both modular (Φ = 0) and integrated networks (Φ>0). However, out of the 22 independent LODs with perfect fitness only 6 showed no integration of concepts (Φ = 0) at generation 59,504, with the same wiring diagram as shown in Fig. 4C (Task 1) in 5 out of 6 cases. Of the remaining 16 animats with perfect fitness and integrated MCs, half evolved 2 hidden elements and half 3, with 9 different types of wiring diagrams and even higher degeneracy in their evolved logic functions and behavior. This corroborates the fact that evolutionary pressure for more concepts and integration is higher in Task 2 than in Task 1. As in Task 1, degeneracy and variability in network structure and behavior in Task 2 was higher for animats with integrated brains: taking all animats with perfect fitness across all generations and LODs into account, the animats with ΦMax>0 showed 920 different TPMs, leading to 407 different behavioral patterns, implemented by 34 different wiring diagrams, compared to only 235 different TPMs, with 85 different behavioral patterns, implemented by 30 different wiring diagrams for animats with ΦMax = 0.
Although Task 3 and 4 were more difficult, the maximal fitness that was reached with just one hidden element in these tasks was similar to that of Task 2: 78.1% (100/128) in Task 3 and 77.3% (99/128 correct trials) in Task 4. However, even with 2 hidden elements, the highest overall fitness reached was only 96.9% (124/128 correct trials) in Task 3 and 93.8% (120/128 correct trials) in Task 4. While in Task 3 the highest fitness achieved with a modular network without an integrated main complex (Φ = 0) was 96.1%, in Task 4 it was only 89.8%. The wiring diagrams of the fittest animats of both tasks are displayed in Fig. 5. In both cases, the animats developed brains with more than two hidden elements and an integrated main complex. Notably, the fittest animat in Task 4 evolved a main complex that was strongly integrated with <ΦMax> = 1.13 and had many higher order concepts. Fig. 5C shows the conceptual structure of the fittest animat of Task 4 for one representative state. While the MC concepts are always about the elements in the main complex, some may be interpreted from the extrinsic perspective, such as the concept AC = 11, which here could mean “keep going right”. Which concepts exist at a given time depends on the state of the system. In this way, evolved concepts can correlate with and indirectly refer to specific states/events of the environment. A detailed interpretation of the extrinsic and intrinsic meaning of the animats' MC concepts is, however, beyond the scope of this study. Although it cannot be excluded that Task 4 is in principle solvable with 4 hidden elements connected in a non-integrated manner (Φ = 0), these results suggest that evolution strongly prefers integrated brains in Task 4.
In summary, under the constraints of maximally 4 binary, hidden elements, the fittest animats evolved in Task 1 developed modular and integrated wiring diagrams with similar likelihood. With higher memory requirements evolution increasingly selected for integrated networks with ΦMax>0. In Task 4, all animats with>90% fitness (8 LODs) developed an integrated main complex.
Task difficulty and the amount of sequential memory necessary to solve a task depend not only on the environment, but also on the sensor and motor capacities of the animats themselves. Solving the same task with fewer (or worse) sensors and motors requires increased reliance on memory. Consequently, the animats' evolved number of concepts and their integration should increase if the animats' sensor and motor capacities are restricted during adaptation.
To test this hypothesis, 50 additional LODs were evolved in the environment of Task 1 with one of the animats' sensors disabled (set to 0 at each time step and thus rendered useless). As explained above, with two functional sensors the two blocks in Task 1 can be categorized based on momentary sensory data alone (Fig. 2C,D). As a result, Task 1 could be solved equally well with a modular and integrated brain network (Fig. 4). Given only a single sensor, however, the task becomes more complex and requires memory of input sequences for block categorization. Fig. 6 shows the results obtained from the animats with only one sensor compared to Task 1 with two sensors (in black).
The average fitness reached with just one sensor was 82.8±1.4%. Nevertheless, in 4/50 LODs the animats reached 98.4% fitness (126/128 correct trials). As predicted, the animats evolved brains with more concepts, higher ΣφMax, and more integration than those with two sensors at their disposal (U98 = 510.5/514/746.5/749.5/728.5, Z = −5.116/−5.074/−3.591/−3.566/−3.716, p = 0.000, respectively for #concepts/ΣφMax/#MC elements/#MC concepts/ΦMax). Also the number of LODs that correlated positively with fitness was higher with only one sensor: number of concepts and ΣφMax correlated significantly in 46/50 LODs and MC measures in 36/50 LODs (compared to only 34/50 and 12/50, respectively, with two sensors). The increase in concepts and integration due to restricted sensors is even more apparent in the subset of 19 fittest LODs with the same average final fitness as in Task 1 with two sensors (Fig. 6, dark orange).
In terms of network structure, with just one sensor, the maximal fitness achieved with one hidden element was only 67.2% (compared to 92.2% with two sensors) and 95.3% with two hidden elements (100% with two sensors). In three out of the four fittest LODs (98.4% fitness), the animats evolved brains with an integrated main complex (Φ>0). Overall, the results obtained in Task 1 with one sensor are comparable to those of Task 4, the task with the largest block sizes, which requires most sequential memory (Fig. 3, 4th column).
As demonstrated above, restricting the sensor capacities of the animats increased brain integration since Task 1 had to be solved based on memory alone instead of momentary sensor states. Restricting the animats' motor capacities still allows using the sensor state S1S2 = 11 to distinguish blocks of size 3 from size 1. Nevertheless, with just one available motor, reliance on memory should increase, since movements have to be coordinated across several time steps. This, in turn, should lead to more concepts and higher integration. Fig. 7 shows the results of another 50 LODs evolved in Task 1 with one of the animats' motors disabled (set to 0 at every time step).
Overall, restricting the animats' motor capacities to one motor led to larger main complexes with more concepts and higher integration (ΦMax) (U98 = 806/824/741, Z = −3.156/−3.028/−3.618, p = 0.002/0.002/0.000 for #MC elements/#MC concepts/ΦMax). With one motor only, the maximal fitness achieved was 87.5% (112/118 correct trials) in one LOD; average final fitness was 78.8±0.7%. Task 1 with one motor could thus not be compared at the same level of fitness as Task 1 with two motors. Instead, a subset of the 10 fittest animats is plotted in dark green in Fig. 7, in addition to the average across all 50 LODs (light green). In this subset, also the number of modular concepts was significantly increased compared to the standard Task 1 (U58 = 107.5, Z = −2.857, p = 0.004). The maximal fitness reached with one motor and one hidden element was 71.8%. 24/50 animats evolved the same wiring diagram as shown in Fig. 4B, but with only one motor element. The fittest animat (112/128 correct trials) evolved an integrated main complex with at most 3 elements and <ΦMax> = 0.38. Positive correlation with fitness was also higher given just one motor: the number of concepts and ΣφMax correlated significantly in 40/50 LODs and MC measures in 34/50 LODs.
Finally, evolutionary pressure for more memory should also arise with sensory data that are less reliable. Consequently, more concepts and higher integration are expected to evolve in an environment where sensor inputs are noisy, if compensating mechanisms are developed. To test this prediction, we simulated 50 additional LODs of Task 1 with 1% sensor noise for each of the two sensors (Fig. 8), meaning that the state of each sensor had a probability of 1% to be flipped. During evolution with noise, each trial was repeated 20 times and the next generation of animats was selected based on the average fitness across repetitions.
On average (Fig. 8, pink), animats evolved in the noisy environment developed brains with similar number of concepts and integration as those evolved in the noise-free environment (black). Presented with the noise-free Task 1, their average final fitness was lower than for those animats that had adapted to the noise-free environment (88.1±1.0% compared to 94.2±0.7%). Given the limited size of the animats' brains, it is possible that during 60,000 generations no compensatory mechanisms could be developed and the sensor noise only reduced the animats' performance without adaptive influence on their network structures. However, when fitness is evaluated in the environment with 1% sensor noise, the animats that had adapted to the noisy environment reached 79.0±0.8% fitness at generation 59,904, while the animats that had evolved without sensor noise only reached 76.3±0.7% fitness. This indicates that in a subset of the 50 evolutionary runs, the animats adapted to compensate for the sensor noise, at least in part. We thus evaluated the subset of 20 LODs evolved under noise with highest fitness in the noisy environment, shown in purple in Fig. 8. In line with the above predictions, this subset of LODs indeed showed more concepts and a trend for higher ΣφMax, and larger main complexes with more MC concepts than the animats that evolved without sensor noise (U68 = 299.0/368.0/380.0/382.0, Z = −2.638/−1.716/−1.658/−1.630, p = 0.008/0.086/0.097/0.103, respectively for #concepts/ΣφMax/#MC elements/#MC concepts), although their fitness in the noise-free Task 1 was very similar (first panel, Fig. 8). Note that, due to the data processing theorem [31], introducing sensor noise would generally decrease standard (Shannon) measures of information processing across the communication channel between the environment and the animat, regardless of compensatory mechanisms in the system. By contrast, measures of information integration may actually increase, since they take into account the noise compensation mechanisms implemented by the intrinsic causal structure of the animat. Taken together, the results presented in this section show that the number of concepts and their integration not only increase with the complexity of the environment, but also with the complexity of the environment relative to the sensor and motor capacities of the organism. This confirms the hypothesis that, if more reliance on memory is required to reach high levels of fitness and the number of elements is restricted, evolutionary pressure favors more integrated network structures.
In this study, we analyzed how the causal structure of simulated neural networks (animats) evolves during adaptation to environments of increasing complexity. To that end, we first evaluated all concepts (modular and integrated) specified by the brain elements of each animat and measured their integrated information φMax. Second, we identified the animat's main complex (MC), the set of elements in an animat's brain that generates the maximally integrated conceptual structure, and computed its associated integrated conceptual information ΦMax.
We investigated the evolution of animats in four environments (Task 1–4) with different levels of task difficulty and requirements for sequential memory. Task difficulty (assumed to be inversely related to the average evolved fitness after 60,000 generations) was lowest for Tasks 1 and 2 and highest for Task 4. The requirements for sequential memory were low for Task 1, intermediate for Task 2 and 3, and high in Task 4. In accordance with the predictions of IIT, the animats evolved on average more concepts and larger, more integrated main complexes (higher Φ) the more sequential memory was necessary to solve a task. Similar results were obtained in a second set of simulations, in which the animats' sensor or motor capacities were restricted while the animats adapted to Task 1. This increased the reliance on memory and led, as predicted, to more concepts and more integrated conceptual structures. Taken together, these results point to an active evolutionary trend towards more concepts and integrated conceptual structures if the environment's causal structure is complex and there are constraints on the number of sensors, motors, and hidden elements.
The notions of information and complexity play an important role in recent attempts to understand evolutionary success [2]–[4], [6], [7], [13], [20], [21], [32]. For example, Marstaller et al. [13] presented a measure of “representation”, defined in information-theoretic terms as the mutual information between (coarse-grained) states of the environment and internal “brain” states, given the states of the sensors. Applied to animats adapting to a block categorization task similar to Task 1, representation of a set of salient environmental variables was shown to increase during adaptation [13]. Another recent study examined how sensory-motor mutual information (ISMMI) [20], predictive information (IPred) [21], and integrated information as defined in [6], change over the course of adaptation to a single environment with fixed statistical properties (traversing random mazes) [6], [7].
The mutual information between sensors and motors quantifies the degree of differentiation of the observed input-output behavior [20], [32]. Thus, ISMMI reflects the richness of a system's behavioral repertoire (behavioral complexity), which should be advantageous in a complex environment. Predictive information [21]—the mutual information between a system's past and future states—measures the differentiation of the observed internal states of a system. Thus, IPred reflects the richness of a system's dynamical repertoire (dynamical complexity), which is also expected to promote adaptation to complex environments. ISMMI, IPred, and integrated information as defined in [6] all increased during evolutionary adaptation to the maze environment [6], [7]. Moreover, these indices showed a positive correlation with fitness and positive lower bounds pointing to a minimal, necessary amount of complexity for a given fitness [7]. In the present simulations, IPred always increased during evolution and was highest for Task 4 (see S4–S6 Figs.). However, changes in ISMMI with adaptation as measured in [6], [7], [13] varied with the task. Specifically, in Task 1 and 2, after an initial maximum ISMMI actually decreased with increasing memory capacity, as also observed in [13].
The present approach extends previous investigations in several ways. In addition to aggregate measures of information applied to the animat's brain as a whole, we evaluated all the individual concepts specified by the elements of each animat, taken alone or in various combinations (as specified in IIT 3.0 [15]). In essence, concepts characterize the irreducible input-output functions performed by a mechanism in a state [15]. Assessing concepts requires a perturbational approach that reveals a mechanism's causal properties within a system under all possible initial states [14], [15]. Thus, a concept expresses the entire set of causal dispositions or “powers” conferred by a mechanism in a given state to the system to which it belongs. This analysis thus picks up causes and effects, not just correlations, and does so for the entire set of possible circumstances to which an animat may be exposed, not just for those that happen to be observed in a given setting. Importantly, the causal analysis performed here also shows that combinations of elementary mechanisms (higher-order mechanisms) may specify additional concepts, thus greatly enriching the causal powers of an animat for a given number of elements. Crucially, higher-order concepts only count if they are integrated (φ>0), indicating that their causal power cannot be reduced to the causal power of their parts. For each animat in the present study the IIT 3.0 measures were evaluated for every brain state with p>0 and averaged, weighted by each state's probability of occurrence while the animat is performing the task. The finding that successful adaptation to more complex environments leads to the development of an increasing number of concepts fits well with the notion that, everything else being equal, different concepts provide different causal powers, thereby increasing the substrate available to selective processes.
The present results also show that complex environments lead not only to an increasing number of concepts available to an animat, but also to the formation of integrated conceptual structures within the animats' brains. If a conceptual structure specified by a set of elements is maximally irreducible to the conceptual structures specified by subsets of elements (ΦMax), the set of elements constitutes a main complex (MC) [15]. The conceptual structure specified by the main complex of an animat thus corresponds to a local maximum of causal power. In this way, the main complex forms a self-defined causal entity, whose borders are determined based on the causal powers of its own mechanisms. Importantly, while the concepts within a main complex are specified over hidden elements (the cause-effect repertoires are all within the MC), they do reflect previous input from the sensors and they can, of course, influence the motors. In this way, an integrated conceptual structure can combine current inputs and outputs with past ones and with the state of internal elements that may reflect past memories as well as future goals. All the concepts specified by the main complex over itself thus reflect a system's intrinsic complexity.
Complexity and fitness are often associated, though not invariably [6], [7], [10], [33], [34]. In particular environmental niches, simple systems can be very successful, while complex systems may be selected against if, for example, increased energy requirements trump higher behavioral flexibility (e.g., [35]–[38]). For the evolution of intrinsic complexity investigated in this article, it is thus important to understand under which environmental conditions integrated conceptual structures become advantageous.
Overall, the results of the present simulations indicate that, given constraints on the number of elements and connections, integrated systems can have a selective advantage if the causal structure of the environment is complex. This was shown, first, by the finding that the highest fitness in the more complex tasks (2,3 and especially 4) was achieved by animats with (highly) integrated conceptual structures. By contrast, in a simpler task (Task 1), high fitness was achieved by both integrated and modular systems. Accordingly, correlations between measures of integration and fitness were low in Task 1, but increased progressively over Tasks 2–4 (Table 1, S1 Fig). The relative simplicity of Task 1 is illustrated by the rapid achievement of close to maximum fitness in most evolutionary histories and by the minimal requirement for sequential memory (in Tasks 2–4, a longer sequence of sensor inputs needs to be stored inside the animat's brain to perform adequately). Second, when Task 1 was made more difficult without changing the environment, by reducing the number of sensors and motors, animats had to rely more on sequential memory to achieve high fitness. In this case, animats that evolved highly integrated conceptual structures had once again a selective advantage.
Why is this so? Given limitations on the number of hidden elements, integrated brains can implement more functions (concepts) for the same number of elements, because they can make use of higher-order concepts, those specified by irreducible combinations of elements (see also [26]). Moreover, integrated brains with functions specified by hidden elements over hidden elements, or combinations of input, hidden, and output elements, are able to rely more on memory. Note that given an upper limit, or cost on the number of sensors, motors, and hidden elements (and the speed of interaction between them), an empirical positive lower bound of Φ will exist for higher fitness values in complex task environments, as observed for the informational measures evaluated in [7] (ISMMI, IPred, and integrated information as defined in [16]). Note also, however, that any task could, in principle, be solved by a modular brain with Φ = 0 given an arbitrary number of elements and time-steps (see in particular Fig. 21 in [15] and [39]–[41]).
Another potential advantage of integrated brains is related to degeneracy [25]. Degeneracy is the property according to which a given function can be performed by many different structures [25], [42], [43], and it is ubiquitous in biology [44]. Degenerate structures show equivalent behavior in certain contexts, but can perform different functions in different contexts. Degeneracy contrasts with redundancy, where many identical structures perform the same function under every circumstance. Systems that show high degeneracy usually are well-suited to integrating information [14], [25]. Indeed, our results are in line with higher degeneracy for animats having high Φ, both at the population level and within each individual animat brain. The number of different neural architectures, logic functions, and behaviors developed by animats with integrated brains (Φ>0) that solved Task 1 and 2 was much higher than for animats with modular brains (Φ = 0). More potential solutions with Φ>0 provide a probabilistic selective advantage for integrated structures and lead to higher variability due to neutral mutations (S3 Fig.) and more heterogeneous populations. This suggests that populations having high Φ and high degeneracy should be better at adapting rapidly to unpredictable changes in the environment and more robust to mutations, because some animats are likely to be available that are already predisposed to solve new problems.
A similar advantage is provided by degeneracy in the concepts available to each individual animat. In integrated brains, selective pressure may favor the emergence of particular concepts. However, in such brains higher order concepts will also become available at no extra cost in terms of elements or wiring, and they may prove useful to respond to novel events. How the evolution of integrated conceptual structures with high degeneracy is affected by changing environments, or by environments with multiple connected niches and coevolution of different species [45] will be the subject of future work.
To conclude, rich environments that put a premium on context-sensitivity and memory, such as competitive social situations, should favor the evolution of organisms controlled by brains containing complexes of high Φ. This is because the integrated conceptual structures specified by complexes of high Φ can accommodate a large number of functions in a way that is more economical and flexible than what can be achieved with modular or nearly-modular architectures. Moreover, since according to IIT integrated conceptual structures underlie consciousness [14], [15], [18], [23], the finding that such structures offer a selective advantage in complex environments could provide a rationale as to why and how consciousness evolved.
Animat brains consist of 8 binary elements: 2 sensors, 4 hidden elements, and 2 motors (left, right) that can loosely be referred to as neurons. The sensors are directed upwards with a space of one unit between them and activated (set to 1) if a falling block is located directly above a sensor (Fig. 1). Otherwise the sensor element is set to 0. All elements are updated from time step t to t+1 according to a transition probability matrix (TPM). In general, the TPM could be probabilistic with transition probabilities between 0 and 1. In the present work, however, the animats' TPMs are purely deterministic, i.e., transition probabilities are either 0 or 1. The brain elements can thus be considered as binary Markov variables, whose value is specified by deterministic logic gates (just as the Markov brains in [13]). Note that the elements are not limited to classic logic gates, such as ANDs, ORs, or XORs, but can potentially specify any deterministic logic function over their inputs. If only one of the motors is updated to state 1, in the next time step the animat will move one unit to the right (motor state 01) or left (motor state 10), respectively. Since no other movement was required of the animat, motor state 11 (both motors on) was chosen to be redundant with motor state 00, for which the animat will not move.
To evaluate the number of different TPMs and connectivity matrices for animats with perfect fitness in Task 1 and 2, the TPMs and connectivity matrices were compared in “normal form”, i.e., independent of the labels of their elements and only potentially causal connections were included in the analysis (meaning, hidden elements with only inputs or outputs to the rest of the system were excluded). To that end, for a given matrix all elements were permuted and the resulting permuted matrices were ordered lexicographically. The first permuted matrix was then chosen as the “normal form”.
All animat brains are initialized without connections between their elements. Connectivity evolves indirectly during adaption to the environment as outlined below, following a genetic algorithm that selects, mutates, and updates the animat's genome at each new generation. The animats' genes encode hidden Markov gates (HMGs), which in turn determine the connectivity and transition table of each brain element: each HMG has input elements, output elements, and a logic table that specifies the elements' transition table (see [6], [13] for details). In this study, the ancestral genome (generation 0) of all animats does not encode any HGM. Different from previous publications [6], [7], [13], evolution is thus not “jump-started”, which avoids random causal connections in the animats' brains, but requires more generations to reach high levels of fitness.
The animats' genomes consist of at least 1,000 and at most 20,000 loci, where each locus in the genome is an integer value ∈ [0,255]. The beginning of a gene is marked by a start codon (the consecutive loci 42 and 213), followed by two loci that respectively encode the number of inputs and outputs of one HMG. The next eight loci are used to determine where the inputs come from and the outputs go to. Because gates are allowed to have at most 4 in- and at most 4 outgoing connections, 8 loci are reserved, and used according to the 2 preceding loci. The subsequent loci encode the transition table of the HMG, determining the input and output elements and their logical relations. This encoding is robust in the sense that mutations that change the input-output structure of an HMG only add or remove the respective parts of the HMG's logic table, while the rest of the table is left intact. Encoding the connectivity and logic functions of the animats' brain elements with HMGs allows for recurrent connections between hidden elements and also self-connections. Feedback from the hidden elements to the sensors, and also from the motors to the hidden units is however prohibited by zeroing out the sensors and motors at each time-step respectively before the new sensor input arrives and after the movement was performed.
The animat is located at the bottom row of a 16×36 unit world with periodic boundary conditions (Fig. 2B). We chose the height of 36 units to allow the animats enough time to assess the direction and size of the falling blocks from each initial condition. Each animat is tested in 128 trials: all 16 initial block positions, with blocks moving to the right and left, and four potentially different block sizes. Note that in Task 1 (“catch size 1, avoid size 3”) and Task 2 (“catch size 1, avoid size 2”) the two different block sizes are thus shown 2×32 = 64 times, while in Task 3 (“catch size 1+4, avoid size 2+3”) and Task 4 (“catch size 3+6, avoid size 4+5”) each block size is shown 32 times. In each trial a block of a certain size falls from top to bottom in 36 time steps, moving 1 unit downwards and sideways always in the same direction (left or right). If at time-step 36 at least one of the animat's units overlaps with the block, it is counted as “caught”, otherwise as “avoided”.
In Task 1, sensor state S1S2 = 11 unambiguously distinguishes size 3 blocks from size 1 blocks. In all other cases, whether a block should be caught or avoided cannot be decided based on a momentary sensor input state.
An animat's fitness F at each generation is simply calculated as the percentage of successfully caught and avoided blocks out of all possible 128 test trials. Starting from a set of 100 ancestral animats without HMGs and thus without connections between elements, the animats adapt according to a genetic algorithm across 60,000 generations. At each generation, fitness is assessed for all animats in a population of 100 candidates. The most successful candidates are selected probabilistically for differential replication according to an exponential fitness measure S = 1.02F*128. For every successfully caught or avoided block the score is thus multiplied by 1.02. The 100 candidate animats are ranked according to S and selected into the next generation with a probability proportional to S and thus to their fitness (roulette wheel selection without elite). After this replication step, the new candidate pool is mutated in three different ways: a) by point mutations, which occur with a probability of p = 0.5% per locus, causing the value to be replaced by a random integer drawn uniformly from [0,…,255]; b) by deletion: with 2% probability, a sequence between 16 and 512 adjacent loci is deleted; c) by duplication: with 5% probability a sequence between 16 and 512 adjacent loci is duplicated and inserted at a random location within the animat's genome, where the size of the sequence to be deleted or duplicated is uniformly distributed in the range given. Since insertions are more likely than deletions, genomes tend to grow in size during evolution. Deletions and duplications are, however, constrained so that the genome remains between 1,000 and 20,000 loci. All genes are expressed. Some of the genes may give rise to redundant HMGs, which, however, will not be robust to mutation. Under fitness selection, the number of genes thus tends to converge to a balanced level (roughly the number of possible elements). Under random selection, only very few rapidly changing random connections between elements appear, and existing network structures decompose within less than 1,000 generations [7].
For each task, 50 evolutionary runs of 60,000 generations are performed. At the end of each evolutionary run, the line of descent (LOD) [19] of a randomly chosen animat from the final generation is traced back to its initial ancestor at generation 0. For each evolutionary run one LOD is obtained, which captures the run's particular evolutionary history. Since reproduction is asexual, without crossover, a unique LOD can be identified for an animat from the final generation. Because, moreover, all animats are part of the same niche, it makes almost no difference which animat is chosen in the final generation, since going backwards across generations their different LODs quickly coalesce to a single line [6]. We performed the full IIT analysis across each line of descent every 512 generations starting from 0.
The most recent mathematical formulation of the integrated information theory (“IIT 3.0”) is presented in detail in [15]. In the following we will summarize the main principles and measures relevant to this study, illustrated in simple examples of neuron-like logic gates mechanisms (Fig. 9).
Table 1 shows the average (nonparametric) Spearman rank correlation coefficients across all 50 LODs for all evaluated IIT measures in Task 1-4. In S1 Fig. complementary histograms are shown of the correlation coefficients of all individual LODs. Correlation coefficients were calculated based on ranked variables (i.e., using Spearman's instead of Pearson's correlation coefficients), since the amount by which fitness increases is not expected to depend linearly on any of the causal measures. Initial increases in fitness can be large, simply because initially there is more room for large improvements than at later generations where the animat already has a high percentage of fitness.
Error margins throughout this article denote SEM.
Since none of the measured variables was found to be normally distributed for all task conditions (Kolmogorov-Smirnoff test for normality) and variances between tasks differed for some of the measures, statistical differences were evaluated using a Kruskal-Wallis test, the non-parametric equivalent of a one-way ANOVA. For all statistical tests across task conditions after adaptation, measures were averaged over the last 3,000 generations (6 data points).
Task 1–4 were compared (see Fig. 3), first, taking all 50 independent LODs of each task into account, despite the lower average fitness reached in Task 3 and 4. In this set, statistical differences were found for the number of concepts, ΣφMax, and ΦMax (p = 0.001/0.002/0.016), but not for the number of MC concepts and MC elements. Second, Task 1–4 were compared at the same level of fitness, taking only a subset of LODs with high final fitness into account in Task 3 and 4 (9 and 7 fittest LODs, respectively). The respective subsets of LODs were selected as the set of fittest LODs in Task 3 and 4, whose average fitness across the last 5,000 generations was closest to that achieved on average in Task 1. Compared at the same level of fitness, all IIT measures showed statistical differences (p = 0.000/0.000/0.003/0.003/0.000 for #concepts/ΣφMax/#MC elements/#MC concepts/ΦMax).
Moreover, the standard Task 1 was compared to Task 1 with one sensor only, one motor only, and 1% sensor noise (Fig. 6–8). All measures showed significant difference (p = 0.000) when all 50 LODs of each condition were taken into account and also when a subset of LODs with high fitness was compared (again, p = 0.000 for all measures).
Differences between pairs of task conditions reported in the results section were assessed by post-hoc Mann-Whitney U tests.
Custom-made MATLAB software was used for all calculations. The program to calculate the complex of a small system of logic gates and its constellation of concepts is available under [51]. EMD calculations within the IIT program were performed using the open source fast MATLAB code of Pele and Werman [49]. The IBM SPSS software package was used for statistical analysis.
|
10.1371/journal.ppat.1006918 | Mutations in the pantothenate kinase of Plasmodium falciparum confer diverse sensitivity profiles to antiplasmodial pantothenate analogues | The malaria-causing blood stage of Plasmodium falciparum requires extracellular pantothenate for proliferation. The parasite converts pantothenate into coenzyme A (CoA) via five enzymes, the first being a pantothenate kinase (PfPanK). Multiple antiplasmodial pantothenate analogues, including pantothenol and CJ-15,801, kill the parasite by targeting CoA biosynthesis/utilisation. Their mechanism of action, however, remains unknown. Here, we show that parasites pressured with pantothenol or CJ-15,801 become resistant to these analogues. Whole-genome sequencing revealed mutations in one of two putative PanK genes (Pfpank1) in each resistant line. These mutations significantly alter PfPanK activity, with two conferring a fitness cost, consistent with Pfpank1 coding for a functional PanK that is essential for normal growth. The mutants exhibit a different sensitivity profile to recently-described, potent, antiplasmodial pantothenate analogues, with one line being hypersensitive. We provide evidence consistent with different pantothenate analogue classes having different mechanisms of action: some inhibit CoA biosynthesis while others inhibit CoA-utilising enzymes.
| The coenzyme A (CoA) biosynthetic pathway is under investigation as a target for the development of drugs aimed at several infectious agents, including malaria parasites. To synthesise CoA, the parasite scavenges the essential precursor pantothenate (vitamin B5). Several pantothenate analogues possess potent (nM) activity against the parasite, but their exact mechanism of action is unknown. We have generated mutant parasites that are resistant or hypersensitive to various pantothenate analogues. These parasites harbour mutations in a gene we now show codes for the first enzyme in the CoA biosynthesis pathway. This enzyme is not the target of the analogues, but instead converts them into antimetabolites that, depending on the analogue, either inhibit a CoA biosynthesis enzyme or downstream CoA-utilising enzymes.
| In recent years, the effort to roll back malaria has shown encouraging progress through the increased use of insecticide-treated mosquito nets, improved diagnostics and artemisinin-based combination chemotherapies (ACTs) [1]. Evidence of this includes the decreasing worldwide malaria incidence (266 million cases in 2005 down to 212 million cases in 2015) and mortality (741,000 deaths in 2005 down to 429,000 deaths in 2015) over the past decade [1]. However, there is an alarming trend of ACT-resistant parasites emerging in multiple Asian countries where the disease is endemic [2]. Recently, there have also been multiple reports of patients contracting ACT-resistant Plasmodium falciparum malaria from various African countries [3,4], which exemplify the clear risk of artemisinin resistance developing in the continent. This threat to the efficacy of ACTs highlights the requirement for a new armoury of antimalarial medicines, with several compounds representing different chemotypes entering the preclinical trial stage. However, the antimalarial drug-discovery pipeline is reliant on just a few known drug targets and the probability of successfully producing a new blood-stage medicine remains low [5]. In order to manage the threat of parasite drug resistance, there needs to be a continued effort to identify new classes of antimalarials. One metabolic pathway that has garnered recent interest for drug-development is the parasite’s coenzyme A (CoA) biosynthetic pathway [6,7].
Early seminal studies have shown that the asexual stage of intra-erythrocytic P. falciparum absolutely requires an exogenous supply of vitamin B5 (pantothenate; Fig 1) for survival [7–9]. Pantothenate is taken up by the parasite [10,11] and converted into CoA, an essential cofactor for many metabolic processes [7]. This conversion is catalysed by a series of five enzymes, the first of which is pantothenate kinase (PfPanK), an enzyme that phosphorylates pantothenate to form 4’-phosphopantothenate [11]. By performing this step, the parasite traps pantothenate within its cytosol and commits it to the CoA biosynthetic pathway [10]. The additional four steps are, in turn, catalysed by phosphopantothenoylcysteine synthetase (PfPPCS), phosphopantothenoylcysteine decarboxylase (PfPPCDC), phosphopantetheine adenylyltransferase (PfPPAT) and dephospho-CoA kinase (PfDPCK) [6]. Putative genes coding for each of the enzymes in the pathway (with several enzymes having multiple putative candidates) have been identified in the P. falciparum genome [12,13] and have also been shown to be transcribed during the intraerythrocytic stage of the parasite’s lifecycle [14]. In order to capitalise on the pathway as a potential target for drug-discovery, however, it is crucial to ascertain the exact identity of each of these putative genes. This will allow the process to become more efficient and targeted.
Investigations aimed at discovering antiplasmodial agents that act by interfering with the parasite’s CoA biosynthetic pathway identified several antiplasmodial pantothenate analogues, including pantothenol (PanOH) and CJ-15,801 [9,15,16] (see Fig 1 for structures). Subsequent studies identified pantothenamides as antiplasmodial pantothenate analogues with substantially increased potency [17–19]. Unfortunately pantothenamides are unstable in vivo because they are degraded by the serum enzyme pantetheinase [17]. Recent reports of structural optimisations of lead pantothenamides have described two compounds, N5-trz-C1-Pan (compound 1e in Howieson et al. [20]) and N-PE-αMe-PanAm (see Fig 1 for structures), that are potent antiplasmodials (with nanomolar IC50 values) and also resistant to pantetheinase-mediated degradation [20,21]. However, although these compounds have been shown to target CoA biosynthesis or utilisation, their exact mechanism(s) of action has not been elucidated.
In this study, we have used continuous drug-pressuring with PanOH or CJ-15,801 to generate a number of P. falciparum parasite lines that are several-fold resistant to these pantothenate analogues. Whole-genome sequencing revealed mutations in one of the two putative Pfpank genes of all of the clones, Pfpank1. Complementation experiments confirmed that these mutations are responsible for the resistance phenotypes. Characterisation of the effects of the mutations on parasite growth in culture and also PfPanK function, generated data consistent with the mutated gene coding for an active PanK in P. falciparum and with the gene being essential for normal parasite development during the intraerythrocytic stage of its lifecycle. Additional characterisation of the PanOH and CJ-15,801-resistant lines revealed that antiplasmodial pantothenate analogues have at least two distinct mechanisms of action, targeting CoA biosynthesis or utilisation. Both of these mechanisms can be influenced by the Pfpank1 mutations identified here. Furthermore, our study provides genetic evidence validating the importance of the metabolic activation of pantothenate analogues in the antiplasmodial activity of these compounds.
P. falciparum parasites were maintained in RPMI 1640 media supplemented with 11 mM glucose (final concentration of 22 mM), 200 μM hypoxanthine, 24 μg/mL gentamicin and 6 g/L Albumax II (referred to as complete medium) as previously described [22]. Clonal parasite populations were generated through limiting dilution cloning as reported previously [23], with modifications. Parasite lysates were prepared from saponin-isolated mature trophozoite-stage parasites as described previously [10].
Several plasmid constructs were generated through the course of this study to be used for different lines of investigations. The strategies used to generate the Pfpank1-pGlux-1, Pfpank1-stop-pGlux-1 and ΔPfpank1-pCC-1 plasmids are detailed in the SI and the primers used in these strategies are listed in S1 Table. The constructs were transfected into ring-stage parasites and positive transfectants were selected by introducing WR99210 (10 nM) [24].
The pantothenate analogues CJ-15,801 [25], N5-trz-C1-Pan [26] and N-PE-αMe-PanAm [21], used in this study were synthesised as reported previously.
The effect of various compounds on the in vitro proliferation of the different parasite lines were tested using a previously-reported SYBR Safe-based fluorescence assay [17], with minor modifications (SI). In vitro pantothenate requirement experiments were performed similarly (SI), except instead of a test compound, ring stage-infected erythrocytes were incubated in pantothenate-free complete RPMI 1640 medium (made complete as described above; Athena Enzyme Systems) supplemented with 2-fold serial dilutions of pantothenate. IC50 and SC50 (defined in the Results section) values were determined from the sigmoidal curves fitted to each data set using nonlinear least squares regression (SI).
Two independent drug-pressuring cultures were initiated for each of the pantothenate analogues, PanOH and CJ-15,801. Pressuring was initiated by exposing synchronous ring-stage Parent line parasites (10 mL culture of 2 or 4% parasitaemia and 2% haematocrit) to either analogue at the IC50 values obtained for the Parent line at the time (PanOH = 400 μM and CJ-15,801 = 75 μM). Parasites were then exposed to cycles of increasing drug-pressure that lasted about 2–4 weeks each (SI). When the pressured parasites became approximately 8 × less sensitive than the Parent line to the selecting analogues, they were cloned and cultured in the absence of the analogues for the remainder of the study.
In order to compare the fitness of the mutant clones with that of the Parent, we set up three competition cultures, each containing a mixture of one mutant line and the Parent line. Equal number of parasites (5 × 108 cells in the first experiment and 2.5 × 108 cells in the second experiment) from each line were mixed into a single culture. Aliquots (3 to 5 mL) of these cultures were immediately used for a PanOH SYBR Safe-based parasite proliferation assay (to generate Week 0 data) as described above. The cultures were then maintained under standard conditions as detailed above for a period of 6 weeks before they were used to perform another PanOH proliferation assay (to generate Week 6 data).
Next generation whole genome sequencing was performed by the Biomolecular Resource Facility at the Australian Cancer Research Foundation, the Australian National University. Samples were sequenced with the Illumina MiSeq platform with version 2 chemistry (2 × 250 base pairs, paired-end reads) Nextera XT Kit (Illumina). To determine the presence of any single nucleotide polymorphisms (SNPs) in the genomes of the drug-pressured clones, the genomic sequencing data were analysed using an integrated variant calling pipeline, PlaTyPus, as previously described [27], with minor modifications to resolve operating system incompatibility. As PlaTyPus does not detect insertions-deletions polymorphisms (“indels”), the Integrated Genome Viewer (IGV) software (Broad Institute) was used to manually inspect the gene sequences of all putative enzymes in the CoA biosynthetic pathway for indels.
To determine the proportions of parasites (i.e. Parent versus mutant clone) that make up each competition culture, we performed qPCR to measure the amount of Parent and mutant genomic DNA (gDNA) at week 0 and week 6 of the competition assay. gDNA was extracted using a DNeasy Plant Mini Kit (QIAGEN) or QIAamp DNA Blood Midi Kit (QIAGEN) following the manufacturer’s instructions, unless otherwise specified (detailed in SI). Six different primer sets were designed, each to detect the specific variants (SNPs or indel) in the wild-type and mutant Pfpank1 alleles (S1 Table). The qPCRs were performed using QuantiTect 2 × SYBR Green PCR Master Mix (QIAGEN) essentially as reported previously [28] with the following specifications: The reactions (20 μL final volume) for “Parent vs PanOH-A” and “Parent vs PanOH-B” contained 2 μL of 10 ng/μL gDNA stocks while that for “Parent vs CJ-A” contained 4 μL of 10 ng/μL gDNA stocks. The PCR program for “Parent vs PanOH-A” and “Parent vs CJ-A” entailed an initial DNA polymerase activation step (95°C for 15 min) followed by 40 cycles of denaturation (94°C for 15 s), annealing (60 or 55°C, respectively, for 30 s) and extension (72°C for 30 s) with the detection step (green channel) set during the extension step of each cycle. For “Parent vs PanOH-B” the PCR was programmed as above except that it was set for 45 cycles, the annealing step was carried out at 58°C and included an additional detection step (65°C for 15 s, green channel) in each cycle. Following each reaction, a melt curve analysis was performed starting with a temperature that is 1°C lower than the annealing temperature and increased to 99°C at 1°C/s. Control reactions that each contained defined proportions of the Parent and mutant gDNA as templates, set at the same final DNA concentration as detailed above, were included in each qPCR experiment to enable preparation of a standard curve. The standard curve was generated by plotting the threshold cycle values obtained for the control reactions against the corresponding log10 gDNA amount and fitting a straight line (y = y0 +ax, where y denotes threshold cycle value, x denotes log10 DNA amount and a denotes primer efficiency). The concentration of Parent and mutant gDNA in each sample was determined by comparing their threshold cycle against the appropriate standard curve. All reactions were performed in duplicate.
The structure of PfPanK1 minus its parasite-specific inserts was predicted by homology modeling using the AMPPNP and pantothenate-bound human PanK3 structure (PDB ID: 5KPR [29]) as a template. PfPanK1 shares 28% sequence identity with human PanK3 over the parts of the protein that have been modeled. The model was generated using the one-to-one threading module of the Phyre2 webserver (available at http://www.sbg.bio.ic.ac.uk/phyre2) [30].
Erythrocytes infected with trophozoite-stage 3D7 strain P. falciparum parasites expressing PfPanK1-GFP were observed and imaged either with a Leica TCS-SP2-UV confocal microscope (Leica Microsystems) using a 63 × water immersion lens or a Leica TCS-SP5-UV confocal microscope (Leica Microsystems) using a 63 × oil immersion lens. The parasites were imaged as fixed or live cells as described in the SI.
The Pfpank1 disruption plasmid, ΔPfpank1-pCC-1 (SI), was transfected into wild-type 3D7 strain P. falciparum, and positive transfectants were selected as described above. P. falciparum parasites have previously been shown to survive equally well in a pantothenate-free complete RPMI 1640 medium supplemented with ≥100 μM CoA as compared to standard complete medium, consistent with them having the capacity to take up exogenous CoA, hence bypassing the need for any PfPanK activity [7]. Therefore, to support the growth of any Pfpank1 gene-disrupted parasites generated with the ΔPfpank1-pCC-1 construct, parasites were continuously maintained in complete medium supplemented with 100 μM CoA following transfection. Positive and negative selection steps (with WR99210 and 5-fluorocytosine (5-FC), respectively) were performed to isolate ΔPfpank1-pCC-1-transfectant parasites in which the double crossover homologous recombination had occurred (detailed in SI).
gDNA samples (~2 μg) extracted from ΔPfpank1-pCC-1-transfectant parasites isolated through the positive and negative selection steps were digested with the restriction enzyme AflII (New England Biolabs), before being analysed by Southern blotting using the digoxigenin (DIG) system (Roche) according to the Roche DIG Applications Manual for Filter Hybridisation. The results from the Southern blot were confirmed with PCR (the primers used are shown in S1 Table).
The phosphorylation of [14C]pantothenate by parasite lysates prepared from the Parent and mutant clonal lines was measured using Somogyi reagent (which precipitates phosphorylated compounds from solution) as outlined previously [31], with some modifications (detailed in SI).
Cultures of predominantly trophozoite-stage P. falciparum-infected erythrocytes (Parent line) were concentrated to >95% parasitaemia using magnet-activated cell sorting as described elsewhere [32]. Following recovery, trophozoite-infected erythrocytes were treated with N5-trz-C1-Pan (1 μM) or a solvent control (0.01% v/v DMSO) before the metabolites in these samples were extracted and processed for liquid chromatography-mass spectrometry (LC-MS) analysis. Metabolite samples were analysed by LC-MS, using a Dionex RSLC U3000 LC system (ThermoFisher) coupled with a high-resolution, Q-Exactive MS (ThermoFisher), as described previously [33] (detailed in SI). LC-MS data were analysed in a non-targeted fashion using the IDEOM workflow, as described elsewhere [34]. Unique features identified in N5-trz-C1-Pan-treated samples were manually assessed by visualising high resolution accurate mass LC-MS data with Xcalibur Quanbrowser (ThermoFisher) software.
To measure PfPanK1 levels, parasite samples were prepared from Parent and mutant cultures as previously described with minor modifications [35]. Briefly, mature trophozoites were saponin-isolated from a culture of infected erythrocytes (10% parasitaemia, 2% haematocrit in 30 mL) by resuspending the pellet in 1 × phosphate buffered saline (PBS; Sigma-Aldrich) containing 0.05% (w/v) saponin, 1 × complete mini protease inhibitor cocktail (Roche), 20 mM sodium fluoride and 0.1 mM sodium orthovanadate. The isolated trophozoites were pelleted (2,000 × g, 8 min), washed (15,850 × g, 30 s, supernatant discarded after each wash) once in 1 mL of the above solution excluding saponin, and twice with 1 mL 1 × PBS. The trophozoite pellet was then stored at −80°C until required. For sample processing, a total of 500 μg of protein, accurately determined using the Pierce BCA protein assay kit (ThermoFisher), was incubated overnight with sequencing-grade trypsin (1:50 dilution; Promega). On the following day, trypsin activity was quenched using 5% formic acid (FA) and the detergent (sodium deoxycholate) used for protein solubilisation was removed. The samples were then dried and resuspended in 20 μL of 2% (v/v) acetonitrile (ACN) and 0.1% (v/v) FA for LC-MS/MS analysis. For facilitating retention-time (RT) alignments among samples, a RT kit [36] (iRT kit, Biognosys) was spiked into all samples (1:20 dilution).
LC-MS/MS was carried out as described previously [35], with minor modifications. Briefly, samples were loaded at a flow rate of 15 μL/min onto a reversed-phase trap column (100 μm × 2 cm; Acclaim PepMap media, Dionex), which was maintained at a temperature of 40°C. Peptides were then eluted from the trap column at a flow rate of 0.25 μL/min through a reversed-phase capillary column (75 μm × 50 cm; LC Packings, Dionex). The HPLC gradient was set to 158 min using a gradient that reached 30% ACN after 123 min, then 34% ACN after 126 min, 79.2% ACN after 131 min and 2% ACN after 138 min, at which it was maintained for a further 20 min. The mass spectrometer was operated in a data-independent acquisition (DIA) mode with a 25-fixed-window setup of 24 m/z effective precursor isolation over the m/z range of 375–975 Da.
For Spectronaut processing, raw files were loaded into Spectronaut (version 11, Biognosys) with an in-house P. falciparum-infected red blood cell library and default settings. Briefly, RT prediction type was set to dynamic iRT and the correction factor for the window was set to one. Mass calibration was set to local mass calibration. Interference correction was on MS2 level. The false discovery rate was set to 1% at peptide precursor level. For quantification, the interference correction was activated and a cross run normalisation was performed using the total peak area as the normalisation base. A significance filter of 0.01 was used. The peptides used in the identification and quantitation of PfPanK1 and the four control proteins are listed in S2 Table.
Statistical analysis of means was carried out with unpaired, two-tailed, Student’s t tests using GraphPad 6 (GraphPad Software, Inc) from which the 95% confidence interval of the difference between the means (95% CI) was obtained. All regression analysis was done using SigmaPlot version 11.0 for Windows (Systat Software, Inc).
The 3D7 P. falciparum strain was cloned through limiting dilution, and a single parasite line (henceforth referred to as the Parent line) was used to generate all of the subsequent lines tested in this study (unless otherwise specified). This was done to ensure that all of the parasite lines generated during the course of this study would share a common genetic background. Using the Parent line, three independent drug-pressuring cultures were set up (two with PanOH and one with CJ-15,801). When these parasites had attained approximately 8-fold decrease in sensitivity (~11–13 weeks of continuous pressuring), they were subsequently cloned by limiting dilution and maintained in the absence of drug pressure. In this manner, three parasite clones were generated: PanOH-A and PanOH-B were generated from the two independent PanOH-pressured cultures while CJ-A was generated from the CJ-15,801-pressured culture. The clones are significantly resistant (95% confidence interval (CI) exclude 0) to the pantothenate analogues they were pressured with. The 50% inhibitory concentration (IC50) values of PanOH against PanOH-A and PanOH-B, and the IC50 value of CJ-15,801 against CJ-A are approximately 7–8-fold higher than those measured against the Parent line (Fig 2A and 2B and S3 Table). Significant cross-resistance towards the other pressuring analogue was observed for these clones, as compared to the Parent line (95% CI exclude 0). The PanOH-A and PanOH-B lines were found to be 4–6-fold less sensitive to CJ-15,801 while CJ-A was found to be 13-fold less sensitive to PanOH (Fig 2A and 2B and S3 Table). To ensure that the clones did not develop a general drug-resistance phenotype during the selection process, we tested their sensitivity to chloroquine, an antiplasmodial with a mechanism of action that is unrelated to the parasite’s CoA biosynthetic pathway [37]. We found that all of the drug-pressured lines have chloroquine IC50 values that are indistinguishable from that of the Parent line (95% CI include 0; Fig 2C and S3 Table).
The PanOH and CJ-15,801 resistance phenotypes observed in the clones were stable for several months of continuous culture in the absence of the pressuring analogue (≥ 3 months), consistent with a genetic alteration in these parasites. To determine the mutation(s) responsible for these phenotypes, gDNA was extracted from each clone and subjected to whole genome sequencing. All of the drug-resistant clones were found to harbour a unique mutation in the putative pantothenate kinase gene, Pfpank1 (PF3D7_1420600), as shown in Fig 3A. Other non-synonymous mutations were detected for each clone (S4 Table) but we did not find another gene that is mutated in all three clones. The mutation found in the Pfpank1 of PanOH-A results in the substitution of Asp507 for Asn. The other two drug-resistant clones have a mutation at position 95 of the protein: the Pfpank1 of PanOH-B harbours a deletion of an entire codon leading to a loss of Gly from the PfPanK1 protein, while the PfPanK1 of CJ-A has a Gly to Ala substitution. Since the structure of PfPanK1 has not yet been resolved, we generated a three-dimensional model in order to map the mutations within the enzyme. Fig 3B shows a PfPanK1 model structure (pink) based on the solved structure of human PanK3 in complex with adenylyl-imidodiphosphate (AMPPNP) and pantothenate (PDB ID: 5KPR), overlaid on this structure (blue). The spheres shown in the model indicate the positions of the mutated residues, while the bound AMPPNP and pantothenate indicate the active site of the enzymes. Although the mutations are far apart in the primary amino acid sequence of PfPanK1, they are positioned in closer proximity to each other in the folded protein and are situated adjacent to the active site.
To confirm that the resistance phenotypes observed for the clones are directly caused by the mutations in Pfpank1, each clone was transfected with an episomal plasmid (Pfpank1-stop-pGlux-1) that enables the parasites to express the wild-type copy of Pfpank1 (in addition to the endogenous mutated copy). These complemented lines are indicated with a superscripted “+WTPfPanK1”. From Fig 4 (and S3 Table), it can be observed that the complemented mutant clones (grey bars) are significantly less resistant to PanOH (Fig 4A) and CJ-15,801 (Fig 4B) compared to the non-complemented mutant clones (black bars; 95% CI exclude 0). As expected, the relative sensitivity of the mutant clones to chloroquine is unchanged by the presence of the PfPanK1-encoding plasmid (95% CI include 0; Fig 4C). Transfection of the PfPanK1-encoding plasmid into the Parent line did not alter its sensitivity to PanOH, CJ-15,801 or chloroquine (95% CI include 0; S3 Table). These data are consistent with the mutations observed in Pfpank1 being responsible for the resistance phenotype observed in the mutant clones. An alternative strategy to confirm that the mutant PfPanK1 is responsible for the observed phenotypes would be to express the mutant forms of PfPanK1 in the Parent line. This has not, however, been attempted.
We performed a competition assay to determine whether the Pfpank1 mutations impart a fitness cost to the parasite clones. Each of the competition cultures was set up by mixing an equal number of parasites from the Parent line and one of the mutant clonal lines, and was maintained under standard conditions for a period of 6 weeks (Fig 5A). The sensitivity of each competition culture to PanOH was tested on the day the lines were mixed (Week 0) and again at the end of the 6-week period (Week 6). As expected, each Week 0 (dashed line) PanOH dose-response curve is between those obtained for the Parent and the respective mutant clone (dotted lines). A shift of the dose-response curve obtained at week 6 (solid line) towards the dose-response curve of the Parent line would indicate that the mutant PfPanK1 imparts a fitness cost on the clone. As shown in Fig 5B, the Week 6 curve for the PanOH-A competition culture only exhibits a marginal leftward shift from Week 0, whereas those for PanOH-B (Fig 5C) and CJ-A (Fig 5D) exhibits a more substantial shift, almost reaching the dose-response curve of the Parent line (dotted line, white circles). These trends are supported by qPCR analysis designed to determine the proportions of mutant versus Parent gDNA at week 0 and week 6 (S1 Fig). These results are consistent with the mutations at position 95 in the PfPanK1 of PanOH-B and CJ-A having a greater negative impact on the in vitro growth of the parasites. We also investigated the importance of PfPanK1 expression for parasite growth by attempting to disrupt the Pfpank1 locus in wild-type 3D7 parasites through homologous recombination (S2A Fig). However, using Southern blots, we failed to detect the presence of transfectants with the expected gene-knockout integration event (S2B Fig), consistent with this gene being essential during the parasite’s intraerythrocytic stage. The lack of integration of the knockout construct into the PfPanK1 locus, even in a small sub-population of the parasite culture, was confirmed using PCR (S2C Fig).
The phosphorylation of radiolabelled pantothenate by lysates prepared from each of the mutant clones and the Parent line was measured to determine if the mutations in the putative Pfpank1 gene affect PanK activity, and hence whether the gene codes for a functional PanK. As shown in Fig 6, at the end of the 75 min time-course, the lysate prepared from PanOH-A phosphorylated approximately 3 × more [14C]pantothenate than the lysate prepared from the Parent line, while the lysate of PanOH-B generated about 3 × less phosphorylated [14C]pantothenate compared to the Parent line. By comparison, the lysate prepared from CJ-A only produced a small amount of phosphorylated [14C]pantothenate in the same time period. The inset in Fig 6 demonstrates that PanK activity could be detected in CJ-A lysates when the experiment was carried out in the presence of a 100-fold higher pantothenate concentration (200 μM) and for an extended time (420 min). These observations provide strong evidence that the Pfpank1 gene codes for a functional PanK. Further, we generated from the wild-type 3D7 strain a transgenic parasite line that episomally expresses a GFP-tagged copy of PfPanK1 in order to localise the protein within the parasite. We found that PfPanK1 is largely localised throughout the cytosol of trophozoite-stage parasites, and is not excluded from the nucleus (Fig 7).
To investigate further the effects of the PfPanK1 mutations present in the PanOH- and CJ-15,801-resistant clones, we analysed the PanK activity profiles using lysates prepared from each mutant and the Parent, and determined their kinetic parameters from the Michaelis-Menten equation (Fig 8). The apparent pantothenate Km values of the mutant clones are 26–609-fold higher (95% CI exclude 0) than that of the Parent line. The maximal velocity (Vmax) of pantothenate phosphorylation by lysates prepared from PanOH-A, PanOH-B and CJ-A are also significantly higher (95% CI exclude 0) than that of the Parent. To eliminate the possibility that the elevated Vmax values are due to increased PfPanK1 levels in the mutant parasites, we used DIA-MS to compare the PfPanK1 level in each of the mutant lines with that of the Parent parasite. We found that PfPanK1 levels are unchanged in the CJ-A line (95% CI include 0) and are only elevated by 26 ± 5% and 27 ± 4% (mean ± SEM) in PanOH-A and PanOH-B, respectively (S3 Fig). This modest (or lack of an) increase in PfPanK1 levels is insufficient to explain the larger increases in Vmax values observed in lysates prepared from the mutant lines, as the increase in protein levels would need to be ~20-fold for PanOH-A and ~2-fold for PanOH-B to account for the altered Vmax values. We have therefore not adjusted the observed Vmax values to account for these small increases in the PfPanK1 levels. We also calculated the PfPanK relative specificity constant for each parasite line, which indicates the catalytic efficiency of each variant of PfPanK relative to that of the Parent line. The relative specificity constant obtained for PanOH-A (0.74 ± 0.04, mean ± SEM) is not significantly different (95% CI include 0) from that of the Parent. However, those of PanOH-B (0.058 ± 0.004, mean ± SEM) and CJ-A (0.019 ± 0.005, mean ± SEM) are significantly lower (95% CI exclude 0). These data are consistent with all three PfPanK1 mutations observed reducing the enzyme’s affinity for pantothenate, although the associated increase in the enzyme Vmax compensates for the reduced affinity: fully in the PanOH-A clone, but to a much lesser extent in PanOH-B and CJ-A (resulting in a 17-fold and 52-fold reduction in the enzyme’s catalytic efficiency, respectively).
An extracellular supply of pantothenate is essential for the in vitro proliferation of the intraerythrocytic stage of P. falciparum [8]. Given the impact that the PfPanK1 mutations have on PanK activity (Fig 8), we set out to determine whether a higher extracellular concentration of pantothenate is required to support the proliferation of the different mutant clones relative to that required by the Parent line. As observed in Fig 9A, the proliferation of the Parent line (white circles) increased as the extracellular pantothenate concentration was increased, reaching the 100% control level (parasites maintained in the presence of 1 μM pantothenate, the concentration of pantothenate in the complete RPMI medium used to maintain all of the parasite cultures, which is within the physiologically relevant range in human blood [38,39]) at approximately 100 nM. In order to compare the extracellular pantothenate requirement between the different lines, we determined the SC50 (50% stimulatory concentration; i.e. the concentration of pantothenate required to support parasite proliferation to a level equivalent to 50% of the control level) values for the mutants (with and without complementation) and Parent. From Fig 9B, it can be seen that the SC50 values of PanOH-A and PanOH-B are not different from that of the Parent line (95% CI include 0). Conversely, as illustrated by the rightward shift in its dose-response curve (black diamonds, Fig 9A), the pantothenate SC50 of CJ-A is approximately 3-fold higher than that of the Parent (Fig 9B; 95% CI = 2.7 to 30.9). Furthermore, consistent with the data from the complementation experiments (Fig 4), the SC50 value of CJ-A+WTPfPanK1 (grey diamonds, Fig 9A) is comparable to that of the Parent line and also the control line, Parent+WTPfPanK1 (Fig 9B; 95% CI include 0), indicating that the episomal expression of wild-type PfPanK1 is sufficient to reverse the phenotypic effects of the mutation.
To determine whether the resistance of the clones to PanOH and CJ-15,801 extends to other pantothenate analogues, we tested the mutant clones against the two recently-described, modified pantothenamides with potent antiplasmodial activities, namely N5-trz-C1-Pan [20] and N-PE-αMe-PanAm [21]. We found that PanOH-A is 3-fold more sensitive to N5-trz-C1-Pan, PanOH-B is 2-fold more resistant and CJ-A is 9-fold more resistant when compared to the Parent line (95% CI exclude 0; Fig 10A and S5 Table, left side). Similarly, we found that relative to the Parent line, PanOH-A was more sensitive to N-PE-αMe-PanAm (~2-fold), while CJ-A is 2-fold more resistant (95% CI exclude 0; Fig 10B and S5 Table, left side). The sensitivity of PanOH-B to N-PE-αMe-PanAm was statistically indistinguishable from that of the Parent line (95% CI = -0.046 to 0.005; Fig 10B and S5 Table, left side). These results indicate that PfPanK1 can influence the sensitivity of the parasite to multiple antiplasmodial pantothenate analogues. Remarkably, the mutation at position 507 of the PfPanK1 in PanOH-A makes the parasite resistant to the antiplasmodial activity of certain pantothenate analogues (PanOH and CJ-15,801) whilst at the same time hyper-sensitises the parasite to pantothenate analogues of a different class (modified pantothenamides, N5-trz-C1-Pan and N-PE-αMe-PanAm).
Previous work has shown that, in bacteria, pantothenamides are metabolised by the CoA biosynthetic pathway to form CoA antimetabolites [40], consistent with PanK activity being important for metabolic activation of pantothenamides. Additionally, it has been reported recently that pantothenamides are also phosphorylated by the PanK in P. falciparum [41], in line with their metabolic activation in bacteria. We therefore set out to determine whether the modified, pantetheinase-resistant, pantothenamides are metabolised and to what extent. In order to do so, we exposed intact Parent line P. falciparum-infected erythrocytes to N5-trz-C1-Pan (at ~10 × the IC50 for 4 h) and subjected lysates from the treated (and control) samples to untargeted LC-MS. Among the metabolites extracted from parasite-infected erythrocytes treated with N5-trz-C1-Pan (but not untreated control samples) were molecules with masses corresponding to phosphorylated N5-trz-C1-Pan ([M-H]- m/z 378.1668), a dephospho-CoA analogue of N5-trz-C1-Pan ([M-2H]2- m/z 353.6099) and a CoA analogue of N5-trz-C1-Pan ([M-H]- m/z 787.1858) as shown in S4 Fig. This is consistent with N5-trz-C1-Pan being metabolised within infected erythrocytes to generate a CoA antimetabolite.
Lastly, we investigated the ability of the different pantothenate analogues to inhibit the phosphorylation of [14C]pantothenate by parasite lysates prepared from the various mutant clones. These data are shown in Fig 10C–10F and S5 Table, right side. All of the analogues tested are significantly less effective (95% CI exclude 0) at inhibiting pantothenate phosphorylation by lysates generated from the mutant clones compared to their ability to inhibit pantothenate phosphorylation by lysates prepared from the Parent line. The exception being N-PE-αMe-PanAm, which did not reach statistical significance when tested against lysate prepared from PanOH-A. Additionally, the effectiveness of the analogues at inhibiting pantothenate phosphorylation by lysates prepared from the mutant lines all observe the following order: PanOH-A > PanOH-B > CJ-A.
When PfPanK1 is compared to type II PanKs from other organisms in a multiple protein sequence alignment, the three nucleotide-binding motifs characteristic of this superfamily can be seen to be conserved in PfPanK1, consistent with it being a functional PanK [7]. However, biochemical confirmation of its putative function as a PanK has not been demonstrated. In this study, we show that mutations in the putative PfPanK1 lead to substantial changes in pantothenate kinase activity (Fig 8) providing, for the first time, biochemical evidence that the cytosolic PfPanK1 (Fig 7) is a functional pantothenate kinase. Furthermore, the fact that multiple independent experiments aimed at generating parasites resistant to PanOH and CJ-15,801 always selected for mutations in PfPanK1 (Fig 3A and S4 Table) is consistent with the kinase being the primary PanK involved in the metabolic activation of pantothenate analogues, at least during the intraerythrocytic stage. Although it is intriguing that PfPanK1 is not excluded from the nucleus, this phenomenon has been reported for some PanKs in other organisms [42].
Although it is clear that the PfPanK1 residues at position 95 and 507 are required for normal PfPanK1 activity (Fig 8), and the PfPanK1 model structure (Fig 3B) shows that both residues are situated adjacent to the enzyme active site, their exact role(s) in modifying the activity of the protein is less obvious. The Gly residue at position 95 is conserved in eukaryotic PanKs [7] and is the residue at the cap of the α2-helix in the inactive conformation of the protein (S5A Fig). One possibility is that the change to Ala at this position could affect the structure of the helix and consequently the overall stability of the protein (at least when the protein is in the inactive state), as Gly has been shown to be much better than Ala at conferring structural stability when located at the caps of helices [43]. Alternatively, Gly residues have been shown to be present at a higher frequency in the active sites of some enzymes where they likely confer the flexibility to alternate between open and closed conformations [44]. Although the Gly95 residue is not part of the PfPanK1 active site per se, it is within close proximity to the site (Fig 3B), and the α2-helix certainly undergoes a conformational change when PanK switches from the inactive conformation to the active one [29]. This conformational change is demonstrated in S5A Fig, which shows an overlay of the human PanK3 crystal structures in the active [29] and inactive [45] state (Gly95 in PfPanK1 is equivalent to Gly117 in the human enzyme). It is worth noting that acetyl-CoA (an inhibitor of the enzyme) can only be accommodated in the binding site when the enzyme is in the inactive state, as when the enzyme is in the active state, the α2-helix encroaches on the space occupied by acetyl-CoA (S5A Fig). This lends further support to the importance of this helix for the enzyme to transition from the inactive to the active state (and vice versa) and, therefore, to the role that Gly95 could play in this process. Either way, these suggestions may explain why a mutation at this position has a greater impact on PfPanK1 function than the mutation at position 507. The Asp residue at position 507 is replaced by a different amino acid (Glu) in most other eukaryotic PanKs, although they are both negatively-charged [7]. A substitution to the uncharged Asn could disrupt any important salt-bridges or hydrogen bonds with the residue at this position. In human PanK3, the amino acid equivalent to Asp507 is Glu354. In the enzyme’s inactive state, Glu354 is within ionic bonding distance of Arg325, which in turn interacts with the 3’-phosphate of acetyl-CoA (S5B Fig) [45] and may therefore stabilise the inactive state of the enzyme. Conversely, in the enzyme’s active state, Glu354 and Arg325 are not within ionic bonding distance (S5B Fig). If Asp507 in PfPanK1 plays a similar role to that proposed for Glu354 in human PanK3, the change at position 507 to an Asn (abolishing the negative charge) could prevent stabilisation of the inactive state, providing an explanation for the increased activity observed in PanOH-A (Fig 6). Determining the crystal structure of PfPanK1 bound with pantothenate may provide a better understanding of the roles these residues play in PanK function.
It has been established that almost all of the 4’-phosphopantothenate found in P. falciparum-infected erythrocytes is generated within the parasite by PfPanK as part of its metabolism into CoA [10], which is in line with PfPanK activity being essential for the parasite’s survival. In the present study, our inability to knock out PfPanK1 (S2 Fig) is consistent with the protein being essential for the intraerythrocytic stage of P. falciparum, although we cannot exclude the unlikely possibility that the regions we targeted for the required double-crossover recombination event are genetically intractable. Our observation that clones PanOH-B and CJ-A, which harbour mutations at position 95 of PfPanK1, can be outcompeted by the Parent parasites in competition assays over approximately 20 intraerythrocytic cycles (Fig 5) is consistent with the mutations incurring a fitness cost. This, in turn, indicates that PfPanK1 is essential for normal parasite development, at least during the blood stage of its lifecycle. It was also found here that clone CJ-A requires an approximately 3-fold higher extracellular pantothenate concentration in order to survive (Fig 9), coinciding with this clone having the PfPanK with the highest Km. This is not surprising given the reported importance for the substrate concentration to exceed the enzyme Km for optimal enzyme efficiency [46,47]. More importantly, the requirement by this clone for a higher concentration of extracellular pantothenate is also congruent with PfPanK1 being essential for the normal development of P. falciparum during its asexual blood stage.
Our observation that PfPanK1 is essential for normal parasite growth during the blood stage is at odds with recent reports that show both PanK1 and PanK2 from Plasmodium yoelli and Plasmodium berghei are dispensable during the blood stage of those parasites [48–50]. However, both of these murine malaria parasite species preferentially infect reticulocytes [51,52]. Unlike the mature erythrocytes preferred by P. falciparum, reticulocytes have been shown to provide a rich pool of nutrients for the parasite, allowing the murine parasites to survive metabolic or genetic changes that would have been deleterious in P. falciparum [53]. It is therefore conceivable that unlike the condition faced by P. falciparum, the reticulocyte-residing parasites are able to salvage sufficient CoA or CoA intermediates from the host cell for their survival, rendering the two PanK proteins dispensable during their intraerythrocytic stage, a possibility acknowledged by the authors of the P. berghei study [49].
We have presented data consistent with the observed PfPanK1 mutations being the genetic basis for the PanOH and CJ-15,801 resistance phenotypes observed in all of the drug-pressured clones we generated (Fig 4). As shown by the data presented in Fig 10, both PanOH and CJ-15,801 inhibited pantothenate phosphorylation by the mutated PfPanK1 proteins less effectively than their inhibition of pantothenate phosphorylation by the wild-type PfPanK1 (the order of their IC50 values is Parent < PanOH-A < PanOH-B < CJ-A). Importantly, this order is also reflected in the level of resistance of the mutant clones to these two analogues, although the magnitude is not preserved (Fig 2 and S3 Table). These data are consistent with PfPanK1 being involved in the antiplasmodial activity of these analogues either as a target or a metabolic activator. Previous studies have demonstrated that the antiplasmodial activity of both PanOH and CJ-15,801 involves the inhibition of pantothenate phosphorylation by PfPanK [9,15]. More specifically, PanOH has been shown to inhibit PfPanK-mediated pantothenate phosphorylation by serving as its substrate [54]. CJ-15,801 is also likely to be a substrate of PfPanK, especially since it has been shown to be phosphorylated by the S. aureus PanK [55], another type II PanK [56]. In addition, the same study showed that the second enzyme in the CoA biosynthetic pathway, PPCS, subsequently accepts phosphorylated CJ-15,801 as a substrate, and performs the first step of the PPCS reaction (cytidylylation) on it. However, unlike what happens with 4’-phosphopantothenate, this cytidylylated phospho-CJ-15,801 acts as a tight-binding, dead-end inhibitor of the enzyme [55]. A separate study also concluded that PanOH targets the PPCS enzyme in Escherichia coli and Mycobacterium tuberculosis [57]. Based on the data generated in this study and the published reports that PanOH and CJ-15,801 both inhibit PPCS in other systems, we propose a similar mechanism of action for these compounds in P. falciparum, whereby they are phosphorylated by PfPanK1 and subsequently block PfPPCS as dead-end inhibitors (Fig 11). The overlapping pattern of cross-resistance between the two compounds (Fig 2) is also in line with them having a similar mechanism of action. We propose that the observed resistance to PanOH and CJ-15,801 is due to the mutated PfPanK1 having a reduced capacity to phosphorylate these analogues relative to pantothenate. This would have the effect of reducing the amount of phosphorylated PanOH or CJ-15,801 generated relative to 4’-phosphopantothenate, thereby allowing the parasites to survive at higher concentrations of the drugs. Furthermore, our observation that the mutant clones have comparable levels of PanOH and CJ-15,801 resistance (Fig 2 and S3 Table), despite having PfPanK1 proteins of vastly different efficiency (Fig 8), is likely due to the pathway flux control at the PfPPCS-mediated step in the CoA biosynthetic pathway of P. falciparum, as shown in a previous study [32].
N5-trz-C1-Pan and N-PE-αMe-PanAm are pantothenamide-mimics that harbour modifications designed to prevent them from being substrates of pantetheinase, thereby preventing their degradation: N-PE-αMe-PanAm is methylated at the α-carbon [21] while N5-trz-C1-Pan harbours a triazole instead of the labile amide [20]. Our LC-MS data clearly show that N5-trz-C1-Pan is converted into a CoA antimetabolite (S4 Fig), and it is therefore likely to go on to inhibit/inactivate CoA-utilising enzymes, killing the parasite. Such a mechanism has previously been put forward to explain the antibiotic activity of two prototypical pantothenamides (N5-Pan and N7-Pan) whereby the compounds are phosphorylated by PanK and subsequently metabolised by PPAT and DPCK to generate analogues of CoA (ethyldethia-CoA and butyldethia-CoA) [40,58,59]. These CoA analogues then mediate their antibacterial effect/s primarily by inhibiting/inactivating CoA-requiring enzymes and acyl carrier proteins [40,58,59]. Similarly, phospho-N5-trz-C1-Pan is not expected to interact with PfPPCS because it lacks the carboxyl group (Fig 1) required for the nucleotide activation by nucleotide transfer [55]. It is therefore expected to bypass the PfPPCS and PfPPCDC steps of the CoA biosynthetic pathway on its way to being converted into the CoA antimetabolite version of N5-trz-C1-Pan (Fig 11).
Furthermore, the antiplasmodial activity rank order of N5-trz-C1-Pan against the various mutant clones is very similar to that of N-PE-αMe-PanAm – with PanOH-A being hypersensitive to both compounds, CJ-A resistant to both and PanOH-B, by comparison, exhibiting only small changes in sensitivity to the two compounds (Fig 10A and 10B) – and is starkly different to those of PanOH and CJ-15,801 (Fig 2). These are congruent with (i) the antiplasmodial mechanism of action of N-PE-αMe-PanAm being similar to that of N5-trz-C1-Pan and (ii) the antiplasmodial mechanism of action of N-PE-αMe-PanAm and N5-trz-C1-Pan being different to that of PanOH and CJ-15,801.
The order of antiplasmodial activity of N5-trz-C1-Pan and N-PE-αMe-PanAm against the mutant clones can be explained on the basis of (i) the difference in the rate of PfPanK1 activity in the various clones at the concentrations of pantothenate and N5-trz-C1-Pan / N-PE-αMe-PanAm used (Fig 6), (ii) the fact that the PfPPCS-mediated step imposes pathway flux control [32] and (iii) the fact that this pathway flux control is bypassed by N5-trz-C1-Pan (and almost certainly also by N-PE-αMe-PanAm) en route to its conversion into a CoA antimetabolite. As seen in Fig 6, in the presence of 2 μM pantothenate (a similar concentration to the 1 μM present in the antiplasmodial assay), the pantothenate phosphorylation rate of the different clones has the following rank order: PanOH-A > Parent > PanOH-B > CJ-A, approximately the inverse of the antiplasmodial IC50 values of N5-trz-C1-Pan and N-PE-αMe-PanAm (described above). Therefore, PanOH-A, for example, would be expected to generate more 4’-phosphopantothenate and phosphorylated N5-trz-C1-Pan (or N-PE-αMe-PanAm), based on the assumption that the mutation also leads to increased phosphorylation activity towards the pantothenate analogues (point i). Whilst the increased levels of 4’-phosphopantothenate would not be expected to result in a concomitant increase in CoA (due to the pathway flux control at PfPPCS; point ii), the increased production of phospho-N5-trz-C1-Pan would result in increased levels of the N5-trz-C1-Pan CoA antimetabolite as the flux control step is bypassed (point iii). This would explain the increased sensitivity of PanOH-A to both N5-trz-C1-Pan and N-PE-αMe-PanAm and also the sensitivity rank order of the other parasite lines.
In conclusion, our study confirms for the first time that PfPanK1 functions as the active pantothenate kinase in the asexual blood stage of P. falciparum. Our data show that the sites of mutation in PfPanK1 reported here are important residues for normal PfPanK function and are essential for normal intraerythrocytic parasite growth, although further structural and functional studies are required to elucidate their exact role(s). Furthermore, we propose that following phosphorylation by PfPanK1, PanOH and CJ-15,801 compete with 4’-phosphopantothenate and serve as dead-end inhibitors of PfPPCS (depriving the parasite of CoA). In contrast, N-PE-αMe-PanAm and N5-trz-C1-Pan are further metabolised (by PfPPAT and PfDPCK) into CoA analogues that kill the parasite by inhibiting CoA-utilising metabolic processes. Finally, we provide the first genetic evidence consistent with pantothenate analogue activation being a critical step in their antiplasmodial activity.
|
10.1371/journal.ppat.1003231 | Enterovirus 71 Protease 2Apro Targets MAVS to Inhibit Anti-Viral Type I Interferon Responses | Enterovirus 71 (EV71) is the major causative pathogen of hand, foot, and mouth disease (HFMD). Its pathogenicity is not fully understood, but innate immune evasion is likely a key factor. Strategies to circumvent the initiation and effector phases of anti-viral innate immunity are well known; less well known is whether EV71 evades the signal transduction phase regulated by a sophisticated interplay of cellular and viral proteins. Here, we show that EV71 inhibits anti-viral type I interferon (IFN) responses by targeting the mitochondrial anti-viral signaling (MAVS) protein—a unique adaptor molecule activated upon retinoic acid induced gene-I (RIG-I) and melanoma differentiation associated gene (MDA-5) viral recognition receptor signaling—upstream of type I interferon production. MAVS was cleaved and released from mitochondria during EV71 infection. An in vitro cleavage assay demonstrated that the viral 2A protease (2Apro), but not the mutant 2Apro (2Apro-110) containing an inactivated catalytic site, cleaved MAVS. The Protease-Glo assay revealed that MAVS was cleaved at 3 residues between the proline-rich and transmembrane domains, and the resulting fragmentation effectively inactivated downstream signaling. In addition to MAVS cleavage, we found that EV71 infection also induced morphologic and functional changes to the mitochondria. The EV71 structural protein VP1 was detected on purified mitochondria, suggesting not only a novel role for mitochondria in the EV71 replication cycle but also an explanation of how EV71-derived 2Apro could approach MAVS. Taken together, our findings reveal a novel strategy employed by EV71 to escape host anti-viral innate immunity that complements the known EV71-mediated immune-evasion mechanisms.
| Enterovirus 71 (EV71) is the causative pathogen of hand, foot, and mouth disease (HFMD). Since the 2008 outbreak of HFMD in Fuyang, Anhui province, China, HFMD has been a severe public health concern affecting children. The major obstacle hindering HFMD prevention and control efforts is the lack of targeted anti-viral treatments and preventive vaccines due to the poorly understood pathogenic mechanisms underlying EV71. Viral evasion of host innate immunity is thought to be a key factor in viral pathogenicity, and many viruses have evolved diverse antagonistic mechanisms during virus-host co-evolution. Here, we show that EV71 has evolved an effective mechanism to inhibit the signal transduction pathway leading to the production of type I interferon, which plays a central role in anti-viral innate immunity. This inhibition is carried out by an EV71-encoded 2A protease (2Apro) that cleaves MAVS—an adaptor molecule critical in the signaling pathway activated by the viral recognition receptors RIG-I and MDA-5—to escape host innate immunity. These findings provide new insights to understand EV71 pathogenesis.
| When viruses infect host cells, the innate immune response is activated as the first line of defense against viral invasion. Pathogen associated molecular patterns (PAMPs) are sensed by host pattern recognition receptors (PRRs), resulting the expression of type I interferon and proinflammatory cytokines [1], [2]. These cytokines can induce an anti-viral state in the host cells and initiate host adaptive immunity, leading to limitation or clearance of the viral infection. Anti-viral innate immunity can be roughly divided into three phases: (i) the initiation phase, where PRRs recognize viral RNA and recruit specific signaling adaptor molecules; (ii) the signal-transduction phase, where adaptor molecules transduce signaling to activate IKK-related kinases that activate transcription factors, like interferon regulatory factor 3 (IRF3) and nuclear factor-κB (NF-κB); and (iii) the effector phase, where IRF3 and NF-κB translocate to the nucleus and prime type I IFN synthesis. Type I IFNs then activate the signal transducers and activators of transcription (STAT) pathway on neighboring cells to induce synthesis of interferon-stimulated genes (ISGs). RNA viruses are detected by membrane-bound Toll-like receptors (TLRs) and cytoplasmic sensors, including retinoic acid induced gene-I (RIG-I) and melanoma differentiation associated gene (MDA-5). Although RIG-I and MDA-5 are both RNA helicase domain-containing proteins that use mitochondrial anti-viral signaling protein (MAVS, also called VISA, IPS-1, Cardif) to transduce signaling, they specialize in sensing different types of viruses [3]–[6].
Enterovirus 71 (EV71), which belongs to the Picornaviridae family, is a single-stranded, positive-sense RNA virus. EV71 infection usually causes childhood exanthema, also known as hand, foot, and mouth disease (HFMD). Acute EV71 infection can also induce severe neurological disease, including aseptic meningitis, brainstem and/or cerebellar encephalitis, and acute flaccid paralysis [7]. EV71 outbreaks have been reported around the world since the first report in the United States in 1974 [8]. In recent years, the frequency and the severity of EV71 infection are increasing in China and pose a threat to human health and social stability. However, no effective vaccines or specific anti-viral treatments are currently available.
Although the specific molecular mechanism underlying EV71 pathogenesis is not clear, EV71 virulence is associated with circumventing anti-viral immunity. While type I IFN administration protects mice against EV71 infection, anti-IFNα/β neutralizing antibody treatment exacerbates EV71-induced disease [9]. Recent studies show that the EV71-encoded 3C protease (3Cpro) inhibits the RIG-I and MAVS interaction and is able to cleave TIR domain-containing adaptor inducing IFN-β (TRIF), a key TLR3 adaptor molecule, to inhibit type I IFN production [10], [11]. Another recent study showed that 2Apro, another EV71 protease, reduced IFN receptor I (IFNAR1) expression that inhibited type I IFN signaling [12]. Although these known EV71-mediated inhibitory mechanisms affect the initiation and effector phases of the innate immune response, not much is known about the effect of EV71 infection on the signal transduction phase involving TLR3- or RIG-I/MDA5-mediated type I IFN production, a phase that is usually regulated by a sophisticated interplay between host and viral proteins under infection conditions. This study aimed to explore whether and how EV71 inhibits type I IFN production through regulating signal transduction pathways. We found that EV71 inhibited type I IFN responses upstream of IRF3 activation. MAVS, the common adaptor signaling molecule acting upstream of IRF3, was cleaved during EV71 infection. MAVS cleavage was independent of host cellular protease activity, but was dependent on EV71-encoded protease 2Apro, where 2Apro cleaved MAVS at three residues with different degrees of cleavage. EV71 also induced morphological and functional changes to host-cell mitochondria, and the EV71 VP1 protein was found to associate with host-cell mitochondria. Overall, our findings reveal a novel virus–MAVS interaction that inhibits signal transduction induced by anti-viral innate immunity to evade the ensuing immune response.
Previous studies demonstrated that EV71 evolved mechanisms to counteract type I IFN production [10]–[12]. To confirm and further clarify whether and how EV71 inhibits type I IFN production and determine at which step inhibition occurs, type I IFN production was evaluated. First, we measured type I IFN activity in supernatant from Sendai virus (SEV)- or EV71-infected HeLa cells using the type I IFN-responsive 2FTGH-ISRE reporter cell line. While the supernatant from the positive control SEV-infected HeLa cells exhibited time-dependent type I IFN production, supernatant from EV71 infected cells contained negligible type I IFN production over 36 h (Figure 1A). RT-PCR analysis showed that EV71 failed to induce mRNA expression of IFN-β or RANTES, a proinflammatory cytokine, in HeLa cells even though SEV could successfully do so (Supplemental Figure S1A). To confirm these results, a luciferase reporter assay was performed to investigate whether SEV- and EV71-infection induced IFN-β and NF-κB promoter activation. EV71 barely activated the IFN-β and NF-κB promoters (Supplemental Figure S1B). The above results suggest that EV71 inhibitory activity may occur upstream of the effector phase of type I IFN production.
Based on the above results, we next looked at IRF3 dimerization, which is a critical step upstream of IFN-β transcription and production. IRF3 dimerization was monitored by native PAGE, and we found that EV71-infected HeLa cells did not induce IRF3 dimerization even though SEV was able to induce it in a time-dependent manner (Figure 1B). This result indicates that EV71 might inhibit IFN-β production upstream of IRF3 activation. In order to confirm this result, native PAGE was performed on EV71-infected HeLa cells super-infected with SEV at different time points post-EV71 infection. The results showed that EV71 infection led to a pronounced, time-dependent decrease in SEV-induced IRF3 dimerization but did not interfere with SEV replication (Figure 1C). This EV71-mediated suppression of SEV-induced IRF3 dimerization reinforced the idea that EV71 inhibited IFN-β upstream of IRF3 activation.
MAVS is the unique adaptor molecule shared between the RIG-I and MDA-5 cytoplasmic PRRs, which acts upstream of IRF3 [3]–[6]. Many viruses, such as hepatitis C virus (HCV) [6], [13]–[16], GB virus [17], hepatitis A virus (HAV) [18], Coxsackievirus B3 (CVB3) [19], and rhinovirus [20], specifically target MAVS in order to escape host innate immunity. Considering the important function of MAVS in both the RIG-I and MDA-5 signaling pathway, a time-course study was conducted to test MAVS expression levels during EV71 infection by western blot. We found that expression of full-length MAVS declined after EV71 infection, and two fragments appeared at approximately 30 kD in both EV71-infected HeLa cells and rhabdomyosarcoma (RD) cells (Figure 2A–B). This result suggested that MAVS was cleaved during EV71 infection and that more than one cleavage residue may exist. In order to confirm that MAVS was indeed the source of these cleavage bands, two separate antibodies raised against different amino acid sequences of MAVS (E-3 was raised against residues 1–135 of human MAVS, while AT107 was raised against residues 160–450) were used to probe the above-mentioned western blot. Indeed, the cleavage products were recognized by both antibodies, as exhibited by the yellow signal that appeared after merging the green (E3) and red (AT107) western blot images. This result confirmed that MAVS was the source of the cleaved products (Figure 2A–B).
MAVS is localized on the outer membrane of mitochondria, and this sub-cellular localization is crucial for its function in anti-viral signaling. We therefore examined whether any changes to the cellular distribution of its cleavage products occurred during EV71 infection by confocal microscopy. The results showed that MAVS co-localized with Mito-dsRed, an RFP-containing mitochondrial target construct, in mock-infected cells. However, EV71 infection dramatically disrupted this co-localization (Figure 3A). To further confirm this, we separated the mitochondrial protein from the cytosolic protein by differential centrifugation. Western blot analysis was performed to determine the distribution of MAVS and its cleaved fragments; we clearly observed that MAVS was cleaved from the mitochondria, and the cleaved fragments were released into the cytoplasm (Figure 3B).
Viral infection induces cellular apoptosis as a consequence of the battle between the host cells and the virus. Apoptosis has been observed to occur in EV71-infected cells [21]–[23], and the EV71-derived proteases 2Apro and 3Cpro have been reported to induce this process [24], [25]. During virus-induced apoptosis, caspases are activated and lead to cleavage of some cellular proteins like PARP. Innate immune signaling proteins such as RIG-I, MDA-5, and MAVS are also targeted by activated caspases in other viral infections [20], [26]–[28]. These proteins also undergo proteasomal degradation through host- and viral-protein-mediated ubiquitin-ligating proteins, like host-derived RNF125, RNF5, and PCBP2 and the virus-derived hepatitis B virus (HBV) X protein [29]–[32]. To test whether EV71-induced MAVS cleavage is associated with cellular apoptosis and activated caspases, we first examined whether caspase activation occurred after EV71 infection in HeLa cells by western blot analysis of pro-caspase 3, 8, 9, PARP, and EV71-VP1 during an infection time course. EV71 infection led to caspase 3, 8, and 9 activation as well as PARP cleavage. PARP cleavage began at 12 h post-infection and was nearly complete at 24 h (Figure 4A), while MAVS cleavage was similarly detected at both 12 and 24 h post-infection (Figure 2A), suggesting that MAVS cleavage accompanied cellular apoptosis. To further investigate whether MAVS cleavage is the result of activated caspases or proteasome degradation, we tested the effect of pan-caspase inhibitor Z-VAD-FMK and proteasome inhibitor MG132 on MAVS cleavage in mock- or EV71-infected HeLa cells. Western blot analysis showed that PARP cleavage and caspase-3 activation, but not MAVS cleavage, was inhibited by Z-VAD-FMK alone or Z-VAD-FMK in combination with MG132. MG132 alone inhibited EV71 replication (indicated by the decreased VP1 protein, which was also reported in other viral infections [33]–[35]), but could not rescue MAVS cleavage (Figure 4B). Consistent with these results, neither the inhibitors alone nor their combined treatment could rescue IRF3 dimerization in EV71-infected cells as determined by native PAGE (Figure 4C). Taken together, the above results indicate that MAVS cleavage is independent of cellular apoptosis and proteasome degradation.
Mitochondria are well known for their crucial role in energy production, calcium homeostasis, and apoptosis. The presence of MAVS on the mitochondrial outer membrane indicates that this organelle has anti-viral functions. Recently, reports show that mitochondrial dynamics and membrane potential (ΔΨ m) are all required for MAVS-mediated anti-viral signaling, which underscores the importance of the mitochondrial microenvironment in anti-viral signaling [36], [37]. As MAVS was cleaved during EV71 infection and accompanied cellular apoptosis, we evaluated whether other mitochondrial abnormalities were associated with EV71 infection. First, we measured membrane potential using Mito-probe JC-1, a cationic dye that indicates mitochondrial depolarization by red-green fluorescence ratio reduction. Upon EV71 infection, an obvious loss of ΔΨ m began at 12 h (Figure 5A). We next assessed mitochondrial outer-membrane permeability by measuring cytochrome c release, another indicator of mitochondrial abnormality, and found that EV71 infection led to a small amount of cytochrome c release from the mitochondria into the cytoplasm (Figure 5B). To further explore mitochondrial abnormalities, we observed morphological changes by confocal microscopy of Mito-dsRed-transfected HeLa cells infected with EV71. Dramatic morphological changes occurred, as the typical mitochondrial network structure observed in mock-infected cells became diffuse and unclear in EV71 infected cells. Moreover, mitochondria partially stained positive for an anti-EV71 virus antibody, indicating viral co-localization with mitochondria (Figure 5C). Further in-cell western blot analysis demonstrated that the EV71 antibody was against the EV71 structural protein VP2 (Supplemental Figure S2). The extent of this partial co-localization indicated that mitochondria might only function at particular steps during the viral life cycle.
The processed viral components of many viruses, like HBx of HBV [38], NS3/4A and NS4A of HCV [6], [13], [14], [16], [39], 2B of poliovirus [40], and the 3ABC precursor of HAV [18], have been reported to associate with mitochondria to induce morphologic and functional changes in the mitochondria, causing subsequent apoptosis or targeting MAVS to inhibit innate-immune signaling. Based on the above analysis, we tested whether mitochondria were involved in the EV71 viral replication cycle by evaluating whether the EV71 structural protein, VP1, physically associated with mitochondria. Western blot analysis of mitochondria isolated from the cytoplasmic protein fraction showed that VP1 is mainly detected in the crude mitochondria as compared to the cytosol compartment (Figure 6A). In order to exclude the possibility that the VP1 detected in the isolated mitochondria fraction was a result of endoplasmic reticulum (ER) contamination that is believed to be important for picornavirus replication, we performed a more rigorous protocol to isolate mitochondria (using slower centrifugation speeds) and further purified it by Percoll gradient fractionation (Figure 6B) [15], [41]. Using specific markers for ER and mitochondria, western blot analysis demonstrated that the pure mitochondria were not contaminated with ER and that EV71 VP1 still associated with the mitochondrial compartment (Figure 6C). Collectively, the above results strongly indicate that the EV71 viral replication cycle involves the mitochondria, suggesting that viral proteins expressed during EV71 propagation may cause mitochondrial abnormalities and induce MAVS cleavage.
EV71 encodes two proteases, 2Apro and 3Cpro, that are important for processing viral protein precursors; they also reportedly cleave a variety of host-cell molecules that affect fundamental functions of the host cell. Since we found that MAVS was cleaved upon EV71 infection, we speculated that EV71 proteins executed this cleavage, especially as we previously excluded the role of cellular proteases and further detected the presence of viral protein on mitochondria. Since 2Apro has a strong inhibitory effect on host gene expression that makes it difficult to express and test in cultured cells, we first took advantage of a cell-free in vitro cleavage system—considered to be the most straight-forward approach to study picornavirus protease hydrolysis function [42]–[48]—to determine whether EV71-encoded 2Apro and 3Cpro proteases could directly cleave MAVS. We incubated recombinant EV71 2Apro and 3Cpro with HeLa cell extracts and detected MAVS cleavage by western blot using two antibodies that recognize different MAVS epitopes. EV71-infected HeLa cells were used as the positive control. We found that although both proteases generated cleavage bands, only 2Apro generated the same-sized cleavage bands as the EV71-infected cells. The appearance of these cleavage bands, approximately 30 kD in size, correlated with 2Apro treatment in a dose-dependent manner (Figure 7A–B). Another band in both 2Apro and 3Cpro treated cell extracts (Figure 7A–B, indicated by *) was considered to be a non-specific cleavage product and will be discussed later. In order to further scrutinize the role of EV71 3Cpro, we transfected HeLa cells with increasing doses of a plasmid encoding GFP-tagged 3Cpro and 3ABC proteases, as HAV use the 3ABC precursor to cleave MAVS [18]. Neither of these proteins induced MAVS cleavage even when expressed at a high level in HeLa cells (Figure 7C). This result was also consistent with our previous study showing that EV71 3Cpro could not interact with MAVS when over-expressed in live cells [11].
We next explored whether 2Apro exhibited any proteolysis ability on MAVS by transfecting a 2Apro-expressing plasmid into HeLa cells. eIF4GI, a known substrate of 2Apro, was used as an readout to indicate whether 2Apro was functional in this experimental system, as we know that 2Apro expression in this system may be weak since 2Apro protein was difficult to detect by western blot (likely due to the concomitant restriction on its own expression from its inhibition effect on host gene expression). To our surprise, while eIF4GI cleavage was detected in this system, PABP, another 2Apro substrate [20], [26], and MAVS remained intact (Supplemental Figure S3A). We speculated that this difference might be due to the varied sensitivities that these substrates have to 2Apro levels, and we tested this idea by a time-course study in EV71-infected cells. Since all mature EV71 viral proteins are derived from the same poly-protein precursor that undergoes subsequent post-translational cleavage, the amount of VP1 could indirectly reflect the varied expression of 2Apro and was therefore utilized to monitor 2Apro expression in this study. The results showed that eIF4GI cleavage appeared at 6 h after EV71 infection when VP1 protein was expressed at a low, not detectable level; in contrast, PABP and MAVS cleavage was observed at a later time point, at 12 h, when VP1 was abundantly expressed during infection (Supplemental Figure S3B). This result supported our above speculation. Previous attempts to efficiently express target genes in mammalian cells used the prokaryotic T7 RNA polymerase and the internal ribosome entry site sequence (IRES) of encephalomyocarditis virus (EMCV) to avoid host transcription factors and permit mRNA translation in a capping-independent way [49]. Another study showed that the foot-and-mouth disease virus (FMDV), which also belongs to the Picornaviridae family, could be efficiently rescued in a baby hamster kidney cell line (BHK-21) stably expressing T7 polymerase [50]. Considering that FMDV has a similar genomic structure and encodes a similar protease to EV71 2Apro [45], [51] and that our 2Apro-expressing plasmid contained both a T7 promoter and IRES sequence upstream of the 2Apro coding region, we exogenously expressed 2Apro and assessed its cleavage effect on MAVS in BSRT7/5 cells, a derivative cell line from BHK-21 that constitutively expresses T7 RNA polymerase [52]. 2Apro was indeed abundantly expressed in these cells, and the results showed that MAVS decreased with increasing 2Apro expression (Figure 7D). However, the cleavage bands were absent in this system; this absence might be due to the amino acid sequence differences between human and hamster MAVS, or to highly efficient cleavage in this over-expression system, rendering the cleavage fragments unstable or short-lived. An analogous phenomenon was previously reported in CVB 3Cpro-mediated cleavage of MAVS and in HCV-mediated cleavage of TRIF [19]. Taken together, these results suggest that EV71 2Apro, but not 3Cpro, is the protease inducing MAVS cleavage upon EV71 infection.
As EV71 2Apro is a cysteine protease, its major catalytic sites are His21, Asp39, and Cys110. To further confirm that the catalytic enzymatic activity of 2Apro is responsible for cleaving MAVS, we introduced a mutation in 2Apro that changed amino acid 110 from Cys to Ala (named 2Apro-110), which destroyed and inactivated the catalytic site of 2Apro [12], [47], [53]. We incubated 2Apro-110 with HeLa cell extracts and used PABP as a positive control for picornavirus 2Apro enzyme activity. Western blot results showed that the mutated 2Apro lost the ability to induce cleavage of both MAVS and PABP in this cell-free cleavage system (Figure 7E). Taken together, these results suggest that EV71 2Apro mediates MAVS cleavage during EV71 infection, and that the catalytic enzyme activity of 2Apro is required for cleaving the MAVS protein.
In order to identify the 2Apro-targeted cleavage residue(s) within the MAVS protein, we took advantage of the Protease-Glo Assay system to screen the whole extra-membrane region of MAVS. In this system, synthesized oligonucleotides encoding 12-mer polypeptides of MAVS (every 12 amino acids, with 6 amino-acid overlap) were inserted in-frame into a pGlosensor-10F linear vector that contained a genetically engineered firefly luciferase. The constructs were then expressed in a protein expression system labeled with FluoroTect GreenLys and used as substrate for 2Apro. If 2Apro cleaved any of the expressed polypeptides, an increase in luciferase activity would be detected, and two cleaved products at 36 and 25 kD would emerge in gel analysis (Figure 8A). Upon the first round of screening the 86 constructs we generated, we chose any plasmid exhibiting more than a 5-fold increase in luminescence density together with the visualized cleavage products in gel analysis as positive candidates; 10 constructs met this criterion (Figure 8B, Table 1). Some of these positive constructs may be false positives, since the linker region of pGlosensor-10F vector contains a Gly residue that is prone to being recognized as P1′ site of 2Apro substrate and cleaved by 2Apro [47], [54]–[56]. The Gly residues were mutated to Ala, and these vectors were used in the second round of screening. Three constructs remained positive and were found to encode MAVS protein residues 201–212, 243–254, and 255–266 (Figure 8C; Table 2).
Some characteristics are common among picornavirus 2Apro substrates, according to previous studies: the P1 position is preferentially occupied with a hydrophobic residue, and the P2 position is usually a Thr/Ser residue [47], [54]–[56]. The amino acid composition of the three positive polypeptides revealed that the P1′ residues were composed of Gly209, Gly251, and Gly265, respectively. We therefore constructed site-directed mutants (from Gly to Ala) of these potential cleavage residues and designated them as M209, M251, and M265. Upon exposure to 2Apro, the results showed that these mutations conferred resistance to all three constructs (Figure 8D). To further confirm the above results, we constructed plasmids encoding the non-mutated MAVS extra-membrane segment (designated as MAVS-EM) as well as a corresponding MAVS mutant containing Gly to Ala mutations at all three (209, 251, and 265) sites (designated as MAVS-EM-3M). These two plasmids were expressed by in vitro translation in the presence of FluoroTect GreenLys. Gel analysis of the 2Apro-induced cleavage pattern demonstrated that 2Apro hydrolyzed MAVS-EM but failed to hydrolyze MAVS-EM-3M (Figure 8E). Taken together, these results demonstrate that Gly209, Gly251, and Gly265 are the cleavage residues within MAVS that are targeted by EV71 2Apro.
We also tested EV71 3Cpro in both the Protease-Glo assay screening for cleavage sites in the extra-membrane region of MAVS and in the cleavage assay testing cleavage ability on the in vitro translated MAVS-EM. Although two oligo sets (encoding MAVS residues 87–98 and 147–158) appeared to be positive candidates in the Protease-Glo assay (Supplemental Figure S4), 3Cpro failed to cleave in vitro translated MAVS-EM (Supplemental Figure S5). This result was consistent with the results obtained from the 3Cpro and 3ABC over-expressed cells, and again demonstrated the inability of EV71 3Cpro to cleave MAVS. The discrepancy between the results may be explained by MAVS harboring potential 3Cpro cleavage sites that could be cleaved in the linear-polypeptide-based screening assay but not in the whole-protein-based cleavage assay due to conformational structure constraints that might block the approaching of 3Cpro protein.
Considering that MAVS translated in vitro may be slightly different from MAVS expressed in mammalian cells, such as in its protein conformation, stable cell lines were established to express wild-type MAVS (WT-MAVS) and MAVS mutants. Among the MAVS mutants, each residue was mutated individually and designated as m-MAVS-209, m-MAVS-251, and m-MAVS-265, and all three residues were also simultaneously mutated within one mutant (m-MAVS-3M). Cell lysates from the above cell lines were incubated with 2Apro and then subject to western blot analysis to evaluate MAVS cleavage. Wild-type MAVS could be cleaved by 2Apro, resulting in two major cleavage fragments: CF209, a ∼40 kD peptide cleaved from Gly209, and CF251/265, a ∼34 kD product. CF251/265 might contain a mixture of cleavage fragments from CF251 and CF265 cleaved from Gly251 and Gly265, respectively. Gly251 and Gly265 lie close to each other within the protein, which might account for why these two cleavage fragments could not be distinguished from each other in the gel. While m-MAVS-3M is resistant to cleavage by 2Apro, m-MAVS-209, m-MAVS-251, and m-MAVS-265 exhibited different degrees of cleavage after incubation with 2Apro (Figure 9A). 2Apro showed the strongest cleavage ability against Gly251, followed by Gly209 and Gly265. This comes from the evidence that CF209 was more abundant than CF265 in 2Apro-treated cell lysates of m-MAVS-251 but was relatively less than CF251 in 2Apro-treated cell lysates of m-MAVS-265 (Figure 9A, lanes 8&10). These results were also consistent with the luminescence density detection results in the previous Protease-Glo screening assay, which showed that the vector encoding residues 243–254 induced the highest fold increase of luminescence density (68-fold), compared to the vector encoding residues 243–254 (40-fold) and 255–266 (13-fold) (Table 2). Figure 9B schematically summarizes the cleavage fragments and the degrees of cleavage that we could conclude from the above analysis.
To evaluate the cleavage order of each residue by the 2Apro protease, we performed a kinetic analysis of 2Apro on WT-MAVS, m-MAVS-251, and m-MAVS-265. Although all cleavage fragments exhibited a time-dependent increase upon incubation with 2Apro, the time they emerged slightly differed among them. CF251 emerged at 5 min, while CF209 and CF265 began to appear at 15 min (Figure 9C–E). Moreover, this assay verified that CF251 had the strongest band intensity, followed by CF209 and CF265 (Figure 9C–E), consistent with the results from Figure 9A. Taken together, these results suggest that 2Apro exerts varying proteolysis ability on the different cleavage residues contained in MAVS and that Gly251 is the dominant residue that 2Apro most strongly and rapidly cleaves.
Since both MAVS and mitochondria are EV71 targets, we wondered whether normal mitochondria containing full-length MAVS could rescue the EV71-mediated inhibition of IRF3 activation. Zeng et al. had established a cell-free system demonstrating that mitochondria derived from SEV-infected cells could activate IRF3 in cytosol [57], [58]. Taking advantage of this system, we separated the mitochondrial and cytosolic compartments from mock-, SEV-, and EV71-infected cells, and reconstituted the RIG-I signaling pathway by exchanging the different compartments. While the mitochondria from SEV-infected cells dimerized IRF3 in the presence of mock-infected cytosol (Figure 10A, lane 3), mitochondria from EV71-infected cells inhibited this process (Figure 10A, lane 4). Moreover, mitochondria from SEV-infected cells rescued IRF3 activation in EV71-infected cytosol (Figure 10A, lane 6). These results suggest that MAVS cleavage and the associated mitochondrial changes might be a direct cause of EV71-induced inhibition of the innate immune response.
MAVS function requires mitochondrial localization. Since the EV71-induced MAVS cleavage occurred at three different residues between the proline-rich domain and the transmembrane domain, the N-terminal MAVS cleavage fragments would be released from the mitochondria. To test whether these cleavage fragments lost function in inducing type I IFN production, a series of deletion mutants from each cleavage residue was generated (Figure 10B) and transfected into HeLa cells with an IFN-β luciferase reporter plasmid. While full-length MAVS strongly activated the IFN-β promoter (nearly 1200-fold), none of the deletion mutants could activate the promoter, suggesting that the EV71-induced MAVS cleavage inactivated the signaling cascade leading to type I IFN production (Figure 10C).
EV71 is a member of the Enterovirus genus, Picornaviridae family. Its pathogenicity is likely related to its ability to evade host innate immunity. Although both the TLR3 and RIG-I/MDA-5 pathways recognize viral PAMPs and induce host anti-viral signaling during the innate immune response induced upon EV71 infection [1], [2], [59], the type I IFN response usually resulting from these pathways is totally absent [11]. The mechanism behind this observation is not clearly understood, although circumventing strategies have been found in RIG-I and TLR3 pathways [10], [11]. In this report, we reveal that another signaling molecule, MAVS, is cleaved by the EV71 viral protein 2Apro at multiple residues that results in inhibiting type I IFN production. This novel finding can help to explain the influence of EV71 on both RIG-I and MDA-5 signaling transduction pathways and is a good supplement to the current understanding of how EV71 escapes host innate immunity.
The central role of MAVS in innate immunity predisposes it to being a target of many viruses. In recent years, several different viruses were reported to use various strategies to disrupt MAVS function. HCV-derived NS3/4A protease was the first viral protein reported to co-localize with MAVS at mitochondrial membranes and cleave MAVS at Cys508 [6], [13], [14], [16], and HBV-derived HBx protein was reported to bind MAVS and promote its degradation to inhibit IFN-β production [31], [60]. More interestingly, viruses within the Picornaviridae family cleave MAVS through various mechanisms and at different sites. HAV, a picornavirus belonging to the Hepatovirus genus, cleaves MAVS at Gln428 by the protease precursor 3ABC [18]. Rhinovirus cleaves MAVS by its 2Apro and 3Cpro proteases as well as by activated caspase 3. Coxsackievirus B3 (CVB3), another member of Enterovirus genus in the Picornaviridae family, cleaves MAVS at Gln148 by its 3Cpro [19]. Our finding that EV71 2Apro cleaved MAVS at Gly209, Gly251, and Gly265 provides a new insight into how virus-derived proteins and MAVS can interact. To our knowledge, our study is also the first to show that MAVS cleavage occurred at multiple residues to inhibit type I IFN production. All three cleavage residues reside within the region between the proline-rich domain and transmembrane domain of MAVS, and this region is relatively disorganized from a structural point of view and forms a reasonable docking structure for the approaching of 2Apro protease. Mukherjee et al. previously studied MAVS expression in CVB3- and EV71-infected cells. While they found that CVB3 cleaved MAVS into fragments between 40–50 kD, they failed to detect these cleavage products in EV71-infected cells even though MAVS expression was significantly reduced in both cases; they speculated that MAVS was cleaved at other sites during EV71 infection [19]. Our studies confirmed their speculation, as EV71-induced MAVS cleavage not only occurred at other residues but also by a new mechanism. This finding provides new information regarding pathogen diversity as well as host-pathogen antagonism.
Due in part to the identification that mitochondrial-localized MAVS participates in the innate immune response, the idea that mitochondria not only play an important role in energy metabolism and cellular apoptosis but also provide a platform for virus-host interaction is now a generally accepted concept [3]–[6]. Consistent with this, some viral proteins also localize to the mitochondria to cleave MAVS as a way to circumvent innate immunity, like NS3/4A of HCV or the 3ABC precursor of HAV [6], [13], [14], [16], [18]. Also, mitochondrial dynamics and membrane potential have recently been recognized as essential for MAVS-mediated anti-viral signaling [36], [37]. These examples highlight the function of mitochondria as a platform structure in innate immunity, where viruses rely on its membrane structure and constitution to complete replication, and host cells utilize its membrane communication mechanisms to sense viral PAMPs and induce anti-viral immunity. In our study, we detected EV71 VP1 protein on mitochondria, raising the possibility that mitochondria may function at some particular stage of EV71 propagation; our results also further support the idea that EV71 could use this localization to cleave MAVS and destroy mitochondria to evade host innate immunity and provide another example for host-pathogen antagonism occurring on this intracellular-membrane platform. This finding could also explain the previously reported interaction between EV71 3Cpro and RIG-I [11]. RIG-I is recruited to a region nearby the mitochondria upon activation and interacts with MAVS via its CARD domain; the known role of 3Cpro in this process suggests that its presence is proximal to the mitochondria. In the literature, mitochondria have only been identified as a replication site for alphanodavirus flock house virus (FHV) [61], although the mitochondrial localization of HAV-derived 3ABC suggested an association with mitochondria in picornavirus replication [18]. Our current findings that EV71 VP1 co-localizes with mitochondria and that mitochondrial abnormalities were observed in EV71-infected cells strengthen the concept that mitochondria play a role in picornavirus replication. Future studies focusing on the specific mechanisms of mitochondria during the picornavirus life cycle should be carried out to further explore this concept.
EV71-encoded 2Apro and 3Cpro proteases are responsible for processing poly-protein precursors to produce mature structural and non-structural viral proteins. Picornavirus proteases affect numerous host mechanisms. EV71 3Cpro had been identified as a strong antagonist of innate immunity, as it was shown to interact with RIG-I and cleave TRIF to inhibit the RIG-I– and TLR3-mediated anti-viral signaling [10], [11]. Picornavirus 2Apro, on the other hand, has been shown to hijack host-cell gene expression by cleaving eIF4GI, eIF4GII, and PABP, among other things [10], [11]. This gene “shutoff” mechanism also inhibits expression of IFN-stimulated genes and can therefore be considered another mechanism by which picornavirus regulates host innate immunity. Moreover, Enterovirus 2Apro was also previously shown to be essential for its own replication in type I interferon-treated cells [62], and a recent study showed that EV71 2Apro reduces IFN receptor I (IFNAR1) to inhibit type I IFN signaling, indicating that EV71 2Apro functions as an antagonist to anti-viral innate immunity. Our finding that EV71 2Apro strongly cleaves MAVS supports role for this protease in antagonizing innate immunity. Since our study as well as others showed that both EV71-encoded proteases target anti-viral innate immunity at multiple steps, it is possible they may act synergistically to ensure the effective immune-evasion of EV71. Our study also attempted to evaluate the contribution of the different mechanisms used by EV71 2Apro and 3Cpro to antagonize innate immunity. We generated two mutated EV71 infectious clones, M-EV71-2A110 and M-EV71-3C40, that contained mutations at residue 110 of 2Apro and at residue 40 of 3Cpro, respectively, as these sites had previously been demonstrated to be indispensable for innate-immune inhibition by 2Apro and 3Cpro in the above-mentioned study and in our previous study [11]. Unfortunately, we were not able to obtain EV71 mutants with these mutated proteases, as the mutations impeded EV71 production due to the critical nature of these residues in catalytic enzyme activity and in EV71 replication (Supplemental Figure S6).
In this study, we provided direct biochemical evidence that EV71 2Apro protease cleaved MAVS using a cell-free in vitro system. This in vitro cleavage system is widely used and considered to be the most straight-forward approach to study the hydrolysis function of picornavirus proteases [42]–[48]. However, this system presented the following drawbacks as compared to the in vivo system: factors affecting the cleavage process in live cells might be omitted in the in vitro system, such as subcellular location; the in vitro cleavage-reaction buffer is different from the microenvironment in live cells and might cause slight conformational changes of the target proteins; and variation in the amount of recombinant protease, cleavage time, and temperature might induce non-specific cleavage that might confound the results. We speculate that these factors might help to explain the appearance of another MAVS cleavage band (Figure 7A–B, indicated by *) in our in vitro cleavage system that did not appear in EV71-infected cells, which we now think may represent a non-specific product.
When mapping protease cleavage site(s) on a target molecule, the routine method is to construct a series of mutants based on cleavage band size and bioinformatic analysis according to the hydrolyzing characteristics of the protease, followed by co-transfection of the mutants and protease into cells to test the predicted outcome. This approach requires accurate prediction, and missing potential cleavage sites is a possibility, especially when multiple cleavage sites exist. This routine strategy was not appropriate to use in our study for the following additional reasons. First, we cannot successfully express EV71 2Apro at the required levels for verifying the speculated cleavage sites in regular cells, since 2Apro was reported to hijack host-cell gene expression and also affect its own exogenous expression in mammalian cells. Second, we failed to observe the cleavage bands in cells over-expressing MAVS upon EV71 infection, which we originally thought was due to the poor viral replication inhibited by innate-immune activation. Therefore, two strategies were adopted to circumvent these issues, including: (i) establishing HeLa cell lines that stably express MAVS and MAVS mutants in which no sustained IRF3 activation was observed; and (ii) using P2.1 cells to transiently over-express MAVS for EV71 infection experiments. The P2.1 cell line is derived from the HT1080 cell line; it cannot respond to type I and type II IFNs because it lacks functional Jak1 and expresses very low IRF3 levels [63]. Despite these strategies, we still failed to observe cleavage bands from exogenously transfected MAVS (data not shown). Although the underlying reason is not yet clear, we speculate the following possibilities to explain these results: (i) the conformation and distribution of exogenously transfected MAVS might be different from endogenous MAVS; and (ii) exogenously transfected MAVS might have the potential to activate innate immunity and therefore induce and recruit MAVS-associated negative regulators that might prevent its interaction with downstream molecules. This latter possibility was hinted at by a report showing that PCBP2 is a negative regulator of MAVS-mediated signaling [29], and association of MAVS with other proteins might also prevent any effect of EV71. We therefore switched strategies and took advantage of the Protease-Glo assay system to screen the whole MAVS extra-membrane region. Using these methods, we successfully identified three MAVS residues cleaved by the EV71 2Apro and confirmed this in both the in vitro translated MAVS-EM and the stably expressed MAVS in HeLa cells.
When using exogenous MAVS and MAVS mutants expressed in HeLa cells to evaluate MAVS cleavage, the cleavage fragments recognized by the HA antibody are located in the C-terminus of MAVS and its mutants; they are indeed the corresponding counterparts to the endogenous N-terminal cleavage fragments recognized by the anti-MAVS antibodies used in EV71-infected cells (Figure 2A–B). This can be deduced from the molecular weight size and band intensity of the cleavage fragments. Full-length endogenous MAVS is approximately 65 kD in size, and the two cleavage fragments resulting from EV71 infection are both approximately 30 kD, where one appears above the 30 kD molecular weight band (∼31 kD) and the other one appears below the 30 kD band (∼25 kD). These bands seem to be counterparts to and coincident with the observed 34 kD (CF251/265) and 40 kD (CF209) bands in Figure 9C–E, including their respective band intensities.
Overall, we showed in this study that the EV71-derived 2Apro cleaves the key adaptor molecule MAVS as a strategy to evade anti-viral innate immunity at the signal transduction phase. Furthermore, we identified three key residues cleaved by the 2Apro protease activity on the extracellular fragment of MAVS. Our findings therefore reveal a new mechanism of EV71 viral protease-mediated evasion of host innate immunity.
Rhabdomyosarcoma (RD) cells and HeLa cells were purchased from ATCC. RD cells were cultured in MEM supplemented with 10% FBS and penicillin/streptomycin. HeLa cells were cultured in DMEM supplemented with 10% FBS and penicillin/streptomycin. 2FTGH-ISRE cells were a gift from Dr. Zhengfan Jiang (School of Life Sciences, Peking University, China). BSRT7/5 cells were cultured in DMEM supplemented with 10% FBS and 1 mg/mL G418. Enterovirus 71 (EV71) is a Fuyang strain isolated from a child in the city of Fuyang with a clinical diagnosis of HFMD in 2008 (GenBank accession no. FJ439769.1), and was propagated in RD cells. Sendai virus (SEV) was kindly provided by Dr. Zhengfan Jiang and propagated in chicken embryos.
The PGL3-IFNβ-Luc, pNifty-Luc, and pRL-Actin plasmids were gifts from Dr. Zhengfan Jiang. Mito-dsRed was provided by Dr. Xuejun Jiang (Institute of Microbiology, Chinese Academy of Sciences, China). pEGFPC1-EV71-3ABC was constructed by inserting EV71 3ABC cDNA fragment into the Hind III and Sal I sites of the pEGFPC1 vector. The plasmid expressing EV71 2Apro was generated by PCR amplification from PEGFPC1-EV71-2A as described before [11] and cloned into pET 30a (+) vector. Plasmid expressing EV71 2Apro-110 was mutated by PCR using pET 30a (+)-2A as template. The MAVS construct and its mutants were generated by PCR amplification from GFP-MAVS (provided by Dr. Zhengfan Jiang) and cloned into the pcDNA3.1 (+) vector. pcDNA3.1-IRES-2A was a gift from Dr. Shih-Yen Lo (Department of Laboratory Medicine and Biotechnology, Tzu Chi University, Hualien, Taiwan) and described before [53].
Mouse monoclonal antibodies directed against β-Actin (AC-15) and GFP (GSN24) were purchased from Sigma. Rabbit polyclonal antibody against HA was purchased from Bethyl Laboratories. Rabbit polyclonal antibodies against IRF-3 (FL-425) and cytochrome c (7H8) were purchased from Santa Cruz Biotechnology. Mouse anti-MAVS (E-3, monoclonal antibody raised against residues 1–135 of human MAVS) and rabbit anti-MAVS (AT107, polyclonal antibody raised against residues 160–450 of human MAVS) were obtained from Santa Cruz Biotechnology and Enzo Life Sciences, respectively. Another MAVS antibody, which reacts with human, mouse, and rabbit MAVS, was purchased from Signalway Antibody and used in western blot analysis of BSRT7/5 cells. Mouse anti-KDEL (10C3, recognizes GPR78 and GPR94 with particular prominence), mouse anti-mitochondria (MTC02, recognizes a 60 kD non-glycosylated protein component of human mitochondria), rabbit anti-caspase 3, and mouse anti-PABP (10E10) were obtained from Abcam. Rabbit anti-PARP, rabbit anti-caspase 8 (D35G2), and rabbit anti-caspase 9 were obtained from Cell Signaling Technologies. Mouse anti-enterovirus 71 was purchased from Millipore. Mouse anti-enterovirus 71 VP1 (3D7) was purchased from Abnova. Rabbit anti-Sendai antibody was purchased from MBL International Corporation. The general caspase inhibitor benzyloxycarbonyl-Val-Ala-Asp-(OMe) fluoromethylketone (Z-VAD-FMK) and proteasome inhibitor MG132 were purchased from Sigma and Calbiochem, respectively.
HeLa cells (∼2×105) were seeded on 24-well dishes and transfected the following day by Lipofectamine 2000 (Invitrogen) with 200 ng of PGL3-IFNβ-Luc or pNifty-Luc and 5 ng pRL-Actin. Cells were co-transfected with 600 ng of the indicated plasmids or infected with EV71/SEV 24 h post-transfection. In all experiments, cells were lysed and reporter activity was analyzed using the Dual-Luciferase Reporter Assay System (Promega). The type I IFN bioassay was performed as previously reported by Sun et al. [64]. Briefly, the supernatant from SEV- and EV71-infected cells were collected at the indicated times, added directly to 96-well dishes seeded with 2FTGH-ISRE cells, and luciferase activity was measured after 6 h and calculated with reference to a recombinant human IFN-β standard (R&D system).
Native PAGE was carried out as previously described [65]. Native gel (8%) was pre-run with native running buffer (25 mM Tris and 192 mM glycine, pH 8.4) with 0.5% deoxycholate in the cathode chamber for 30 min at 25 mA on ice. Samples were prepared in the native sample buffer (62.5 mM Tris–HCl, pH 6.8, 15% glycerol, and 1% deoxycholate), then loaded onto the gel and electrophoresed at 20 mA for an additional 1 h.
Whole-cell extracts (20–100 µg) were separated by 8%–15% SDS-PAGE. After electrophoresis, proteins were transferred to a PVDF membrane (Bio-Rad). The membranes were blocked for 1 h at room temperature in 5% dried milk and then were probed with the indicated primary antibodies at an appropriate dilution overnight at 4°C. The following day, the membranes were incubated with corresponding IRD Flour 680- or 800-labeled IgG secondary antibodies (LI-COR Biosciences) and were scanned by the Odyssey Infrared Imaging System (LI-COR Biosciences).
Cells were fixed in 4% formaldehyde, permeabilized in 0.5% Triton X-100, blocked in 1% BSA in PBS, and then probed with indicated primary antibodies for 1 h at room temperature. Following a wash, cells were incubated with their respective secondary antibodies for another 1 h. The cells were then washed and stained with 4, 6-diamidino-2-phenylindole (DAPI) to detect nuclei. Images were captured with a laser confocal microscope (Leica).
Mitochondrial isolation was carried out by differential centrifugation. Briefly, cells were harvested and resuspended in HB buffer (210 mM mannitol, 70 mM sucrose, 5 mM HEPES, pH 7.12, 1 mM EGTA, and an EDTA-free protease inhibitor cocktail) and subject to homogenization. After 30 strokes, cell homogenate was centrifuged at 600×g for 10 min at 4°C. The supernatant was saved and subjected to further centrifugation at 10000×g for 10 min at 4°C. The pellet was washed once with HB buffer and designated as the crude mitochondrial fraction. The supernatant was further centrifuged at 12000×g and designated as the cytosol fraction after discarding the final pellet. Mitochondria purification was performed by Percoll gradient fractionation as previously described with minor modifications [41], [66], [67]. A schematic overview of the isolation and purification protocol is displayed in Figure 6B.
Recombinant EV71 3Cpro was produced as described before [68]. To produce EV71 2Apro and 2Apro-110, the respective plasmids were introduced into competent E. coli BL21 (DE3) cells, and protein expression was induced by treatment with 200 µM IPTG at 18°C overnight. 2A-His fusion protein was purified by Ni-Agarose column. In vitro cleavage assay was performed with the indicated amount of recombinant protease incubated together with cell lysates in reaction buffer (50 mM Tris-HCl, pH 7.0, and 200 mM NaCl) at 37°C for 6 h or 30°C for 2 h.
Mitochondrial membrane potential was analyzed using Flow Cytometry Mitochondrial Membrane Potential Detection Kit (BD Biosciences) by a BD FACS Canto II flow cytometer (BD Biosciences). The experiments were carried out according to the manufacturer's instructions.
Synthesized oligonucleotides encoding 12-mer peptides (with six amino-acid overlap between two adjacent 12-mers) for the MAVS extra-membrane region were inserted in pGloSensor-10F linear vector (Promega). The resulting vectors were subjected to in vitro transcription/translation with TNT SP6 High-Yield Wheat Germ Protein Expression System (Promega) and FluoroTect GreenLys in vitro Translation Labeling System (Promega) according to manufacturer's instructions. The reactions were incubated at 25°C for 2 h. Then, 7 µg of recombinant EV71 2Apro or 3Cpro was added to 10 µL reactions with 10 µL 2× digestion buffer (100 mM Tris-HCl, pH 7.0, and 400 mM NaCl). The digestion reactions were incubated for 2 h at 30°C, and a 10 µL aliquot was removed and subjected to 10% SDS-PAGE. The gels were scanned by a Typhoon gel scanner (GE Healthcare) to visualize the fluorescently labeled proteins. The remaining 10 µL was diluted 20-fold, and luciferase activity was measured using the Bright-Glo assay reagent (Promega) according to the manufacturer's instructions. In vitro transcription/translation of the MAVS extra-membrane region was performed by the TNT SP6 High-Yield Wheat Germ Protein Expression System Labeled with FluoroTect GreenLys. The DNA template for this assay was constructed by amplifying the MAVS coding region at residues 1–513 and cloned into the pF3AWG (BYDV) Flexi Vector (Promega).
HeLa cells were transfected with pcDNA3.1-MAVS and its mutants by Lipofectamine 2000 (Invitrogen) and selected in Zeocin (200 µg/mL) to establish the cell lines stably expressing MAVS and MAVS mutants.
In Vitro IRF3 activation assay was carried out as previously described by Zeng et al [57], [58]. Briefly, HeLa cells were resuspended in Buffer A (10 mM Tris-HCl pH 7.5, 10 mM KCl, 0.5 mM EGTA, 1.5 mM MgCl2, 0.25 M D-mannitol, and EDTA-protease inhibitor cocktail) and homogenated. Then, the homogenates were centrifuged at 1000×g at 4°C for 5 min; the supernatants were further centrifuged at 5000×g at 4°C for 10 min to separate the pellets (P5) and the supernatants (S5). P5 was washed once with Buffer B (20 mM HEPES-KOH pH 7.4, 0.5 mM EGTA, 0.25 M D-mannitol, and EDTA-protease inhibitor cocktail) and resuspended in Buffer B. For each reaction, 10 µg P5 and 20 µg S5 were mixed in Buffer C (20 mM HEPES-KOH pH 7.0, 2 mM ATP, 5 mM MgCl2) and incubated at 30°C for 1 h in a 10 µL reaction system. The reaction mixtures were then subjected to native PAGE, and the dimerization of endogenous IRF3 was detected by western blot.
MAVS (HGNC: 29233); RIG-I (HGNC: 19102); MDA-5 (HGNC: 18873); TLR3 (HGNC: 11849); IRF3 (HGNC: 6118); TRIF (HGNC: 18348); Caspase 3 (HGNC: 1504); Caspase 8 (HGNC: 1509); Caspase 9 (HGNC: 1511); PARP (HGNC: 270); PABP (HGNC: 8554); RANTES (HGNC: 10632); IFN-β (HGNC: 5434); PCBP2 (HGNC: 8648); RNF125 (HGNC: 21150); RNF5 (HGNC: 10068); IFNAR1 (HGNC: 5432); Cytochrome c (HGNC: 19986).
|
10.1371/journal.pgen.1004574 | Dopamine Signaling Leads to Loss of Polycomb Repression and Aberrant Gene Activation in Experimental Parkinsonism | Polycomb group (PcG) proteins bind to and repress genes in embryonic stem cells through lineage commitment to the terminal differentiated state. PcG repressed genes are commonly characterized by the presence of the epigenetic histone mark H3K27me3, catalyzed by the Polycomb repressive complex 2. Here, we present in vivo evidence for a previously unrecognized plasticity of PcG-repressed genes in terminally differentiated brain neurons of parkisonian mice. We show that acute administration of the dopamine precursor, L-DOPA, induces a remarkable increase in H3K27me3S28 phosphorylation. The induction of the H3K27me3S28p histone mark specifically occurs in medium spiny neurons expressing dopamine D1 receptors and is dependent on Msk1 kinase activity and DARPP-32-mediated inhibition of protein phosphatase-1. Chromatin immunoprecipitation (ChIP) experiments showed that increased H3K27me3S28p was accompanied by reduced PcG binding to regulatory regions of genes. An analysis of the genome wide distribution of L-DOPA-induced H3K27me3S28 phosphorylation by ChIP sequencing (ChIP-seq) in combination with expression analysis by RNA-sequencing (RNA-seq) showed that the induction of H3K27me3S28p correlated with increased expression of a subset of PcG repressed genes. We found that induction of H3K27me3S28p persisted during chronic L-DOPA administration to parkisonian mice and correlated with aberrant gene expression. We propose that dopaminergic transmission can activate PcG repressed genes in the adult brain and thereby contribute to long-term maladaptive responses including the motor complications, or dyskinesia, caused by prolonged administration of L-DOPA in Parkinson's disease.
| In Parkinson's disease (PD) the motor impairment produced by the progressive death of midbrain dopaminergic neurons is commonly treated with the dopamine precursor, L-DOPA. Utilizing a mouse model of PD, we show that L-DOPA, via activation of dopamine D1 receptors, promotes the expression of genes normally repressed by Polycomb group (PcG) proteins. We propose that this effect is exerted by promoting the phosphorylation of histone H3 on serine 28 at genomic regions marked by tri-methylation of the adjacent lysine 27, generating a H3K27me3S28p double-mark. This event leads to displacement of PcG proteins and aberrant gene expression. These findings reveal a previously unrecognized plasticity of PcG-repressed genes in terminally differentiated neurons. Furthermore, the identification of specific genes whose expression is increased upon prolonged treatment with L-DOPA and the consequential activation of dopamine D1 receptors offer a possibility to design novel therapeutic strategies to treat Parkinson's disease and potentially other disorders caused by dysfunctional dopaminergic transmission in the brain, such as drug addiction and schizophrenia.
| An emerging concept in neurobiology is that many mechanisms implicated in chromatin remodeling and developmental processes retain their plasticity in the adult brain. Indeed, a number of environmental stimuli are known to generate chromatin modifications that have been causally linked to synaptic plasticity and associated behavioral and pathological responses. In this context, core histone modifications [1] have been implicated in cognitive functions, as well as in psychiatric conditions [1], [2].
Polycomb group (PcG) proteins maintain cell type specific gene repression that is established during early embryonic development by regulating chromatin structure [3]. The Polycomb repressive complex 1 (PRC1) mediates histone H2A lysine 119 mono-ubiquitination (H2AK119ub), while PRC2 di- and tri-methylates histone H3 lysine 27 (H3K27me2/3) [4], [5]. Functionally, both PRC1 and PRC2 can be recruited to genomic regions through direct binding to H3K27me3 marked chromatin. Importantly, while dysregulation of PcG binding to target genes has been implicated in serious developmental defects and diseases such as cancer [6], [7], aberrant derepression of PcG target genes have not been associated with pathology of terminally differentiated neurons [2].
Parkinson's disease (PD) is caused by the death of midbrain neurons producing dopamine. This disorder is commonly treated with L-DOPA, which upon conversion to dopamine, relieves the motor symptoms of PD [8]. However, prolonged use of L-DOPA results in the emergence of dyskinesia, involving dystonic and choreic movements [9]. Several lines of evidence indicate that L-DOPA-induced dyskinesia (LID) is caused by abnormal activation of dopamine D1 receptors (D1Rs) located on the medium spiny neurons (MSNs) of the striatum [10], [11]. This leads to increased gene expression through sequential activation of PKA, dopamine- and cAMP-regulated phosphoprotein of 32 kDa (DARPP-32), extracellular signal-regulated kinases (Erk), mitogen- and stress-activated kinase 1 (Msk1) and eventually phosphorylation of histone H3 at serine 10 (H3S10p) [12]–[15].
While the regulation of histone H3S10 phosphorylation has been studied in the adult brain [16], [17], almost nothing is known regarding H3S28 phosphorylation in neurons. However, in non-proliferating human fibroblasts it has been shown that H3K27me3S28 phosphorylation in response to MSK activation can lead to transcription of otherwise stably repressed genes [18]. The initial derepression is caused by displacement of gene repressor complexes containing PcG proteins, followed by transcriptional activation.
In this study, we describe an important link between dopamine signaling, H3K27me3S28 phosphorylation, and aberrant gene expression associated to reduced PcG binding. Using a mouse model of PD, we show that dopamine via D1Rs increases H3K27me3S28 phosphorylation in striatal MSNs via two pathways: 1) activation of Msk1, leading to phosphorylation of H3K27me3S28 and 2) activation of DARPP-32 leading to protein phosphatase 1 (PP1) inhibition and suppression of H3K27me3S28p dephosphorylation. The combined effect is an accumulation of H3K27me3S28p at gene promoters that reduces PcG binding and allows transcription of a subset of genes. The results reveal a previously unrecognized plasticity of PcG-repressed genes in the adult brain, which upon environmental changes can be aberrantly induced via Erk-Msk1 mediated H3K27me3S28 phosphorylation and PKA-DARPP-32-dependent modulation of PP1 activity towards the same histone mark.
The ability of L-DOPA to activate the Erk-Msk1 pathway in striatal MSNs [15], [19], [20], led us to hypothesize that signaling through dopamine receptors would induce phosphorylation of S28 in the context of H3K27me3 marked genomic sites to generate the H3K27me3S28p double histone modification. To test this possibility we turned to an experimental mouse model of PD in which unilateral stereotaxic injection of the neurotoxin 6-OHDA results in the elimination of the dopaminergic innervation to the basal ganglia (Figure 1A) [20], [21]. In the lesioned striatum, MSNs react to the loss of dopamine by developing a remarkable sensitization to D1R agonists and, upon L-DOPA treatment strongly activate Msk1, while the MSNs in the contralateral, unlesioned striatum are unaffected [22]. This unilateral model of PD has the advantage that each mouse can serve as its own within-subject control as the dopamine sensitized MSNs in the striatum of the 6-OHDA lesioned side respond with intense D1R-mediated signaling to administration of L-DOPA, while the MSNs in the unlesioned side are not affected [15], [19], [20].
In the first experiment, lesioned mice were injected with L-DOPA and sacrificed 1 hour later (acute L-DOPA). By Western blotting, we observed that L-DOPA caused a dramatic increase in H3K27me3S28 phosphorylation in the lesioned striatum compared to the unlesioned striatum (Figure 1B).
The striatum contains two main populations of MSNs that are enriched for either D1Rs or dopamine D2 receptors (D2Rs). Activation of these neurons produces opposite behavioral responses, which are related to their distinct connectivity to the output stations of the basal ganglia [23]–[26]. To identify the population(s) of striatal MSNs that responded to L-DOPA with increased H3K27me3S28 phosphorylation we made use of transgenic mice expressing EGFP under the control of regulatory elements of the D1R or D2R (D1R-EGFP or D2R-EGFP) [27]. These mice were lesioned, treated with L-DOPA and perfused after 1 hour. The results showed that, in the lesioned striata of D1R-EGFP mice, H3K27me3S28p co-localized with the EGFP-labeled cell bodies (Figure 1C), while in D2R-EGFP mice, H3K27me3S28p was segregated from EGFP-labeled cell bodies (Figure 1C).
To further confirm that D1Rs activate the signaling cascade inducing the H3K27me3S28p mark, we injected naïve mice with the specific D1R-agonist SKF81297. Indeed, H3K27me3S28p was increased 1 hour after SKF81297 injection (Figure S1A).
We concluded that in hemiparkisonian mice, where the dopamine innervation was eliminated by unilateral injection of 6-OHDA, acute administration of L-DOPA produced a large increase in H3K27me3S28p specifically localized to the nuclei of striatal D1R-MSNs.
Given the dramatic increase in H3K27me3S28p in D1R expressing MSNs, we examined the global distribution of this mark on chromatin. We undertook chromatin immunoprecipitations (ChIPs) for H3K27me3S28p from pooled lesioned or unlesioned striata, after acute administration of L-DOPA (tissue from 45 mice were pooled for each condition), followed by ChIP-sequencing (ChIP-seq). This was also done using antibodies for H3K4me3 and H3K27me3, to define potentially active or PcG-repressed chromatin, respectively. In this way, we identified four genes encoding transcription factors (TFs) where H3K27me3S28 phosphorylation was induced near the transcription start sites (TSS) (Figure 2A). Two of them, Atf3 and Npas4 have previously been found to be implicated in neuronal plasticity, while Klf4 and Hoxa2 are well known PcG target genes implicated in stem cell function and cellular differentiation. In the lesioned striata the peaks observed for H3K27me3S28p at the Atf3, Klf4 and Npas4 genes coincided with the H3K27me3 peaks. Co-occurrence was also observed for H3K27me3S28p and H3K4me3 (Figure 2A). In contrast, chromatin at the Hoxa2 locus was blanketed with H3K27me3 and H3K27me3S28p in the lesioned striata, without any apparent H3K4me3 signal (Figure 2A).
To support these observations, we performed ChIP-qPCR on chromatin from lesioned and unlesioned striata after acute L-DOPA. Antibodies against H3K27me3S28p, H3K27me3, H3K4me3 and Rnf2 (PRC1 subunit), and general IgG as a negative control were used (Figure 2B). The results confirmed the induction of H3K27me3S28p in the lesioned striatum upon L-DOPA stimulation and the presence of H3K27me3 at all genes analyzed, as well as H3K4me3 enrichment in chromatin from the unlesioned and lesioned striata on Atf3, Klf4 and Npas4 genomic loci. In line with our previous findings [18], the induction of H3K27me3S28 phosphorylation on these loci correlated with a reduction in Rnf2 binding.
Whereas L-DOPA induced H3K27me3S28 phosphorylation in the striatum was restricted to D1R-MSNs, the origins of signal in the H3K27me3 and H3K4me3 ChIPs were uncertain, due to the presence of several cell types in the tissue. We have estimated the contribution of D1R-MSNs to the bulk chromatin analyzed in our ChIPs to be approximately 43% (Figure S3). Therefore, D2R-MSNs could account for a significant part of the H3K4me3 and H3K27me3 signals detected by ChIP.
The increase in H3K27me3S28p and the reduced binding of Rnf2 suggested that these genes become de-repressed upon acute L-DOPA treatment. We therefore performed RT-qPCR using RNA isolated from lesioned mice treated with L-DOPA or saline as control (Figure 2C). Indeed, L-DOPA increased the expression of mRNA for Atf3, Klf4 and Npas4 in the lesioned compared to the unlesioned striatum and to the saline controls, whereas the expression of Hoxa2 was unchanged (Figure 2C). The lack of expression from the Hoxa2 locus was supported by the absence of the active H3K4me3 histone mark. Importantly, the increases of Atf3 and Npas4 were furthermore confirmed at the protein level (Figure 2D), suggesting the potential involvement of these TFs in the phenotypic effects produced by L-DOPA in parkinsonian mice.
In summary, our data suggest that administration of L-DOPA in a mouse model of PD promotes H3K27me3S28 phosphorylation on several PcG target genes marked by H3K27me3. This effect occurs in the striatal D1R-MSNs of the lesioned brain hemisphere and correlates with a reduction in Rnf2 binding and increased gene expression.
To estimate the genome-wide extent of H3K27me3S28 phosphorylation induced by L-DOPA, we scored regions +/−1 kb from the transcription start sites (TSS) of all annotated (mm9) mouse transcripts (n = 189,660) for the enrichment of H3K27me3 in chromatin from unlesioned striata (Figure 3A) and H3K27me3S28p in lesioned striata (Figure 3B). For this correlation, the H3K27me3 mark was determined in the unlesioned striata, instead of the lesioned striata, in order to ensure that the H3K27me3 signal was not altered by epitope masking due to H3K27me3S28 phosphorylation [18]. Using a cutoff where only 5% of the regions were expected to score positive by chance (see Materials and Methods), this analysis showed that approximately 20.7% (n = 39,197) of all loci that can give rise to mRNA transcripts were H3K27me3 positive (Figure 3A) and 6.9% (n = 13,148) were H3K27me3S28p positive (Figure 3B). As observed in the Venn diagram presented in Figure 3C the majority (83%) of H3K27me3S28 phosphorylation occurred in genomic regions that were already marked by H3K27me3 before L-DOPA administration, while approximately 17% of the genomic sites had levels of H3K27me3 below our defined cut-off for the analysis and could in principle have gained H3K27me3 upon L-DOPA administration.
Having observed that an impressive 33.5% of all H3K27me3 positive loci that potentially could give rise to mRNA transcripts became enriched for H3K27me3S28p upon L-DOPA stimulation, we next asked if and to which extent these transcripts were actually induced. We therefore performed global RNA-sequencing (RNA-seq) on mRNA isolated lesioned mice treated with L-DOPA. The analysis showed that 1 hour after L-DOPA, 5.4% (n = 10,298) of all transcripts had significantly changed expression in the lesioned striatum compared to the unlesioned striatum (Figure 3D). Importantly, 20.4% of these transcripts also scored positive for H3K27me3 at the chromatin level (Figure 3E; threshold mentioned in Figure 3A). We next examined to which extent the regulated transcripts that were marked by H3K27me3 at their genomic loci also scored positive for H3K27me3S28p (according to Figure 3B) and found that 36.6% matched this criterion (Figure 3F). It was furthermore apparent that most transcripts with high H3K27me3S28p levels near the TSS were induced rather than repressed. Further analysis of the different populations of regulated transcripts showed that, of the 768 that originated from H3K27me3- and H3K27me3S28p-positive loci, 52 were ≥1.5 fold down-regulated, whereas the majority 339 were ≥1.5 fold up-regulated (Figure 3G). This could be compared to the 1,329 regulated transcripts originating from H3K27me3 positive loci lacking H3K27me3S28p of which 230 were ≥1.5 fold down-regulated and 332 were up-regulated (Figure 3H), or to all 10,298 regulated transcripts (regardless of specific histone marks at their loci), where 844 were ≥1.5 fold down-regulated and 2,892 were ≥1.5 fold up-regulated (Figure 3I).
Altogether, regulated transcripts originating from H3K27me3 positive loci that gained the H3K27me3S28p mark were more frequently induced (87%, Figure 3G) than regulated transcripts originating from H3K27me3 positive loci not gaining the H3K27me3S28p mark (59%, Figure 3H) or regulated transcripts in general (77%, Figure 3I). These data highlight a clear correlation between induction of H3K27me3S28 phosphorylation at specific genomic loci and increased transcription in response to acute L-DOPA stimulation in the lesioned striata of parkisonian mice.
Gene ontology (GO) analysis of the group of transcripts up-regulated 1.5 fold or more originating from H3K27me3 positive loci that gained the H3K27me3S28p mark (roman numerical II in Figure 3G), suggested that up-regulation of these gene products could affect overall transcriptional activity and rate of biosynthesis in neuronal cells (Figure 4A, Figure S4 and Table S1). This was in contrast to the group of up-regulated transcripts (≥1.5 fold) originating from H3K27me3 positive loci that did not gain H3K27me3S28 phosphorylation, which enriched for GO-terms involved in immune response (Figure 4B, Figure S4, subpopulation IV).
We have previously shown that MSK1 and MSK2 are the kinases mediating H3K27me3S28 phosphorylation in human fibroblasts, as pharmacological inhibition or shRNA mediated knockdown of MSK1 and MSK2 prevented the induction of H3K27me3S28p [18]. To elucidate which molecular pathways downstream of D1R activation in MSNs are implicated in the regulation of H3K27me3S28 phosphorylation, we lesioned Msk1−/− mice with 6-OHDA and injected them with L-DOPA. In line with our previous findings in fibroblasts, induction of H3K27me3S28p in the lesioned striatum was reduced in Msk1−/− mice compared to wt mice (Figure S5).
The exaggerated D1R transmission induced by L-DOPA in the MSNs of the dopamine-depleted striatum is characterized by elevated cAMP production and PKA activity [11], [13]. PKA further relays the signal via phosphorylation of DARPP-32 at T34 [15], [28]. This converts DARPP-32 into an inhibitor of protein phosphatase 1 (PP1), thereby suppressing dephosphorylation of downstream PP1 targets [29]. We have previously shown that a T34A mutation on DARPP-32 decreases L-DOPA-induced phosphorylation of Erk and histone H3S10 [14]. Therefore, we examined whether DARPP-32-mediated inhibition of PP1 was also involved in the regulation of H3K27me3S28p. When lesioned mice harbouring a T34A mutation in DARPP-32 were injected with L-DOPA we observed a significant less pronounced H3K27me3S28 phosphorylation in comparison to wt mice (Figure 5A). This finding suggested that PP1 is involved in the dephosphorylation of H3K27me3S28p. To test this possibility, we conducted an in vitro phosphatase assay, in which H3 peptides, either unmodified or modified with S28p or K27me3S28p, were incubated with PP1. Changes in the phosphorylation of the peptides after the reactions were detected by dot-blotting using an H3S28p antibody (Figure 5B). This assay showed that PP1 could efficiently dephosphorylate H3K27me3S28p.
To confirm that PP1 is the phosphatase acting on H3K27me3S28p in the striatum, we examined the effect of okadaic acid, which inhibits PP1 and PP2A [30], on H3K27me3S28p in a slice preparation from striatum by Western blotting (Figure 5C). Incubation of striatal slices with 1 µM okadaic acid, a concentration that inhibits both PP1 and PP2A, was sufficient to change the equilibrium towards increased levels of H3K27me3S28p compared to vehicle. In contrast, a concentration of 100 nM okadaic acid, which inhibits PP2A but is insufficient for PP1 inhibition [30], did not affect H3K27me3S28p levels. This supported PP1 as a phosphatase removing S28p from the H3K27me3S28p double-marked chromatin in vivo.
Next, we analyzed the outcome of the global reduction in H3K27me3S28p on regulatory regions of specific genes. Lesioned DARPP-32 T34A and wt mice were treated with L-DOPA and striatal tissue from the lesioned and unlesioned striatum was analysed by ChIP-qPCR. The induction of H3K27me3S28p at the Atf3, Klf4, Npas4 and Hoxa2 genes was significantly reduced in chromatin from the lesioned striatum of DARPP-32 T34A mice compared to wt mice (Figure 5D). Notably, this reduction correlated with decreased expression of Atf3, Klf4 and Npas4 mRNA in the lesioned striatum of L-DOPA injected DARPP-32 T34A mice compared to wt mice (Figure 5E).
Overall these data suggested that D1R stimulation induces transcription associated to H3K27me3S28 phosphorylation via two parallel pathways: 1) activation of Erk-Msk kinases and 2) concomitant PKA-mediated DARPP-32-phosphorylation, leading to inhibition of PP1 and suppression of H3K27me3S28p dephosphorylation.
LID is a serious motor complication caused by prolonged administration of L-DOPA to patients affected by PD [9]. This condition has been linked to persistent hyper-activation of the cAMP/DARPP-32 signaling cascade, produced by L-DOPA acting on sensitized D1Rs [10], [11]. Lesioned mice display dyskinetic behaviour in response to 9 sequential daily L-DOPA injections (chronic L-DOPA) [14], [15]. The severity of LID after chronic L-DOPA administration has been shown to correlate to the level of H3S10 phosphorylation and the induction of specific genes, such as Fosb [12], [15], [19], [31].
To investigate the contribution of H3K27me3S28 phosphorylation to the changes in gene expression associated to LID, 6-OHDA lesioned mice were treated chronically with L-DOPA and the levels of H3S28p and H3K27me3S28p were measured after 1, 3 and 9 days of administration. As expected, L-DOPA increased H3S28p and H3K27me3S28p in the lesioned striatum. However, the induction of these histone marks was progressively reduced during the course of chronic L-DOPA administrations (Figure 6A).
To examine if specific changes in gene expression associated to LID occurred for PcG-repressed genes, we performed global RNA-seq on mRNA isolated from striata of lesioned mice that had been treated chronically with L-DOPA (9 days). Transcripts that were significantly changed in the lesioned striatum after chronic L-DOPA were scored for enrichment of H3K27me3 in the unlesioned striatum and then for H3K27me3S28p in the lesioned striatum after acute L-DOPA, according to Figure 3A–F. From this analysis we concluded that despite the reduced induction of H3K27me3S28p following chronic administration of L-DOPA, regulated transcripts originating from genomic loci marked by H3K27me3 and H3K27me3S28p were still largely induced (76%, Figure S6A, subpopulation II) in comparison to regulated transcripts originating from genomic loci marked by H3K27me3 only (45%, Figure S6B, subpopulation IV) or in comparison to all regulated transcripts (46%, Figure S6C, subpopulation VI).
As the induction of the H3K27me3S28p mark was reduced after chronic L-DOPA compared to acute L-DOPA (Figure 6A), we next examined whether the transcripts from genes marked by H3K27me3 and H3K27me3S28p were less induced after chronic L-DOPA compared to acute L-DOPA treatment. Amongst the regulated transcripts originating from H3K27me3 and H3K27me3S28p positive genomic loci, 53% were ≥1.5 fold less expressed in chronic L-DOPA compared to acute L-DOPA (Figure 6B). In contrast, only 11% of regulated transcripts originating from H3K27me3 positive and H3K27me3S28p negative genomic loci (Figure 6C), and 24% of all regulated transcripts were less expressed in chronic L-DOPA compared to acute L-DOPA. These data suggested that the level of transcription from H3K27me3 and H3K27me3S28p positive genes correlated with the level of induced H3K27me3S28 phosphorylation.
Finally, we examined the individual genes that were ≥1.5 fold induced after chronic L-DOPA and scored positive for H3K27me3 and H3K27me3S28p. We found that a large number of these genes differed from those induced by acute L-DOPA (Figure 6E, Table S1 and Table S2). Thus, a total of 96 genes were induced after acute L-DOPA and 114 genes after chronic L-DOPA, but only 43 genes were commonly induced. To confirm this observation, we performed RT-qPCR on selected genes and could confirm, for instance, that Ppm1n and Galr2 were induced after acute L-DOPA, but not after chronic L-DOPA or saline (Figure 6F; see also Figure S1B showing that the lesion alone does not induce H3K27me3S28p (vehicle control)). Atf3, Klf4, and Npas4 were induced by both acute L-DOPA and, albeit to a lesser extent by chronic L-DOPA (Figure 6G), whereas Nr4a2 (also known as Nurr1), Ngf and Lipg were only induced after chronic L-DOPA (Figure 6H). For these genes we could observe peaks in the ChIP-seq data for H3K27me3 in the lesioned and unlesioned striata, induced H3K27me3S28p in the lesioned striatum and H3K4me3 peaks in the lesioned and unlesioned striata (Figure S7).
As an example of a gene that was induced at the protein level in MSNs expressing the D1-receptor, we stained for Atf3 after chronic L-DOPA stimulation (4 hours timepoint after the last L-DOPA administration) as shown in Figure S6D.
Overall, these data suggested that increased transcription from a subset of PcG-repressed genes was associated with the development of LID. The expression level of H3K27me3 and H3K27me3S28p marked genes was generally lower in chronic L-DOPA stimulated mice compared to acute L-DOPA stimulated mice, correlating with reduced induction of the H3K27me3S28p mark. However, repeated L-DOPA administration resulted in the induction of a unique group of PcG regulated genes that were not induced after a single L-DOPA injection. For the group of genes that gained H3K27me3S28p (subgroup II in Figure 6B) and had significantly higher expression in chronic L-DOPA stimulated lesioned mice compared to acute stimulated mice the most significantly enriched GO-term was “Adult behavior”. However, the term comprises only five genes from subpopulation II (Figure 6B): Nr4a2, Nr4a3, Trh, Npy and Adra1b (out of 93 potential genes in the genome), and although this was a significant (p = 0.0011) 10.7 fold enrichment over the expected number, it did not remain significant when Benjamini-Hockberg corrected for multiple testing.
We have previously shown that H3K27me3S28 phosphorylation causes the displacement of PRC1- and PRC2-complexes from chromatin, leading to expression of a subset of PcG regulated genes in cultured human fibroblasts [18]. Here, we for the first time provide in vivo evidence for the relevance of this mechanism for gene regulation in adult, post-mitotic neurons utilizing a mouse model of PD. We show by genome wide analyses that, in striatal MSNs, signaling through sensitized D1Rs induces H3S28 phosphorylation in the context of H3K27me3 marked genes. Importantly, the H3K27me3S28p mark correlates with a reduction in PcG binding and increased transcription of a subset of genes, several of which have been implicated in neuronal plasticity.
Notably, a systemic injection of a specific D1R agonist was also able to induce the H3K27me3S28p mark in naïve mice. This clearly indicates that the mechanisms described in this study have a general relevance with regard to D1R transmission in the adult brain.
Taking advantage of the pronounced effects produced by dopamine depletion, we mapped putative downstream genomic targets regulated by L-DOPA via D1Rs. Our genome-wide analysis showed that at least 1/3 of all H3K27me3 marked gene loci, that can potentially give rise to a mRNA transcript in the non-repressed state, gained H3K27me3S28 phosphorylation upon acute L-DOPA administration. Most importantly, among regulated transcripts, phosphorylation of S28 in the context of H3K27me3 was a strong indicator of transcriptional activation.
Interestingly, the combined analyses of RNA-seq and ChIP-seq data showed that the majority of transcripts originating from H3K27me3 genomic loci that gained H3K27me3S28 phosphorylation did not change expression upon L-DOPA stimulation. As previously demonstrated, H3K27me3S28 phosphorylation leads to removal of PcG complexes at H3K27me3 marked regions and is considered to derepress the promoter [18]. Therefore it appears that, in order to fully activate genes, H3K27me3S28 phosphorylation requires additional events to take place, which likely would be H3K4 methylation and histone acetylation.
In line with our previous findings, induction of the H3K27me3S28p mark in response to L-DOPA was reduced in 6-OHDA lesioned Msk1−/− mice compared to wt mice (Figure S5). We found that the induction of H3K27me3S28p was less pronounced, but not abolished, in Msk1−/− mice, indicating that in striatal MSNs other histone kinases must be present and actively relay dopamine signaling to chromatin.
The phosphatase responsible for the dephosphorylation of H3K27me3S28p has so far been elusive. Our results, showing that H3K27me3S28 phosphorylation induced by L-DOPA is decreased by abolishing PKA-mediated activation of DARPP-32, pointed to PP1 as a plausible candidate. This idea was corroborated by in vitro and ex vivo experiments showing that PP1 could dephosphorylate H3K27me3S28p (Figure 5B and C). Overall, our results support a model in which inhibition of PP1 via PKA/DARPP-32 works in parallel with Msk1 to promote H3K27me3S28 phosphorylation in response to activation of D1R (see model in Figure 7).
Dyskinesia is a serious motor complication caused by prolonged administration of L-DOPA to parkinsonian patients [9]. It has been proposed that LID depends on the persistent and intermittent hyper-activation of the cAMP/DARPP-32 signaling cascade produced by L-DOPA through sensitized D1Rs. This, in turn, leads to hyper-phosphorylation of histone H3 and aberrant expression of specific genes implicated in dyskinetic behavior [10], [11]. We have previously shown that a T34A mutation on DARPP-32 decreases dyskinetic behaviour in 6-OHDA lesioned mice and that this effect correlates with reduced histone H3S10 phosphorylation [14]. In this study, we found that the induction produced by L-DOPA on H3K27me3S28 phosphorylation at regulatory regions of the Atf3 and Npas4 genes, which are bound and repressed by PcG proteins in MSNs under normal conditions, correlated with increased mRNA- and protein synthesis. We also show that H3K27me3S28 phosphorylation and the associated transcriptional activation are largely reduced in DARPP-32 T34A mutant mice. These observations are particularly interesting in view of the involvement of Atf3 and Npas4 in synaptic plasticity and long-term adaptive responses. Augmented expression of Atf3 in the dorsal striatum has been observed following acute and chronic administration of amphetamine [32], a drug that, similarly to L-DOPA, promotes dopamine transmission. Furthermore, the TF Npas4 has been proposed to promote the expression of other immediate early genes, including Arc, Egr1 and c-Fos [33], which have all been associated to chronic administration of L-DOPA and to the development of LID [15], [34], [35].
Our results show that repeated administration of L-DOPA decreases its ability to induce global phosphorylation of H3K27me3S28 (Figure 6A). This is not surprising since desensitization of kinase signaling pathways following persistent upstream activation and negative feedback is a common phenomenon. Accordingly, previous work showed that repeated administration of L-DOPA to 6-OHDA-lesioned mice is accompanied by a partial normalization of sensitized D1R signaling, reflected in lower levels of Erk activation and reduced global phosphorylation of H3 at S10 [15], [36], [37]. Importantly, despite the attenuated induction of H3K27me3S28 phosphorylation after chronic L-DOPA, our genome-wide analysis showed that there was still a good correlation between the genomic loci gaining the H3K27me3S28p mark and gene induction. This suggested that residual kinase activity was sufficient to drive transcription from PcG-repressed genes in the context of LID (Figure S6A–C).
The decreased ability of L-DOPA to induce H3K27me3S28p observed after chronic administration correlated with reduced expression of transcripts from H3K27me3 and H3K27me3S28p marked genes (Figure 6B), but not of transcripts from H3K27me3 marked genes lacking the H3K27me3S28p mark or transcripts in general (Figure 6C–D). As an exception, IE genes such as Atf3, c-Fos and Egr2-4, which encode TFs, were all induced to at least the same extent after chronic L-DOPA administration compared to acute administration (Table S2). This suggests that these IE genes, which are characterized by transient expression, respond with fast on/off-kinetics presumably due to rapid dephosphorylation of S28 by PP1. This would allow PcG complexes to re-bind, when the signal leading to Msk1 activation ceases, thereby re-setting gene repression until the next activating signal, like for the repeated daily L-DOPA administration in chronically treated parkisonian mice.
The IE genes, Nr4a2 and Nr4a3, also known as Nurr1 and Nor1, encode orphan nuclear receptors that function as TFs. These genes were more strongly induced after repeated L-DOPA administration compared to acute L-DOPA administration. It has been shown previously that the expression of Nr4a2, a gene product involved in the development of dopaminergic neurons [38], is increased in response to prolonged treatment with L-DOPA and that this effect occurs in the MSNs expressing D1Rs [39]. Notably, recent evidence indicates that viral vector-induced overexpression of Nr4a2 in striatal neurons increases dyskinesia in a rat model of PD [40]. Our results are in line with these observations and provide a possible mechanism accounting for the enhanced expression of Nr4a2 in response to chronic L-DOPA administration. In the striatum, Nr4a3 mRNA expression is induced by activation of D1Rs and this effect is prevented by blockade of Erk signaling [41]. Whereas the exact role of Nr4a3 in dopaminergic transmission remains to be elucidated, the present data support the idea that, in PD, this gene is de-repressed via D1R-mediated activation of Erk and increased H3K27me3S28 phosphorylation.
Interestingly, we identified other PcG regulated genes such as neuropeptide encoding genes: galanin (Gal), thyrotropin releasing hormone (Trh) and neuropeptide Y (Npy), which all were only induced upon chronic treatment with L-DOPA (Table S2). In contrast to IE genes, these neuropeptide encoding genes seems to be more tightly regulated, and their activation presumably requires persistent phosphorylation of H3K27me3 marked and PcG bound regions, as well as induction of additional TFs and co-factors to lead to their transcriptional activation.
The regulation of Trh is in line with previous work showing a large increase in the mRNA for this hormone occurring in the striatal MSNs of dyskinetic rats [42]. Subsequent analysis established a correlation between Trh expression in D1R-expressing MSNs and dosage of L-DOPA, which is regarded as a critical factor in the development of dyskinesia [39]. Indeed, enhanced levels of Trh may concur to the development of dyskinesia, since hyperthyroidism is typically associated to hyperkinesia.
Increased levels of NPY mRNA have been found in the striata of Parkinsonian patients treated with L-DOPA [43]. The present data suggest the possibility that this increase may occur in striatal MSNs, although in naïve mice, NPY is mainly expressed in GABAergic interneurons [44]. NPY has been proposed to exert neuroprotective effects on dopaminergic neurons [45], however further studies will be necessary to clarify its potential role in dyskinesia.
Galanin (Gal) and galanin receptors are involved in many neuronal functions including drug addiction [46]. In the striatum, galanin receptors have been localized to cholinergic interneurons and neuronal terminals, which are critically involved in the regulation of the excitability of MSNs [47], [48]. Galanin has also been shown to reduce dopamine release in the striatum [49] and to inhibit spontaneous locomotion by reducing the activity of midbrain dopaminergic neurons [50]. Therefore, the increase in galanin mRNA expression produced by chronic administration of L-DOPA may lead to significant modifications in basal ganglia neurotransmission. Future studies will be necessary to determine the impact of these changes on the development and manifestation of dyskinesia.
The functional implications of differences in reactivation of PcG repressed genes remain to be investigated in more detail. However, it is possible that genes that are activated only in response to acute administration of L-DOPA are implicated specifically in the emergence of LID, whereas genes that are activated only in response to chronic L-DOPA are involved in the consolidation and manifestation of this condition. Interestingly, one of the receptors for the neuropeptide galanin, Galr2 was only induced in response to acute L-DOPA (Table S1 and Figure 6F), while the gene encoding the ligand, Gal, was only induced by chronic L-DOPA stimulation as mentioned above.
Further studies are necessary to examine which permissive determinants are required at genomic sites for H3K27me3S28p mediated expression of PcG target genes. Nevertheless, the fact that many genes regulated by PcG proteins are TFs and among them many are affected by H3K27me3S28 phosphorylation and upregulated by acute and chronic L-DOPA, suggest that LID is, at least in part, a consequence of secondary transcriptional events. These TFs might only need to be induced transiently in order to trigger transcription of other genes that lead to sustained changes in neuronal plasticity associated to long-term maladaptive responses, such as dyskinesia.
Overall our novel findings reveal a previously unrecognized plasticity of PcG-repressed genes in terminally differentiated neurons. The identification of specific genes whose expression is increased upon prolonged treatment with L-DOPA and dopamine D1 receptor stimulation offer a possibility to design novel therapeutic strategies to treat Parkinson's disease and potentially other disorders caused by dysfunctional dopaminergic transmission in the brain, such as drug addiction and schizophrenia.
Female C57BL/6J mice (30 g) were purchased from Taconic (Tornbjerg, Denmark). Bacterial artificial chromosomes transgenic mice expressing EGFP under the control of the promoter for the D2R (Drd2-EGFP) or the dopamine D1R (Drd1a-EGFP) were generated by the Gene Expression Nervous System Atlas program at the Rockefeller University [27] and were crossed on a C57BL/6 background for three generations. DARPP-32 T34A mutant mice [51] and Msk1−/− mice [52] have already been described. The animals were housed in groups of five under standardized conditions with 12 hours light/dark cycle, stable temperature (20°C), and humidity (40–50%). All protocols utilized to generate the model of PD and dyskinesia (including chronic administration of L-DOPA), were approved by the Research Ethics Committee of Karolinska Institutet and by the Swedish Animal Welfare Agency (permit N515/12).
Mice were anesthetized with a mixture of fentanyl citrate (0.315 mg/ml), fluanisone (10 mg/ml) (VetaPharma, Leeds, UK), midazolam (5 mg/ml) (Hameln Pharmaceuticals, Gloucester, UK), and water (1∶1∶2 in a volume of 10 ml/kg) and mounted in a stereotaxic frame (David Kopf Instruments, Tujunga, CA) equipped with a mouse adaptor. 6-OHDA-HCl (Sigma-Aldrich Sweden AB) was dissolved in 0.02% ascorbic acid in saline at a concentration of 3 µg of free-base 6-OHDA per microliter. Each mouse received one unilateral (right hemisphere) injection of 6-OHDA of 1 µl (0.5 µl/min) into the medial forebrain bundle according to the following coordinates (mm): anteroposterior (AP), −1.2; mediolateral (ML), −1.2; dorsoventral (DV), −4.8 (all millimeters relative to bregma) [53]. Noradrenergic neurons were protected by injection of 25 mg/kg desipramine (Sigma) thirty minutes prior to 6-OHDA injection. This procedure leads to a decrease in striatal tyrosine hydroxylase immunoreactivity ≥80% and to a marked akinesia affecting the side of the body contralateral to the lesioned striatum. Animals were allowed to recover for 2 weeks before experimentation.
L-DOPA (10 mg/kg in combination with 7.5 mg/kg benserazide; purchased from Sigma) was dissolved in saline (0.9% NaCl), and injected intraperitoneally in a volume of 10 ml per kilogram of body weight for 1, 3 or 9 days. SKF81297 (3 mg/kg) was purchased from Tocris.
Mouse embryonic stem (mES) cells, wildtype (wt) E14 (provided by Dr. Zhou-Feng Chen and Dr. Helle Færk Jørgensen) and Eed−/− (provided by Dr. Anton Wutz) were cultured on 0.1% (w/v) gelatin-coated plates in ES medium (Glasgow Minimum Essential Medium (Sigma) supplemented with Glutamax-1 (Gibco), non-essential amino acids (Gibco), 50 mM 2-mercaptoethanol, 15% (v/v) ES-cell-qualified FBS (Gibco), and 1% (v/v) penicillin/streptomycin) in the presence of 1,000 U/ml of LIF (Millipore). To induce histone phosphorylation, the mES cells were stimulated with 1 µg/mL anisomycin in DMSO or DMSO only as control.
For ChIP, cells were cross-linked for 10 min at room temperature in culture media containing 1% formaldehyde, 10 mM Hepes (pH 8.0), 0.1 mM EGTA, and 20 mM NaCl. Cross-linking was stopped by addition of glycine to a final concentration of 0.125 M, followed by an additional incubation for 5 min. Fixed cells were washed 3 times with PBS and harvested in SDS lysis buffer (50 mM Tris at pH 8.1, 0.5% SDS, 100 mM NaCl, 5 mM EDTA, 1 mM PMSF, 10 µg/ml leupeptin and 10 µg/ml aprotinin). The cells were then pelleted for 10 min at 2,400 g followed by the same ChIP protocol as for striatal tissue. The included primer sequences are listed in Table S3.
Mice were killed by decapitation, the heads of the animals were cooled in liquid nitrogen for 6 s and the brains were removed. Coronal slices of 1 mm thickness were obtained from a mouse brain dissection matrix (Activational Systems Inc., RBM-2000C), and three striatal punches of 2 mm diameter from sequential slices were dissected out on an ice-cold surface, sonicated in 1% SDS, and boiled for 10 min. Proteins were separated by SDS–polyacrylamide gel electrophoresis and transferred overnight to PVDF membranes (Amersham Pharmacia Biotech, Uppsala, Sweden). The following antibodies were used: H3S28p (Millipore, 07-145), H3K27me3S28p (Hansen lab), histone H3 (Abcam, ab1791), Atf3 (Santa Cruz, sc-188). The Npas4 antibody was provided by Prof. Greenberg, Harvard Medical School, Boston, USA.
Mice were rapidly anaesthetized with pentobarbital (300 mg/kg ip, Sanofi-Aventis, France) and perfused transcardially with 4% (w/v) paraformaldehyde in 0.1 M sodium phosphate buffer (pH 7.5). Brains were post-fixed overnight in the same solution and stored at 4°C. Forty-micrometer-thick sections were cut with a vibratome (Leica, Nussloch, Germany). Free-floating sections were rinsed in tris-buffered saline, permeabilized in 0.2% Triton X-100 in TBS for 20 min and blocked to prevent non-specific binding by incubation in 0.5% Triton X-100, 5% normal goat serum, 1% bovine serum albumin in TBS for 1 hr at RT. Sections were incubated overnight at 4°C with primary antibodies. The following antibodies were used: EGFP (Aves Lab, GFP-1020), NeuN (Abcam, ab138452). Antibodies for histone marks were the same as for Western blotting. Images from the dorsolateral striatum were obtained by sequential laser scanning confocal microscopy (Zeiss LSM 510 Meta). Data was analyzed by one-way or two-way ANOVA when appropriate followed by Tukey's HSD post-hoc test. Unpaired t-test was used when comparing two means. p<0.05 was considered significant.
Naive C57BL/6 mice were killed by decapitation, and the brains were rapidly removed. Coronal slices (250 µm) were prepared with the use of a vibratome (Leica, Nussloch, Germany). Dorsal striata were dissected out from each slice under a microscope. Two slices were placed in individual 5-ml polypropylene tubes containing 2 ml of Krebs-Ringer bicarbonate buffer. The samples were equilibrated at 30°C for 30 minutes, followed by incubation of the slices with either vehicle (DMSO), 100 nM or 1 µM okadaic acid (Sigma) in 2 mL fresh buffer for 50 min. After incubation, the solutions were rapidly removed, the slices were sonicated in 1% SDS, and the samples were analyzed by Western blotting as described.
Ten µg of N-terminal H3 peptides representing the first 40 amino acids of histone H3.1 were used per reaction, either unmodified or modified as follows: S28 phosphorylated (H3S28p) or K27 tri-methylated and S28 phosphorylated (H3K27me3S28p). PP1 (1 unit) was added to the reaction containing 10 µg of specified H3 peptide in a 15 µl de-phosphorylation buffer: 50 mM HEPES, pH 7.5, 100 mM NaCl, 2 mM DTT, 0.01% Brij 35 and 1 mM MnCl2. The de-phosphorylation reaction was allowed to proceed at 30°C, for 30 min and stopped by adding EDTA (pH 8.0) to a final concentration of 5 mM. A fraction of each reaction was spotted on a nitrocellulose membrane corresponding to: 1.0 µg, 0.1 µg and 0.01 µg H3.1 peptide. The membrane was blocked as for standard Western blotting and developed using antibodies for H3S28p (Millipore 07-145, 1∶3,000) and secondary anti-rabbit HRP (Vector Laboratories). Dot-blots were added enhanced chemiluminiscence (ECL) after the last wash and exposures were made using a ImageQuant LAS 4000 camera system. The quantifications were made based on the dots containing 1 µg peptide.
Tissue punches for chromatin preparation was obtained as described for Western blotting. The punches were fixed for 12 min in cold 1% formaldehyde/PBS followed by glycine incubation to stop further cross-linking. The fixed punches were then washed 3× with cold PBS containing phosphatase inhibitors (20 nM okadaic acid, 10 µM NaF) and subsequently snap-frozen for later use. Chromatin immunoprecipitation experiments were performed as described [54] with some modifications: Fixed striatal punches were homogenized in a nuclear extraction buffer (10 mM Tris (pH 8.0), 100 mM NaCl, 2 mM MgCl2, 0.3 M Sucrose, 0.25% IGEPAL CA-630) containing protease inhibitors (1 mM PMSF, 0.1 mM aprotinin, 0.1 mM leupeptin) and phosphatase inhibitors (20 nM okadaic acid, 10 µM NaF), by douncing 15 times using a 2 mL loose grind pestle followed by a 30 min incubation on ice. The homogenate was dounced another 50 times using a 2 mL loose grind pestle for nuclear release, followed by 10 min centrifugation at 2,400 g to pellet nuclei. The extracted nuclei were then lysed in a lysis buffer containing 50 mM Tris-HCl (pH 8.0), 10 mM EDTA, 1% (wt/vol) SDS and protease/phosphatase inhibitors, diluted in RIPA buffer (10 mM Tris-HCl (pH 7.5), 140 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 0.1% (vol/vol) Triton-X-100, 0.1% (wt/vol) SDS, 0.1% (wt/vol) Na-deoxycholate) and the DNA was sonicated to an average size of 300–500 bp using a Bioruptor standard device (Diagenode) (10 cycles 30 sec ON, 30 sec OFF, highest setting).
10 µg anti-rabbit IgG (DAKO), 3 µg of anti-H3 (“GERA”, Hansen lab), 2.5 µg of anti-Rnf2 (“NAST”, Hansen lab) anti-H3K27me3 (9756, Cell Signaling) and 2 µg anti-H3K4me3 (Lys4) (9751, Cell Signaling) and anti-H3K27me3S28p (the specificity of the batch #5 of peptide antigen purified H3K27me3S28p antibody used in this study was tested as shown in Figure S2) was incubated at 4°C with 25 µL washed Dynabeads protein A (Invitrogen) and RIPA in a total volume of 100 µL. The bead-antibody complexes were then incubated at 4°C for 2 h with 20 µL chromatin in a total volume of 250 µL. Beads were washed in 3× RIPA, 1× high salt wash buffer (20 mM Tris-HCl (pH 7.5), 500 mM NaCl, 2 mM EDTA, 0.1% Triton-X-100, 0.1% SDS), 1× LiCl buffer (10 mM Tris-HCl (pH 7.5), 250 mM LiCl, 1 mM EDTA, 1% Na-deoxycholate, 1% IGEPAL CA-630) and 1× TE buffer. After washes, DNA was eluted from beads and de-crosslinked in 20 mM Tris-HCl, pH 7.5, 5 mM EDTA, 50 mM NaCl. 1% (wt/vol) SDS and 50 µg/mL protease K at 68°C overnight. For input, 20 µL chromatin was de-crosslinked in 20 mM Tris-HCl, pH 7.5, 5 mM EDTA, 50 mM NaCl 68°C overnight. ChIP and input DNA was then purified and eluted using Minelute PCR purification kit (Qiagen). Enrichments on selected loci were measured by qPCR, 3 technical replicates, (7500 Fast, Applied Biosystems) relative to a 5-point dilution series of input chromatin. The included primer sequences are listed in Table S3. Student's t test were used to compare means of the different conditions.
ChIPs for each experimental condition was performed in at least triplicates. The resulting immunoprecipitated DNA were pooled and prepared for ChIP sequencing using an Illumina kit according to the manufacturer's guidelines. 2 nanogram of starting material, as determined by PicoGreen concentrations, was used in each case. Sequencing was performed on a Genome Analyzer II (Illumina) at the National High-throughput Sequencing Centre in Copenhagen.
Libraries were de-multiplexed and high quality reads (Chastity score > = 0.6) were aligned to the mouse genome (mm9) using Bowtie [55] allowing up to two mismatches. Reads not aligning uniquely to the mouse genome were removed and only unique reads were used for subsequent analysis. Tracks from single genomic loci were presented using the UCSC Genome Browser (http://genome.ucsc.edu/) [56]. Reads were normalized to a library size of 10M reads and converted to wig-files using the program EaSeq (Lerdrup et al, manuscript in preparation).
All quantitation, scoring, gating, and visualization was done in the program EaSeq (Lerdrup et al, manuscript in preparation). A list of all transcripts including genomic coordinates was derived from Genomatix (see RNA-sequencing for details), and the amounts of ChIP-seq signal at the genomic regions corresponding to −1 kbp to +1 bkp was quantified and normalized to dataset size and a region-size of 1 kbp. Transcripts were scored positive or negative by fitting the abundance to the quantified ChIP-seq signal at all transcripts genome-wide to a normal distribution. A threshold was automatically applied at the level where the only 5% of the regions were expected to score positive by chance (thresholds were 6.8 for left hemisphere H3K27me3 and 16.9 for right hemisphere H3K27me3S28P). Transcripts were gated into subpopulations depending on the level of ChIP-seq signal relative to these thresholds and/or fold change in gene expression as well as significance in Benjamini-Hochberg corrected p-values [57].
One striatal punch per hemisphere was dissected out and subsequently put in RNAlater (Qiagen) at +4°C over night to inhibit RNA degradation. Total RNA was extracted using a RNeasy kit (Qiagen) and quantified on a NanoDrop 1000 device. 200 ng of RNA was used for generation of cDNA using a TaqMan Reverse Transcription Reagents kit (Invitrogen). Expression levels for individual transcripts were measured by qPCR and calculated by the ddCt-method using TATA-binding protein (Tbp) mRNA as housekeeping gene. The expression levels were based on 3 biological replicates. The included primer sequences are listed in Table S3.
A TruSeq RNA Sample preparation kit (Illumina) was used for library generation out of 0.5 µg of total RNA per condition. The generated libraries were sequenced on a Genome Analyzer II (Illumina) at the National High-throughput Sequencing Centre in Copenhagen. Reads were mapped to the mouse genome (mm9) using the Genomatix Mining Station software (Genomatix). Differential expression analysis was done on triplicates using the region miner task “Expression Analysis for RNASeq Data” on a Genomatix Genome Analyzer (Genomatix) using the DeSeq algorithm [58].
Gene ontology (GO) enrichment analyses were done from sets of genes with a significant (Benjanimi-Hochberg corrected for multiple testing) increase or decrease in expression of at least 1.5 fold using the DAVID functional annotation tool at http://david.abcc.ncifcrf.gov/ [59], [60].
The Geo accession number for the ChIP-seq and RNA-seq data reported in this paper is GSE60703.
|
10.1371/journal.pcbi.1003556 | Sensing Membrane Stresses by Protein Insertions | Protein domains shallowly inserting into the membrane matrix are ubiquitous in peripheral membrane proteins involved in various processes of intracellular membrane shaping and remodeling. It has been suggested that these domains sense membrane curvature through their preferable binding to strongly curved membranes, the binding mechanism being mediated by lipid packing defects. Here we make an alternative statement that shallow protein insertions are universal sensors of the intra-membrane stresses existing in the region of the insertion embedding rather than sensors of the curvature per se. We substantiate this proposal computationally by considering different independent ways of the membrane stress generation among which some include changes of the membrane curvature whereas others do not alter the membrane shape. Our computations show that the membrane-binding coefficient of shallow protein insertions is determined by the resultant stress independently of the way this stress has been produced. By contrast, consideration of the correlation between the insertion binding and the membrane curvature demonstrates that the binding coefficient either increases or decreases with curvature depending on the factors leading to the curvature generation. To validate our computational model, we treat quantitatively the experimental results on membrane binding by ALPS1 and ALPS2 motifs of ArfGAP1.
| Selective targeting of soluble proteins to cellular membranes relies on different mechanisms such as receptor-mediated recruitment or direct binding to specific lipids. A new paradigm has been recently proposed, according to which membrane binding of some proteins is driven by the geometrical and physical properties of the membranes, namely the membrane curvature and lipid packing in the external membrane monolayer. Specifically, several proteins referred to as the membrane curvature sensors have been shown to preferentially bind strongly curved membranes. This mode of protein binding is especially relevant for such fundamental cell processes as endocytosis and carrier generation from ER and Golgi Complex, which involve shaping initially flat membranes into strongly curved ones. A subset of the curvature sensors contains amphipathic or hydrophobic domains that shallowly insert into the membrane. Here we explore computationally the detailed physical mechanism underlying the membrane binding by such proteins and demonstrate that their membrane affinity is not determined by the curvature per se but rather by the membrane stress, independently of the way the stress has been generated. Hence, the significance of our work is in elucidating the relationship between the membrane binding of peripheral proteins carrying shallowly inserting domains and the membrane stresses.
| Lipid bilayers serving as matrices of biological membranes bear internal elastic stresses. These stresses can be generated by external forces applied to the membrane surface and driving overall membrane deformations such as generation of membrane curvature and stretching-compression of the membrane area [1], and/or by internal factors such as elastic frustrations, which are intrinsic to the membrane structure [2].
Insertion into the membrane matrix of protein domains spanning completely or partially the lipid bilayer interior must interfere with the intra-membrane stresses. This has to result, on one hand, in the stress-dependence of the energy of the protein insertion into the membrane and, on the other, in alteration of the intra-membrane stresses. The former phenomenon results in the stress sensing by these protein domains, which can be manifested as stress-dependence of the protein partitioning between the membrane and the surrounding aqueous solution [3] and/or as regulation by the stresses of the protein conformational transitions and the related protein activity within the membrane (see e.g. [4]). Alteration of the membrane stress caused by the protein embedding can affect the membrane conformation, e.g. by changing membrane curvature [5], [6].
During the last decade, one of the hot topics discussed in the biophysical literature, and referred to as the curvature sensing by proteins, has been the ability of a number of peripheral membrane proteins to bind preferentially to small liposomes with radii of several tens of nanometers. Commonly for most of these curvature sensing proteins, their binding to membranes has been mediated by shallow insertion into the membrane matrix of an amphipathic or hydrophobic domain [7]. In most cases, such domain is an amphipathic helix [8], but can also be a short hydrophobic loop [9]. The curvature sensing has been demonstrated for numerous proteins involved in intracellular membrane shaping and remodeling such as the N-BAR (Bin-amphiphysin-Rvs) domain-containing protein amphiphysin playing a key role in endocytosis [10]; the GTPase dynamin driving membrane fission [9]; synaptotagmin implicated in membrane fusion [11]; α-synuclein [12]; the lipid droplet enzyme CTP∶phosphocholine cytidylyltransferase (CCT) [13], synapsin I [14], and the autophagosomal protein Barkor/Atg14(L) [15]. A most thorough study of curvature sensing has been performed for a particular kind of amphipathic helices contained in proteins such as Arf1 GTPase-activating protein (ArfGAP1), responsible for the disassembly of the COPI coat [16]–[19]; the golgin GMAP-210; the sterol sensor/transporter Osh4p/Kes1p; and the nucleoporin Nup133 [18]. These helices, which are characterized by bulky amino acids in the non-polar face and small uncharged amino acids in the polar face (mainly serine and threonine), have been demonstrated not only to bind, selectively, to highly curved membranes of small liposomes [16], but also to sense mismatch between the actual membrane curvature and the curvature preferred by the specific lipids composing the outer membrane monolayer [17]. This led to the suggestion that these helices sense membrane curvature by recognizing the curvature dependent defects in lipid packing (see [7], [20], [21] and references therein) and to calling them the amphipathic lipid-packing sensors (ALPS). Since generation and alterations of membrane curvature as well as formation of the lipid packing defects are intimately related to the intra-membrane stresses, it is reasonable to expect that the observed apparent curvature sensing by the insertion-containing proteins is a manifestation of a more general phenomenon of intra-membrane stress sensing.
The goal of the present work is to analyze quantitatively the interplay between the protein insertions imitating amphipathic helices and the membrane stresses produced either by the intra-membrane elastic frustrations or by the external forces leading to different kinds of overall membrane deformations including curvature generation. The major statement of the work is that binding of the insertion-containing proteins to the membrane depends primarily on the local intra-membrane stresses existing within the region of the protein embedding, rather than on the way these stresses have been generated. Particularly, concerning the suggested curvature sensing, we predict that the insertion-containing proteins can exhibit similar binding to membranes of different curvature provided that the membrane stresses in the protein-embedding region are similar. Conversely, these proteins are predicted to bind differently to membranes of similar curvature provided that this curvature is achieved by diverse combinations of the intra- and extra-membrane forces and, hence, corresponds to different intra-membrane stresses. Hence the shallow insertions have to be seen as sensors of the intra-membrane stress rather than the membrane curvature. We substantiate our conclusions by demonstrating that the computational approach we use provides quantitative description of the experimental results on differential binding of ALPS domains to liposomes of various diameters and diverse lipid compositions.
We consider amphipathic helix-like protein domains shallowly embedded into the membrane matrix and refer to these domains as the protein insertions. We model such insertions as cylindrical rods of about one nanometer cross-sectional diameter embedded into the outer membrane monolayer such that the rod axis lies parallel to the membrane plane. The typical embedding depth is about 40% of the monolayer thickness [22] (Fig. 1, left cartoon).
We define as the membrane stress sensing by protein insertions the dependence of the insertion binding to the membrane on the membrane stress. To formalize this definition, we consider a system consisting of Np protein insertions partitioning between the aqueous solution and the outer monolayers of lipid membranes, which are subject to elastic stresses and can have curved shape. We quantify the insertion binding to the membranes by the binding constant, KB, defined as a ratio between the number of the insertions remaining in the aqueous solution, , and the number of the membrane bound insertions, ,(1)Since the binding constant is measurable experimentally [18], the dependence of KB on the membrane stress is a convenient quantitative measure of the stress sensing.
The physical reason for the stress sensing is the dependence of the total free energy of the insertion binding, εbind, on the membrane stress. The binding energy εbind, which is determined as the change of the free energy of the whole system resulting from one insertion binding, has a major contribution, ε0, from a number of essential membrane-insertion interactions such as the hydrophobic, hydrogen bonding and electrostatic interactions. On top of that, embedding of the insertion generates intra-membrane strains and the related change of the membrane elastic energy, εel, such that,(2)Below, εel will be referred to as the elastic binding energy. It follows from thermodynamics of the insertion binding (see Text S1) that for the insertion number, Np, much smaller than the numbers of water, Nw, and lipid, Nl, molecules, , , the binding constant can be presented as(3)where , is the stress-independent part of the binding constant, R is the radius of the membrane mid-plane and δ is the monolayer thickness. The correction is relevant only if the radius is so small that the difference between the amounts of the lipid molecules in the outer and inner membrane monolayers becomes considerable.
Eq. 3 determines a strong exponential dependence of the binding constant on the elastic binding energy, εel, which, in turn, depends on the values of the intra-membrane stresses and their distribution over the membrane thickness.
To reveal the factors that determine the elastic binding energy, εel, we dissect the embedding event into two steps. The first step is embedding of the insertion into the membrane while keeping the initial membrane shape unchanged. The variation of the membrane elastic energy at this stage, εV, is related to creation of a void in the membrane matrix necessary for the insertion accommodation (Fig. 1). This is accompanied by perturbation of the strains and stresses within the membrane. The second step is a partial relaxation of the stress perturbation due to the change of the membrane shape, which is accompanied by another change of the elastic energy, εR (Fig. 1).
The energy of the first step, εV, can be seen as thermodynamic work performed against the membrane stresses in the course of the void generation, which can be presented as a sum of two contributions,(4)In essence, is the work of the void formation performed against the initial stresses existing in the membrane before insertion, while accounts for the energy of the stress perturbation.
Summarizing, the elastic binding energy can be presented as consisting of three contributions(5)
As shown below, in all relevant cases represents the major part of the energy of void formation. In addition, turns out to be the most convenient value accounting for the distribution of the initial unperturbed stresses in the context of the void formation. Therefore, we will refer to as the void energy and use it as a variable characterizing the stressed state of the membrane before insertion.
Since the protein insertions we are considering do not span the whole membrane but rather get shallowly embedded into the membrane matrix, the total energy of the void formation, εV, and hence the elastic binding energy εel, depend on the character of the stress distribution through the membrane thickness. This distribution can be described by the trans-membrane stress profile σ(z) [2] (Fig. S1). In Text S1 we discuss the model assumptions concerning the properties of the trans-membrane stress profile and the relationships between σ(z) and the overall force factors determining the membrane stressed state, namely, the lateral tension γ and the bending moment τ. These determine the ways of the stress profile generation by application to the membrane of external forces or by changing the monolayer spontaneous curvatures through variations of their lipid compositions (see for review [23], [24] and references therein). Specifically, the void energy representing, as mentioned above, the thermodynamic work against the initial stresses needed for the void formation, can be related to the initial stress profile existing within the outer membrane monolayer before the insertion embedding,(6)where the integration is performed over the volume of the void.
Whereas, according to Eq. 6, the void energy will be calculated by a direct integration of the initial stress profile, the additions to the energy related to the emerging strains, , and the energy of relaxation εR, require a more involved numerical computation including determination of the relaxed membrane shape and will be performed based on the relationships Eqs. S10–S13 using Comsol Multiphysics [5].
We address the sensing by the protein insertions of the intra-membrane stress generated by several specific ways that are experimentally feasible and biologically relevant.
First, we consider the stress resulting from the spontaneous curvature Js, which is produced in the membrane monolayers by changing the monolayer lipid composition (see Text S1) and consider three different situations:
Second, we consider application of an external torque to an initially flat bilayer, which results in generation of a bilayer curvature J (see Text S1). This corresponds to the experimental procedures of generation of small liposomes in vitro by means of sonication or extrusion. The external torque produces in the outer monolayer a trans-monolayer stress profile with a bending moment(10)while the stress-profile in the inner monolayer corresponds to a bending moment (Fig. 2D).(11)
Finally, we analyze the case where the membrane stress is produced by applying a stretching force to the flat monolayer, which generates an overall lateral tension, γ, related to the trans-membrane stress profile (see Text S1) (Fig. 2E). Such a force can be produced as a result of, e.g., osmotic stretching of the liposomal membrane.
We model the membrane as consisting of two monolayers each characterized by a bending modulus of [25]. The monolayers can be laterally uncoupled, meaning that there is a reservoir of material for each monolayer with which the lipid molecules can be exchanged. This is the case in most of the biologically relevant situations where the insertions are restricted to a small membrane patch for which the surrounding membrane plays a role of a lipid reservoir. Alternatively, the monolayers can be laterally coupled if, e.g., some rigid barrier restricts the lipid exchange between the membrane patch accommodating the insertions and the surrounding membrane, or if the proteins are recruited to the entire area of a closed membrane, so that there are no lipid reservoirs to exchange with, as it occurs in common in vitro assays.
An insertion is modeled as a rigid cylindrical rod with a radius of 0.5 nm that partially embeds into the outer membrane to a depth of 0.8 nm, which imitates the typical size and insertion depth of amphipathic helices [22]. In general, we consider the length of the insertion along the membrane plane to be 2 nm, characteristic for some amphipathic helices [22], [26].
Our goal is to describe quantitatively the membrane stress sensing by insertions in all five above-mentioned cases of membrane stress generation. We will compute the dependence of the insertion binding constant, KB, Eq. 3, on the void energy, , Eq. 6. The absolute value of the binding constant depends, according to Eq. 3, on the stress-independent factor B, whose value is unknown since it accounts for a combination of the electrostatic, hydrophobic and hydrogen-bonding interactions between the membrane and the insertion, , (Eq. 2). We will therefore compute the relative binding constant, where characterizes the insertion embedding into the initial unstressed flat membrane. The relation between and the elastic binding energy is(12)where is the elastic binding energy prior to the stress generation and R is the curvature radius of the membrane in the stressed state under assumption that the membrane shape is spherical.
For the cases where generation of the membrane stress is accompanied by membrane curvature variations, we will illustrate the relationship between the stress sensing and the earlier suggested curvature sensing by presenting the binding constant, , as a function of the curvature.
The equilibrium distributions of the membrane stresses and strains before and after insertion embedding have been found by solving the set of partial differential equations for the displacement field, and , as explained in [5]. Briefly, for the case of two-dimensional deformations where the membrane adopts a tubular shape with the y-axis laying along the tube, the equations to be solved are(13)where , , and are the trans-monolayer profiles of the monolayer elastic moduli [5], [27]. This approach accounts in a continuous manner for variations of the local pressures and elasticities at distances of sub-nanometer scale. This is equivalent to usage of intra-membrane force field for modeling membrane processes by molecular dynamic simulations, which proved to provide a quantitative description of the membrane behavior. The implemented trans-membrane pressure and elasticity profiles represent a simplified version of those computed recently by the state-of-the-art molecular dynamic simulation using Martini force field [27]. Therefore, our predictions are expected to be of at least semi-quantitative accuracy. For a further discussion about the advantages and disadvantages of both continuum and simulation approaches see Ref. [28].
The equations (Eq.13) were solved for a membrane element of length L and thickness 2h, where h is the monolayer thickness, with the following boundary conditions. First, the insertion is assumed to be much more rigid than the lipid bilayer and hence imposes a horizontal displacement that corresponds to the insertion shape. Second, the top and bottom surfaces of the bilayer are set free, implying that the stresses vanish there. Finally, the right boundary for each monolayer is a symmetry plane, which remains straight but can rotate with respect to the left boundary and can also have a certain constant displacement in both the horizontal and vertical directions. The rotation angle and the displacements are found from minimization of the elastic free energy change upon insertion.
The set of equations (Eq.13) was solved by a finite element method scheme using the commercial software Comsol Multiphysics, allowing one to represent the membrane deformation, calculate the elastic free energy change upon insertion, as well as the void energy. The membrane shape was discretized for the finite element method using a triangular mesh starting with at least 1908 elements, and refined using and adaptive mesh refinement to at least 5514 elements.
For simulation of an initial membrane stress created by a combination of different monolayer spontaneous curvatures, the lateral stress profile has been taken according to Eq. S17. For simulation of the application of an external torque, a constant torque has been applied to the right boundary of the bilayer for both laterally coupled and uncoupled monolayers. Finally, a constant force perpendicular to the right boundary has been applied to simulate the case of a stretched or compressed membrane. In all these cases, the free energy minimization has been acquired by taking into account the work of deformation produced by the externally applied forces.
The computed dependences of the elastic binding energy, , and the relative binding constant, , on the void energy, , for all five aforementioned scenarios of the stress generation are presented in Fig. 3.
Remarkably, the results for obtained for different ways of the stress generation collapse to a single straight line with a slope equal to one (Fig. 3A). This infers, based on Eq. 5, that the contributions from the stresses emerging in the course of insertion, , and the shape relaxation, , have a negligibly small dependence on . While the void energy is determined solely by the intra-membrane stresses preceding the insertion, and are expected to depend on the scenario of the stress generation. Although is expected to be a small correction to the binding energy, the shape relaxation part of the binding energy, , is of the same order of magnitude as the void energy . Our results show that the shape relaxation is independent of the stress distribution along the membrane in the initial state. As a result, the elastic binding energy is practically independent of the way the stress is produced.
Also the dependences of the relative binding constant, , on the void energy, , computed for the five scenarios of the stress generation collapse to a unique curve described by the exponential function (Fig. 3B). According to Eqs. 12 and 5, this is the result of the above-obtained negligibility of the dependences of the energies and on the void energy, , and of the smallness in most cases of the ratio between the membrane thickness and the curvature radius, . Hence, also the relative binding constant, , quantifying the stress sensing does not depend on the scenario of stress generation.
Amphipathic helices of different proteins have various dimensions and could, potentially, get embedded to different depths into the lipid monolayer matrix. Insertion induced curvature depends substantially on the insertion size and the embedding depth [5], which indicates that also the elastic binding energy, , and, hence, the relative binding constant, , may depend on these parameters.
Fig. 3C presents a comparison of the computed dependences of and on the void energy, , for insertions with cross-sectional radii of 0.75 nm and 0.5 nm for the five ways of stress generation. Both types of insertions are assumed to be embedded to the same depth of 0.8 nm, and have the same length of 2 nm. Fig. 3D shows the results obtained for different embedding depths. In both cases, whereas the values of the elastic binding energy do depend on the insertion radius (Fig. 3C, left panel) and embedding depth (Fig. 3D, left panel), the variation of as a function of is always represented by a straight line with slope equal one. This means that although the stress-independent part of the elastic binding energy varies with the insertion size and the embedding depth, the stress-dependent part does not and is, practically, equal to . As a result, the dependence of the relative binding constant, , on the void energy, , and, hence, the stress sensitivity are independent of the cross-sectional radius of the insertion (Fig. 3C, right panel) and of the embedding depth (Fig. 3D, right panel).
Summarizing, the protein insertions are predicted to be universal sensors of membrane stresses existing in the region of the insertion embedding.
In three out of five considered scenarios of the stress generation, building up of the stress is accompanied by emergence of membrane curvature. Fig. 4 presents examples of the computed shapes, which are adopted by an initially flat bilayer as a result of the stress generation (left panels) followed by the insertion embedding (right panels). If the stress is produced as a consequence of inducing the spontaneous curvature in the outer (Fig. 4A) or inner (Fig. 4B) monolayer, or by application of an external torque (Fig. 4C), the membrane acquires curvature prior to the insertion embedding such that the insertion interacts with a bent membrane. In case the stresses result from the spontaneous curvature induced symmetrically in both monolayers (Fig. 4D), or from an overall membrane stretching (Fig. 4E), the insertions get embedded into a flat membrane, whereas the curvature builds up only at the latest stage as a result of the shape relaxation.
Since, as emphasized in the Introduction, an extended literature has been devoted to the curvature sensing by proteins, we show here the correlation between the strength of the insertion binding and the membrane curvature existing prior to the insertion embedding for the three relevant scenarios of the stress generation (Fig. 4A–C).
Fig. 5A,B presents the dependence of the elastic binding energy, , and the relative binding constant, , on the pre-insertion membrane curvature, J. We performed our analysis for a large range of membrane curvatures in order to take into account the highly curved membranes generated, e.g., during transport carrier formation and endocytosis. If the stress is generated by inducing the spontaneous curvature of the inner monolayer or by application of an external torque, the elastic binding energy, , decreases and the binding constant, , increases with the curvature. Thus, the insertion binding is predicted to be stronger for small rather than large liposomes, in agreement with the experimental results of ALPS binding [17]. Opposite prediction corresponds to the case where the stresses are produced by the spontaneous curvature generation in the outer monolayer. In this situation, , increases and the binding constant, , decreases with growing curvature meaning that the insertions are expected to bind stronger to large rather than to small liposomes. These results can be qualitatively understood by considering the stresses in the external part of the outer monolayer for the different scenarios of curvature generation and their influence on the elastic binding energy .
Hence, the dependence of the insertion binding on the membrane curvature and, therefore, the apparent curvature sensing, depend on the way the curvature is produced and can be opposite for different scenarios of curvature generation.
It is convenient to quantify the apparent curvature sensitivity by the slope, , of the line representing, approximately, the elastic binding energy as a function of the membrane curvature J, so that . Following Eq. 12, and under the assumption of smallness of the membrane curvature with respect to the inverse monolayer thickness, , the relative binding constant can be then expressed as . Positive values of the curvature sensitivity, , correspond to preferable insertion binding to membranes with larger curvature (small liposomes), while negative curvature sensitivity, , means preferable binding to membranes with smaller curvature (large liposomes).
Based on the results above, the sign of the curvature sensitivity is determined by the way the membrane curvature is produced. The absolute value of the curvature sensitivity depends on the insertion cross-sectional radius and the embedding depths. These dependences are presented in Fig. 5E,F for the three ways of stress generation leading to membrane curvature. The model predicts that the absolute value of the curvature sensitivity increases with the insertion cross-sectional radius (Fig. 5E). Interestingly, is predicted to change non-monotonously as a function the embedding depth (Fig. 5F). It reaches a maximum for the positive, , and minimum for the negative, , values of the curvature sensitivity at some intermediate embedding depths, the latter varying between the different ways of curvature generation (see Fig. 5F).
Summarizing, the protein insertions cannot be considered as universal curvature sensors since the character of the curvature sensing depends on the specific curvature generating factors.
To validate our proposal of the intra-membrane stress sensing by protein insertions, we used the suggested computational model to treat the quantitative experimental results on membrane binding by the two ALPS motifs of ArfGAP1, ALPS1 and ALPS2, which fold within membranes into amphipathic helices. These studies address the dependence of the ALPS binding on the liposome radius [19] and lipid composition [29].
The quantity measured in [19] was the percentage of ALPS1 and ALPS2 amphipathic helices bound to liposomes of 34 nm, 42 nm, and 90 nm radii. Based on these data, we first found the values that can be obtained from the experimental data and accessible to determination by our model. The absolute values of the binding constant, , are unaccessible, since they depend on the part of the binding energy, , accounted by the parameter B (see Eq. 3) that is not stress-dependent but rather determined by a combination of strong interactions such and hydrophobic, electrostatic and hydrogen-bonding interactions. However, we can eliminate the unknown stress-independent parameter B by using the relative binding constants and , presented in Table 1, and compare them with the experimental results.
To obtain the corresponding values computationally, we took into account several aspects of the experimental system [19]. First, the liposomes were formed by extrusion or sonication meaning that the bending moment and the corresponding curvature, J, were generated by an external torque applied to the membranes. Second, the protein motifs were inserted along the whole membrane area rather than locally [19]. Therefore, the liposome monolayers must be seen as laterally coupled since they could not exchange lipid molecules with any lipid reservoir. The modifications of the computational method needed to account for the monolayer coupling were introduced in [5]. Finally, the length and, especially, the embedding depth of ALPS1 and ALPS2 amphipathic helices could be estimated based on structural data but have not been precisely determined and, therefore, had to be considered as fitting parameters.
Fig. 6 presents the dependences of the computed binding constant ratio on the liposome radius R for different insertion lengths (Fig. 6A) and embedding depths (Fig. 6B). The cross-sectional radii of the amphipathic helices were taken to be 0.5 nm for all computations. Fitting the computed values (Fig. 6A,B) to those derived from the experiments (Table 1), we find for each ALPS motif the relationship between the length and embedding depth guaranteeing a quantitative agreement between the experimental and theoretical results (Fig. 6C). The expected lengths of the amphipathic helices, estimated based on the structural data, vary between 4–6 nm for ALPS1 and 3–4 nm for ALPS2 [19] as presented by the shaded region in Figure 6C. Comparison of the computed and the expected values (Fig. 6C) predicts that ALPS1 and ALPS2 amphipathic helices embed to a depth close to 0.4 nm with a tendency of ALPS1 to penetrate the membrane a little deeper than ALPS2.
In [29] the ArfGAP1 ALPS binding was studied in dependence on the membrane lipid composition, which was modified by symmetric addition to the two membrane monolayers of diacylglycerol (DAG) and phosphatidylethanolamine (PE), the lipids generating a negative monolayer spontaneous curvature [23]. The percentage of the membrane bound ArfGAP1 was measured as a function of the mole fraction of these lipids within the membrane. As explained above (see also Text S1) and according to Eq. 9, symmetric generation of spontaneous curvature of the membrane monolayers leaves the membrane flat but produces stresses in each monolayer. These stresses are expected to modulate the amphipathic helix binding.
To enable the comparison of the experimental results [29] with the model predictions, we first plot the measured fraction of bound protein as a function of the monolayer spontaneous curvature, , (Fig. 7A). The latter is assumed to be related to the monolayer lipid composition by the relationship, , where and are, respectively, the spontaneous curvature and the intra-monolayer area fraction of the individual lipid components and the summation is performed over all lipid components [30]. The area fractions of the constituent lipids are taken from [29], while the lipid spontaneous curvatures are taken to be (see [23] and references therein).
A convenient quantity to be derived from the experimental data is the ratio between the protein binding constants for certain monolayer spontaneous curvatures , and for the background spontaneous curvature of corresponding to 70% PC and 30% PS. The values of this ratio in dependence on are presented in Table 2 and Fig. 7B. The same ratio of the binding constants was computed based on our model using the insertion length and the embedding depth as fitting parameters and assuming, as mentioned above, that the monolayers are laterally coupled. The relationship between the ArfGAP1 insertion length and its embedding depth that best fits the experimental data in Fig. 7B is presented in Fig. 7E where the shaded region corresponds to the feasible values of these parameters. According to these results, for a realistic total insertion length of 4 nm, the required embedding depth is about 0.4 nm, which is consistent with the above estimations (Fig. 6).
It has been proposed and extensively discussed in the literature that some peripheral membrane proteins are able to sense large membrane curvatures [7]. In the experimental studies devoted to verification of this idea, the curvature sensing was manifested by a preferential binding of such proteins to small liposomes of few tens of nanometer radii [17], [31]. The reason for the attractiveness of the concept of curvature sensing by proteins is a straightforward and, therefore, feasible mechanism it suggests for interplay between the geometry and protein composition of cell membrane patches. Such interplay including a positive feedback between the membrane bending and the local protein concentration may have far reaching consequences for the mechanisms of such intra-cellular processes as endocytosis [32] and generation of intra-cellular membrane carriers from the endoplasmic reticulum and the Golgi complex [33], which involve membrane shaping and remodeling by proteins [34].
Two classes of protein domains have been proposed to sense membrane curvature: hydrophilic intrinsically curved domains, such as BAR domains, able to bind the membrane surface and referred to as the membrane scaffolds [10]; and small amphipathic or hydrophobic domains, such as amphipathic α-helices, which get shallowly embedded into the lipid monolayer matrix and are referred to as the hydrophobic insertions [16].
The potential importance of the curvature sensing by proteins raises a question about the mechanism of this phenomenon. The mechanism by which the protein scaffolds sense membrane curvature is straightforward and related merely to the membrane bending energy. The closer the membrane curvature is to that of the scaffolding protein domain, the less membrane bending deformation is required for the protein attachment to the scaffold and, hence, the less bending energy is consumed for the attachment event making it more energetically favorable. Hence, the scaffolding protein domains must sense the membrane curvature per se.
The situation with the hydrophobic insertions, which are not characterized by a curved shape and penetrate the membrane interior rather than stick to the membrane surface, appears to be more complicated. The mechanism of curvature sensing by the insertions has to be related to the internal membrane stresses, which can arise from various membrane deformations rather than, solely, from the overall membrane bending. Here we analyzed numerically the changes of the membrane elastic energy related to the insertion embedding with a goal to understand whether the insertions sense, indeed, the membrane curvature per se, or, alternatively, they sense the intra-membrane stresses independently of the way the stresses are generated.
A protein domain can be considered as a curvature sensor per se if its binding to the membrane is influenced by the membrane curvature, and the curvature-dependence of the binding coefficient is the same for different ways of the membrane curvature generation. Our calculations showed that this is not the case for the protein domains, such as amphipathic α-helices, which get shallowly inserted into the membrane matrix. As illustrated in Fig. 5A the binding constant of such domains increases with increasing curvature for the cases where the curvature is produced by an externally applied torque (black asterisks), or by addition of lipids with negative spontaneous curvature to the inner monolayer (blue tilted squares), but decreases if the curvature is produced by addition of lipids with a positive spontaneous curvature to the outer monolayer (red squares). Hence the curvature sensing is not a universal property of the protein insertions.
At the same time, according to our model, the protein insertions are universal sensors of the intra-membrane stresses within the region of the insertion embedding. The dependence of the insertion binding coefficient on these stresses does not depend on the way the stresses are generated (Fig. 3A). The mechanism of this stress sensing is based on the elastic energy coming from formation of a void in the membrane matrix necessary to accommodate the insertion (Fig. 1). The thermodynamic work of the void formation is performed against the internal stress existing within the membrane matrix, which is equivalent to an intra-membrane pressure (taken with opposite sign). As a result the mechanism of the stress sensing by hydrophobic insertions can be seen as a “pushing the walls” mechanism.
It has to be noted that the stress-sensing mechanism must underlie also the curvature sensing by transmembrane proteins spanning the whole membrane thickness [35]. Distribution of transmembrane proteins between different regions of the same membrane must be determined by the thermodynamic work, which has to be performed against the intra-membrane stresses in order to create a void accommodating the protein. In case this work varies along the membrane, the transmembrane proteins must partition accordingly. A specific example of such situation is lateral partitioning of trans-membrane proteins characterized by asymmetric cone-like effective shapes along membranes with varying curvature [35] or surface concentration of non-bilayer lipids. The difference between trans-membrane proteins and shallow insertions is that the curvature sensitivity by the former cannot be related to the protein binding coefficient since such proteins are very hydrophobic and, therefore, insoluble in aqueous solutions.
The suggested mechanism changes considerably the view on the potential role of proteins domains serving as hydrophobic insertions in the protein targeting and the mode of their action in membrane shaping processes. Further in vitro experimentation aimed at quantitative characterization by biochemical methods of binding of different proteins containing hydrophobic insertions to liposomes of different lipid composition, curvature, or membrane tension, would provide stronger evidence of the stress-sensing mechanism proposed here.
|
10.1371/journal.pntd.0001645 | Dengue-1 Envelope Protein Domain III along with PELC and CpG Oligodeoxynucleotides Synergistically Enhances Immune Responses | The major weaknesses of subunit vaccines are their low immunogenicity and poor efficacy. Adjuvants can help to overcome some of these inherent defects with subunit vaccines. Here, we evaluated the efficacy of the newly developed water-in-oil-in-water multiphase emulsion system, termed PELC, in potentiating the protective capacity of dengue-1 envelope protein domain III. Unlike aluminum phosphate, dengue-1 envelope protein domain III formulated with PELC plus CpG oligodeoxynucleotides induced neutralizing antibodies against dengue-1 virus and increased the splenocyte secretion of IFN-γ after in vitro re-stimulation. The induced antibodies contained both the IgG1 and IgG2a subclasses. A rapid anamnestic neutralizing antibody response against a live dengue virus challenge was elicited at week 26 after the first immunization. These results demonstrate that PELC plus CpG oligodeoxynucleotides broaden the dengue-1 envelope protein domain III-specific immune responses. PELC plus CpG oligodeoxynucleotides is a promising adjuvant for recombinant protein based vaccination against dengue virus.
| Dengue is a mosquito-borne disease. Infection of dengue virus can cause clinical manifestations ranging from self-limiting dengue fever to potentially life-threatening dengue hemorrhagic fever or dengue shock syndrome. In recent years, dengue has spread to most tropical and subtropical areas, making it a global health concern. Specific approaches for dengue therapy do not exist; the development of a dengue vaccine would represent a major advance in the control of the disease. Currently, no licensed dengue vaccine is available. Subunit vaccines provide a great safety strategy for developing dengue vaccine. However, the major weaknesses of subunit vaccines are low immunogenicity and poor efficacy. Here we employed dengue-1 envelope protein domain III as a model vaccine candidate and described a newly developed water-in-oil-in water multiphase emulsion system to overcome the inherent defect of subunit vaccines. We showed that emulsification of dengue-1 envelope protein domain III and CpG oligodeoxynucleotides synergistically broadened immune responses and potentiated the protective capacity of dengue-1 envelope protein domain III. These results provide valuable information for development of recombinant protein based vaccination against dengue virus and future clinical studies.
| Dengue is the most important mosquito-borne flavivirus disease. People living in the tropical and subtropical areas are at risk of dengue virus infection, and more than 50 million dengue infected cases occur worldwide each year [1], [2]. Vaccine inoculation is a cost-effective way of combating the threat of infectious diseases. In the past six decades, tremendous effort has been made to develop a dengue vaccine [3]–[5]. However, despite these efforts, no licensed dengue vaccines are currently available.
Many advanced biological technologies have been applied to dengue vaccine development, and numbers of vaccine approaches are currently in pre-clinical or clinical development. These approaches include chimerization with other flaviviruses or the deletion of portions of the genomes to obtain live attenuated dengue vaccines, viral vector vaccines, DNA vaccines, and recombinant subunit vaccines [3]–[5]. All of the approaches are associated with different advantages and disadvantages. Among these approaches, the recombinant subunit vaccine provides the greatest degree of safety.
Dengue envelope protein domain III has been shown to be involved in host receptor binding [6], [7], and several neutralizing epitopes have been identified within this domain [8]–[13]. These characteristics of the envelope protein domain III indicate that it would be a promising dengue vaccine candidate [14]. Several subunit vaccines based on recombinant dengue envelope protein domain III have been developed to protect against dengue viral infection [15]–[23]. Formulating dengue subunit vaccine candidates with proper adjuvants [15]–[17], [21]–[23] or expressing vaccine candidates in a lipoprotein [18]–[20] was necessary to enhance their immunogenicity. These results indicate that one of the major weaknesses of subunit vaccines is their low immunogenicity and that appropriate adjuvants or delivery systems are required to overcome this weakness.
Adjuvants and delivery systems have noticeably improved over the past several years. We previously developed a bioresorbable diblock tri-component copolymer poly(ethylene glycol)-block-poly(lactide-co-ε- caprolactone) mixed with squalene and Span®85 to produce homogeneous nano-particles (PELC). This water-in-oil-in-water multiphase emulsion system can be utilized for vaccine delivery [24]–[26]. In addition, we demonstrated that a formulation of inactivated influenza virus and CpG oligodeoxynucleotides (CpG) could enhance both the overall immune response and cross-clade protective immunity [27]. These results indicate that PELC-formulated vaccines has improved potential efficacy.
In this study, we evaluated the potential of aluminum phosphate, CpG, PELC, and PELC plus CpG as adjuvants to enhance the immunogenicity of recombinant dengue-1 envelope protein domain III (D1ED III). We demonstrated that recombinant D1ED III formulated with PELC plus CpG induced stronger and broader immune responses than using other adjuvant formulations. These results provide valuable information for future clinical studies.
Animal studies were carried out in strict accordance with the recommendations from Taiwan's Animal Protection Act. The protocol was approved by the Animal Committee of the National Health Research Institutes (Protocol No: NHRI-IACUC-098014) and were performed according to their guidelines.
A consensus sequence for D1ED III from dengue-1 viruses was obtained by aligning five amino acid sequences from different isolates of the dengue-1 virus [21]. According to the amino acid sequence of D1ED III, the DNA sequence of the D1ED III gene was derived using codon usage of Escherichia coli and was fully synthesized using the assembly PCR method [28]. The product of the assembly PCR was then amplified by conventional PCR. To generate an expressing plasmid for recombinant D1ED III, the following primers were used: forward primer, 5′-GGAATTCCATATGaaaggcatgagctatgtgatgt -3′ (Nde I site, underlined); reverse primer, 5′- CCGCTCGAGgctgctgccttttttaaa -3′ (Xho I site, underlined). The PCR product was cloned into the expression vector pET-22b(+) (Novagen, Madison, WI), using Nde I and Xho I sites to produce the pDen1E3 plasmid. As a result, the C-terminus of the recombinant protein contained an additional hexahistidine tag (HisTag). The Escherichia coli strain BL21 (DE3) (Invitrogen, Carlsbad, CA) was transformed with the expression plasmid pDen1E3 for protein expression. The transformed cells were cultured at 37°C overnight, and protein expression was induced by adding 1 mM isopropylthiogalactoside for 20 hours at 20°C.
The recombinant D1ED III was purified by disrupting the harvested cells in a French Press (Constant Systems, Daventry, UK) at 27 Kpsi in homogenization buffer (20 mM Tris (pH 8.0), 50 mM sucrose, 500 mM NaCl and 10% glycerol). The cell lysate was clarified by centrifugation (80,000× g for 40 min). The majority of recombinant D1ED III was found in the soluble fraction. The recombinant D1ED III was purified using immobilized metal affinity chromatography (IMAC) columns. The eluent from the IMAC column was then polished using an anion exchange column (Q sepharose fast flow; GE) after dialysis against Q buffer (20 mM Tris-Cl, 1 m MEDTA, pH 8.0). E membrane (Pall, USA) was used to remove the endotoxin. The endotoxin levels of the purified recombinant D1ED III were determined by the Limulus amebocyte lysate assay (Associates of Cape Cod, Inc. Cape Cod, MA), and the resulting endotoxin levels were found to be below the detection limit of the kit (<3 EU/mg). The fractions from each step were analyzed by SDS-PAGE gel stained with Coomassie blue (Coomassie Brilliant Blue R-250) and were immunoblotted with anti-HisTag antibodies.
The purified recombinant D1ED III was dialyzed against 5 mM ammonium bicarbonate, pH 8.5. Dialyzed samples were mixed with trypsin (Promega Co., Madison, WI, USA) at a 50∶1 ratio (wt/wt) in 25 mM ammonium bicarbonate, pH 8.5. The reaction allowed to continue for 2 hours and stopped by adding formic acid at a final concentration of 1.2%. The tryptic peptides were analyzed by MALDI-TOF (Burker) mass spectrometry.
Murine CpG was synthesized by Invitrogen Taiwan Ltd and was given as a 10 µg dose dissolved in the sterile water or in the antigenic media. The CpG sequence used was 5′-TCCATGACGTTCCTGACGTT-3′ with all phosphorothioate backbones. Aluminum phosphate suspension was kindly provided by the Taiwan CDC and given as a 300 µg dose in acidic media (pH 6). PELC is a squalene W/O/W nanoemulsion stabilized by Span®85 (sorbitan trioleate, Sigma-Aldrich, Steinheim, Germany) and PEG-b-PLACL, the latter consisting of 75 wt-% of hydrophilic bioabsorbable PEG and 25 wt-% of lipophilic biodegradable PLACL with molecular weight of 7,000 daltons as previously described [24]–[26]. Briefly, 120 mg of PEG-b-PLACL, 0.8 mL of phosphate buffer saline (PBS), and 1.1 mL of oily solution consisting of squalene (Sigma-Aldrich, Steinheim, Germany) and Span®85 (85/15 v/v) were emulsified using Polytron®PT 3100 homogenizer (Kinematica AG, Switzerland) at 6,000 rpm for 5 min. The emulsified PELC formulation was stored at 4°C until use. PELC-formulated vaccine was investigated by re-dispersing 0.2 mL of stock emulsion into 1.8 mL of aqueous solution and mixed with a test-tube rotator (Labinco LD-79, Netherlands) at 5 rpm for at least 1 hour before injection. Recombinant D1ED III and/or CpG were introduced in the aqueous solution, respectively.
Dengue-1 (Hawaii) was used for this study. The virus was propagated in Vero cells, and viral titers were determined by focus-forming assays with BHK-21 cells.
Five BALB/c mice (6–8 weeks of age) were immunized subcutaneously with recombinant D1ED III (10 µg per dose, unless indicated). Mice were given one or two immunizations at a two-week interval with the same regimen. To detect the anamnestic response generated by immunization, immunized mice were inoculated intraperitoneally with 3×106 focus-forming units (FFU) of live dengue-1 virus. Blood was collected from each mouse at different time points, as indicated. Sera were prepared and stored at −20°C until use.
The numbers of IFN-γ- and IL-4-producing cells were determined using mouse IFN-γ and IL-4 ELISPOT kits (eBioscience), respectively. All of the assays were performed according to the manufacturer's instructions. Briefly, 96-well plates with PVDF membranes (Millipore) were coated with capture antibody and incubated at 4°C for 18 hours. The plates were washed twice and blocked with RPMI medium supplemented with fetal bovine serum (10%) for one hour to prevent nonspecific binding in later steps. Splenocytes were seeded at a concentration of 5×105 cells/well and stimulated with D1ED III (10 µg/mL) for 3 days at 37°C in a 5% CO2 humidified incubator. After incubation, the cells were removed from the plates by washing three times with 0.05% (w/v) Tween 20 in PBS. A 100 µL aliquot of biotinylated detection antibody was then added to each well. The plates were incubated at 37°C for 2 hours. The washing steps were repeated as above, and after a 45-minute incubation at room temperature with the avidin-horseradish peroxidase (HRP) complex reagent, the plates were washed three times with 0.05% (w/v) Tween 20 in PBS and then three times with PBS alone. A 100 µL aliquot of 3-amine-9-ethyl carbazole (Sigma-Aldrich) staining solution was added to each well to develop the spots. The reaction was stopped after one hour by placing the plates under tap water. The spots were counted using an ELISPOT reader (Cellular Technology Ltd.). The values presented in the results are mean ± standard deviation of each group.
The levels of anti-D1ED III IgG in the serum samples were determined by titrating the samples. Sera were diluted using 3-fold serial dilutions (starting at 1∶33). Briefly, purified D1ED III was coated on 96-well plates. In some experiments, supernatant obtained from dengue-1 virus infected Vero cells was coated on 96-well plates (2×104 ffu virus/well). Bound IgG was detected with HRP-conjugated goat anti-mouse IgG Fc. After the addition of 3, 3′, 5, 5′-tetramethylbenzidine (TMB), the absorbance was measured with an ELISA reader at 450 nm. For measurement of IgG1 and IgG2a anti-D1ED III subclass, biotin-conjugated rat anti-mouse IgG1 and rat anti-mouse IgG2a were used as detectors, and avidin-HRP was then added. Color was developed as described above. ELISA end-point titers were defined as the serum dilution that gave a 0.5 OD value. The serum dilution was obtained from the titration curve by interpolation, unless the OD value was less than 0.5 at the starting dilution (1∶33).
BHK-21 cells were infected with dengue-1 virus. Three days after infection, the cells were fixed for 15 min in 3.7% formaldehyde/PBS. After washing with PBS, the cells were permeabilized with 0.1% Nonidet P40/PBS for 15 min and blocked with 3% bovine serum albumin (BSA)/PBS for 30 min. Viruses in the infected cells were detected by mouse pre-immune and immune sera (from D1ED III-immunized mice). After washing with PBS, antibody-labeled cells were detected using a secondary antibody conjugated with fluorescein isothiocyanate (FITC). Cellular DNA was labeled by Hoechst stains.
Sera were diluted using 2-fold serial dilutions (starting at 1∶8) and the sera were heat inactivated prior to testing. A monolayer of BHK-21 cells in 24-well plates was inoculated with dengue-1 virus that had been pre-mixed at 4°C overnight with pre-immunization or post-immunization sera to a final volume of 0.5 mL. The virus titer prior to pre-mixing was about 20–40 FFU per well. Viral adsorption was allowed to proceed for 3 hours at 37°C. An overlay medium containing 2% fetal bovine serum and 0.8% methylcellulose in DMEM was added at the conclusion of adsorption. The infected monolayer was incubated at 37°C. After 72 hours of infection, the overlay medium was removed from the wells, and the BHK cells were washed with cold PBS. The cells were fixed for 15 min in 3.7% formaldehyde/PBS. After washing with PBS, the cells were permeabilized with 0.1% Nonidet P40/PBS for 15 min and blocked with 3% bovine serum albumin (BSA)/PBS for 30 min. Infected cells were detected by a monoclonal anti-dengue antibody (American Type Culture Collection, No. HB-114). After washing with PBS, antibody-labeled cells were detected using a secondary antibody conjugated to HRP. The labeling was visualized using TMB. The FFUs were counted, and the neutralizing antibody titer FRNT50 (or FRNT80) was calculated as the reciprocal of the highest dilution that produced a 50% (or 80%) reduction of FFU compared with control samples containing the virus alone. For calculation purpose, the neutralizing antibody titer was designated as 22 when neutralizing antibody titer was less than 23.
To test whether D1ED III blocks dengue virus infection of BHK-21 cells, virus was pre-mixed with different amount D1ED III or control BSA protein as indicated for 10 min at 4°C. Viral adsorption was allowed to proceed for 3 hours at 37°C. The FFUs were determined as described above.
The statistical analysis was conducted using GraphPad Prism version 5.02 (GraphPad Software, Inc.). Data from D1ED III blocking dengue-1 viral infection were processed by a two-tailed Student's t-test. Data from ELISPOT assay were analyzed by the Mann-Whitney test. Data from ELISA and FRNT were performed by the ANOVA Bonferroni post test. Differences with a p value of less than 0.05 were considered statistically significant.
The D1ED III amino acid sequence is a consensus sequence of dengue virus type 1 aligned from selected target sequences (Accession number: P27909, P27913, P17763, P33478 and P27912) [21]. The DNA sequence of the D1ED III gene was derived using codon usage with Escherichia coli and was fully synthesized using the assembly PCR method. The PCR product was cloned into the pET22b vector for recombinant D1ED III expression. The recombinant protein contained an additional HHHHHH sequence (HisTag) at the C-terminus and was expressed under the control of the T7 promoter (Figure 1A). The recombinant D1ED III was purified using an immobilized metal affinity chromatography (IMAC) column and an anion exchange column (Figure 1B, lanes 1–4). Recombinant D1ED III was detected with anti-HisTag antibodies (Figure 1B, lanes 5–8). After the lipopolysaccharide (LPS) was removed (less than 3 EU/mg), purified recombinant D1ED III was comparatively analyzed for its immunogenicity and efficacy in animal models. Recombinant D1ED III was also digested with trypsin to assess its peptide mass fingerprinting. The molecular weight of recombinant D1ED III is 12377. All major peaks in the spectrum at m/z 983.41, 1050.51, 1184.68, 1310.55, 1314.58, 1442.68, 1637.88, and 2559.29 cover over 80% of recombinant D1ED III. The results confirmed that the major peaks in the mass spectra were derived from recombinant D1ED III (Figure 1C).
Dengue envelope protein domain III has been show to be involved in cellular receptor binding [6], [7]. We hypothesized that if purified recombinant D1ED III exists in a suitable conformation, then soluble D1ED III should interfere with dengue viral infections. As shown in Figure 2, the ability of dengue-1 virus to infect BHK-21 cells was inhibited in the presence of D1ED III in a dosage-dependent manner. Greater than 80% reduction of focus number was observed when D1ED III added to cells at a concentration of 0.15 mg/mL. BSA did not inhibit dengue-1 focus formation at concentrations as high as 1.5 mg/mL, which is 10-fold higher than D1ED III. These results suggest that D1ED III can block the cellular binding sites of the dengue-1 virus.
The purified recombinant D1ED III vaccine candidate was formulated with different adjuvants and then tested for its ability to induce both T- and B-cell immune responses in mice. Groups of BALB/c mice were immunized with the different formulations two times with a two-week interval between immunizations. Animals immunized with PBS alone served as controls. One week after the second immunization, splenocytes were harvested and examined for IFN-γ and IL-4 secretion in response to three days of D1ED III stimulation. The ability of the different formulations to induce cytokine secretion varied greatly, as shown in Figure 3. The frequency of IFN-γ spots per 106 splenocytes in mice immunized with recombinant D1ED III alone was 4.9±4.4 (n = 5) spots. Splenocytes from mice immunized with recombinant D1ED III formulated with aluminum phosphate (21.6±17.4, n = 9, p<0.05 compared to D1ED III alone), CpG (7.6±4.3, n = 4, p>0.05 compared to D1ED III alone), and PELC (24.9±14.3, n = 4, p>0.05 compared to D1ED III alone) showed modest increases in IFN-γ spots. Mice immunized with recombinant D1ED III formulated with PELC plus CpG elicited the highest number of IFN-γ spots (85.9±70.3, n = 9, p<0.05 compared to D1ED III alone, alum, or CpG) (Figure 3A). Although the formulation using PELC plus CpG elicited fewer IL-4 spots than the aluminum phosphate formulation (89.7±48.6 vs. 140.7±87.0), the difference was not statistically significant (p = 0.1903). Interestingly, the PELC plus CpG formulation elicited more IL-4 spots than the formulations using PELC (28.3±9.7, p<0.05) or CpG (25.1±10.9, p<0.05) alone (Figure 3B).
Next, we evaluated the IgG antibody responses following three immunizations with two-week intervals between immunizations. Serum samples were collected from the immunized mice at different time points, as indicated in Figure 4A. Formulations of recombinant D1ED III with PELC or PELC plus CpG were highly immunogenic and generated stronger antibody responses than the other formulation groups (p<0.05 by the ANOVA Bonferroni post test). All of the antibody responses peaked between week 4 and week 8 after the first immunization. Importantly, substantial levels of anti-D1ED III IgG antibodies were detectable for over 20 weeks after the initial priming. Sera collected from different formulation groups at week 6 were analyzed for presence of IgG1 (Figure 4B) and IgG2a (Figure 4C). Mice immunized with the CpG formulation generated lower levels of IgG1 antibody than mice that received the aluminum phosphate, PELC, and PELC plus CpG formulations (p<0.05 by the ANOVA Bonferroni post test). The PELC and PELC plus CpG groups had significant levels of IgG1 in comparison with the other groups (p<0.05 by the ANOVA Bonferroni post test). The combination of PELC and CpG induced the highest levels of IgG2a (p<0.05 by the ANOVA Bonferroni post test). As shown in Figure 4D, the ratios of IgG2a to IgG1 in mice receiving PELC plus CpG formulation were higher than those observed in mice receiving the antigen alone, aluminum phosphate, or PELC formulations (p<0.05 by the ANOVA Bonferroni post test). These results indicate that combination of CpG and PELC could improve the IgG2a response.
As the preceding experiments showed that the various formulations of recombinant D1ED III used could elicit considerable antibody responses, we wondered whether these antibodies would recognize the dengue-1 virus. To address this, we employed indirect immunofluorescence staining to evaluate antibody specificity in the sera of mice immunized with the various formulations of recombinant D1ED III. As shown in Figure 5A, naïve serum did not produce immunofluorescent reactivity with dengue-1 virus infected cells. Weak immunofluorescence signals were observed in sera obtained from mice immunized with recombinant D1ED III alone or formulated with CpG. In contrast, strong immunofluorescence signals were observed in sera obtained from mice immunized with the aluminum phosphate, PELC, and PLEC plus CpG formulations. These results suggest that the antibodies induced by recombinant D1ED III can react with dengue-1 virus.
To further examine whether D1ED III raised antibodies can recognize native dengue-1 virus, ELISA was performed by using dengue-1 virus coated 96-well plates. In comparison with naïve serum, sera obtained from mice immunized with D1ED III alone or with various adjuvants produced significantly higher OD values (p<0.05 by the ANOVA Bonferroni post test). These results demonstrated that antibodies induced by D1ED III can recognize native dengue-1 virus (Figure 5B).
The major objective of this study was to explore whether any of the formulations of recombinant D1ED III could induce neutralizing antibodies responses. To evaluate the dengue-1 virus neutralizing ability of the antibodies induced by vaccination, antisera from the mouse immunized with the various formulations were collected, and the neutralizing antibody titers were assessed by focus reduction neutralization tests (FRNT). As shown in Table 1, mice immunized with recombinant D1ED III formulated without adjuvant or with aluminum phosphate, CpG or PELC could not generate significant neutralizing antibody responses (neutralizing antibody titers FRNT50<23). In contrast, mice received PELC plus CpG formulation elicited detectable neutralizing antibody titers (FRNT50 = 24.6).
As the formulation of recombinant D1ED III with PELC plus CpG could elicit the strongest cellular and humoral immune responses and neutralizing antibody responses of all the formulations we tested, we further evaluated the protective efficacy of the D1ED III with PELC plus CpG formulation. Groups of BALB/c mice were immunized with various amounts (3, 10, or 30 µg/dose) of D1ED III three times at two-week intervals. The animals were then injected intraperitoneally with dengue-1 virus (3×106 FFU/mouse) 26 weeks after the first immunization. Serum samples were collected from the mice at the indicated time points, and the neutralizing capacity against dengue-1 virus was examined. As shown in Table 2, no significant neutralizing activity was detected in the sera obtained from naïve mice before and after viral infection (neutralizing antibody titers <23). Mice immunized with 3, 10, or 30 µg of D1ED III produced neutralizing antibody responses at week 6 after initial priming. The neutralizing antibody titers waned at week 24 after the first immunization. At 6 days post-viral challenge, the neutralizing antibody titers FRNT50 were 23.8, 24.4, and 24.2 in mice received vaccination with 3, 10, or 30 µg/dose, respectively. These results provide tangible evidence that an efficient anamnestic neutralizing antibody response was induced in mice immunized with recombinant D1ED III formulated with PELC plus CpG.
Adjuvants containing aluminum (alum) are currently the most widely used adjuvants in human vaccines. However, formulations of dengue subunit vaccines using alum were unable to induce complete protection against dengue virus infection [15], [29]–[32]. In the present study, we prepared D1ED III as a dengue subunit vaccine candidate (Figure 1). The purified D1ED III formed in the proper conformation, which could occupy the cellular binding sites to reduce dengue virus infection (Figure 2). We also found that D1ED III formulated with aluminum phosphate could not induce a significant neutralizing antibody response (Table 1). Altogether, these results suggest that alum may not be suitable for dengue subunit vaccines. In the present study, we evaluated the suitability of PELC-based adjuvants to potentiate the neutralizing antibody capacity of the D1ED III in the mouse model.
All the mice immunized with D1ED III formulated without or with various adjuvants could induce a D1ED III-specific antibody response (Figure 4A). We also noticed that antibodies elicited in all of the immunized groups could recognize dengue-1 virus infected cells (Figure 5A and 5B). However, none of antibodies exhibited a significant neutralizing capacity aside from the antibodies obtained from PELC plus CpG immunized mice (Table 1). In addition, sizeable anamnestic neutralizing antibody responses were observed in mice immunized with D1ED III formulated with PELC plus CpG (Table 2). These results indicate that PELC plus CpG is a promising potential adjuvant for dengue subunit vaccines.
The antibody-dependent enhancement (ADE) of flavivirus infection can be inhibited by complement protein C1q. The inhibition effect is IgG subclass-specific. ADE induced by an IgG2a monoclonal antibody but not by an epitope-matched IgG1 monoclonal antibody could be inhibited by purified C1q [33]. Our results show that different adjuvant formulations can alter the ratio of IgG1/IgG2a (Figure 4D). Interestingly, a significant level of IgG2a was induced in mice immunized with the PELC plus CpG formulation (Figure 4C). These results indicate that D1ED III formulated with PELC plus CpG induces an IgG2a-biased response that may possibly reduce the risk of ADE when sufficient serum C1q levels are present.
Alum predominantly induces a Th2 polarized response [34]–[37] featuring IL-4 production. Consistent with these findings, D1ED III formulated with aluminum phosphate induced higher IL-4 production than any of the other adjuvant formulations that we tested (Figure 3B). Interestingly, the PELC plus CpG formulation induced both vigorous IFN-γ and IL-4 responses (Figure 3). IFN-γ has been shown to play an important role in antiviral activity against dengue virus [38], [39]. The induction of strong IFN-γ responses by the PELC plus CpG formulation will be greatly advantageous in protecting hosts against dengue virus.
There are some of limitations for in vivo protection studies due to the lack of a relevant mouse model of dengue infection. In the present study, we utilized dengue-1 virus as an antigenic challenge model to evaluate memory neutralizing antibody responses. Our results show that the low neutralizing antibody titers were induced and diminished at 24 weeks after immunization of D1ED III formulated with PELC plus CpG (Table 2). Interestingly, quick anamnestic neutralizing antibody responses were evoked when stimulated with dengue-1 virus in mice immunized with D1ED III formulated with PELC plus CpG but not naïve mice (Table 2). These results suggest that memory neutralizing antibody responses were induced in mice immunized with D1ED III formulated with PELC plus CpG.
Our findings show that the PELC plus CpG formulation improves both the intensity and quality of the immune responses against D1ED III. Moreover, the immune responses induced by the PELC plus CpG formulation are beneficial to host protection against dengue viral infection. In conclusion, PELC plus CpG is an attractive adjuvant for dengue subunit vaccines based on recombinant envelope protein domain III. Future work should expand to test the suitability of PELC plus CpG formulations in non-human primate studies.
|
10.1371/journal.ppat.1006705 | Membrane alterations induced by nonstructural proteins of human norovirus | Human noroviruses (huNoV) are the most frequent cause of non-bacterial acute gastroenteritis worldwide, particularly genogroup II genotype 4 (GII.4) variants. The viral nonstructural (NS) proteins encoded by the ORF1 polyprotein induce vesical clusters harboring the viral replication sites. Little is known so far about the ultrastructure of these replication organelles or the contribution of individual NS proteins to their biogenesis. We compared the ultrastructural changes induced by expression of norovirus ORF1 polyproteins with those induced upon infection with murine norovirus (MNV). Characteristic membrane alterations induced by ORF1 expression resembled those found in MNV infected cells, consisting of vesicle accumulations likely built from the endoplasmic reticulum (ER) which included single membrane vesicles (SMVs), double membrane vesicles (DMVs) and multi membrane vesicles (MMVs). In-depth analysis using electron tomography suggested that MMVs originate through the enwrapping of SMVs with tubular structures similar to mechanisms reported for picornaviruses. Expression of GII.4 NS1-2, NS3 and NS4 fused to GFP revealed distinct membrane alterations when analyzed by correlative light and electron microscopy. Expression of NS1-2 induced proliferation of smooth ER membranes forming long tubular structures that were affected by mutations in the active center of the putative NS1-2 hydrolase domain. NS3 was associated with ER membranes around lipid droplets (LDs) and induced the formation of convoluted membranes, which were even more pronounced in case of NS4. Interestingly, NS4 was the only GII.4 protein capable of inducing SMV and DMV formation when expressed individually. Our work provides the first ultrastructural analysis of norovirus GII.4 induced vesicle clusters and suggests that their morphology and biogenesis is most similar to picornaviruses. We further identified NS4 as a key factor in the formation of membrane alterations of huNoV and provide models of the putative membrane topologies of NS1-2, NS3 and NS4 to guide future studies.
| Positive-strand RNA viruses induce membrane alterations harboring the viral replication complexes. In the case of human noroviruses (huNoV), the major cause of acute viral gastroenteritis, these are induced by the ORF1 polyprotein, which is post-translationally processed into the functional nonstructural (NS) proteins. Partly due to the lack of efficient cell culture models, little is known so far about membrane alterations induced by huNoV belonging to the most clinically relevant genogroup II, genotype 4 (GII.4), nor about the function of individual NS proteins in their formation. We therefore expressed ORF1 proteins of GII.4 and individual NS proteins in cells to study their contribution to viral replication complex formation. Expression of ORF1 proteins of GII.4 induced vesicular membrane alterations comparable to those found in infected cells and similar to picornaviruses and hepatitis C virus (HCV). GII.4 NS1-2, NS3 and NS4 are contributing to viral membrane alterations. Our work provides new insights into their function in huNoV induced replication complex formation while identifying NS4 as the most important single determinant. This knowledge might provide novel attractive targets for future therapies inhibiting the formation of the membranous viral replication complex, as exemplified by the efficacy of HCV NS5A inhibitors.
| Human noroviruses (huNoV) are the most frequent causative agent of acute gastroenteritis worldwide, responsible for over 30% of all cases, subsequently resulting in over 200,000 deaths per annum [1]. Still, no vaccine or specific antiviral therapy is available to counteract huNoV infections. Noroviruses are divided into seven different genogroups (GI-GVII) and further subdivided into numerous genotypes [2]. Noroviruses grouped into GI, GII and GIV mainly infect humans but also other species, while GV infects mice. The GII genotype 4 (GII.4) cause the majority of infections with novel outbreak strains emerging every 2–3 years, likely in a response to an immunological pressure of herd immunity [3–5].
Noroviruses belong to the Caliciviridae family and have a positive-sense single-stranded RNA genome of approximately 7.5 kilobases (kb) (reviewed in [6]). The huNoV genome contains three open reading frames (ORFs), where ORF1 encodes the non-structural proteins (NS1-7) involved in viral replication, ORF 2 encodes the capsid protein and ORF3 encodes a small structural protein. Murine noroviruses (MNV) additionally encode an ORF4 from an alternative reading frame located in ORF2, termed virulence factor 1 (VF1), involved in antagonism of the host innate immune response [7]. The 5’ end of the genome contains a short 5 nucleotide untranslated region (UTR) and the 3’end contains a short UTR and poly-A tail (reviewed in [8]). The norovirus genome is covalently linked at the 5’end with the viral protein VPg (also termed NS5). ORF1 is translated from the full-length genomic RNA, whereas ORF2, ORF3, and ORF4 are mainly translated from a VPg linked subgenomic RNA (reviewed in [8]).
ORF1 encodes a large, approximately 200 kDa, polyprotein that is processed by the viral protease NS6, giving rise to 6 mature nonstructural proteins involved in viral replication and several precursor proteins with potentially additional, yet poorly defined functions (reviewed in [8]). The function of the most N-terminal protein (termed NS1-2 or p48) is unclear. huNoV NS1-2 varies in size (approximately 40–48 kDa) and contains an N-terminal disordered region and a C-terminal predicted trans-membrane domain [9]. The central domain further shows homology to the NlpC/p60 superfamily of enzymes, with diverse hydrolase functions [10]. Genogroup I NS1-2 has been shown to localize to the Golgi complex and induce Golgi disassembly, dependent upon the C-terminal hydrophobic region [11]. MNV NS1/2 contains 2 sites cleaved by murine caspase 3 and has been shown to localize to the endoplasmic reticulum (ER) upon transient expression [12,13]. NS3 (also termed NTPase, 2C-like and p41) has been demonstrated to function as an NTPase in vitro for GI [14]. NS3 has also been shown to co-localize with double stranded RNA (dsRNA) during MNV infection [15]. NS4 (also called p20, p22 or 3A-like) function remains unclear, although NS4 has been demonstrated to disrupt ER to Golgi trafficking resulting in Golgi disassembly during norovirus replication [16]. NS4 has also been shown to inhibit actin cytoskeleton remodeling in an epithelial cell line upon transient expression [17]. Upon MNV infection, NS4 was shown to localize to the replication complex [15], and upon transient expression shown to localize to endosomes [13]. NS5 is linked to the 5’ end of the genome and plays an integral role in the initiation of translation through its interaction with eukaryotic initiation factors and likely primes genome and subgenomic RNA synthesis [18]. The viral protease NS6 (also called Pro or 3C-like) is a well characterized cysteine protease and responsible for the cleavage and processing of the viral ORF1 polyprotein [12,19]. Lastly, NS7 (also called RdRp) functions as the RNA dependent RNA polymerase in viral replication and transcription of subgenomic RNAs [20,21]).
Membrane rearrangements play a key role in the establishment of viral replication complexes for positive strand RNA viruses. In principle these membrane alterations can be subdivided into two morphotypes (reviewed in [22,23]). First, the “invagination type” consists of single membrane invaginations of a donor membrane which stay connected to the cytoplasm via a pore and are represented by alphaviruses and flaviviruses. Viral replication takes place inside these vesicles and the viral RNA contributes to their morphology [24], with the exception of Brome mosaic virus 1a protein which generates spherules in absence of RNA replication [25]. Second, the “DMV-type” consists of vesicular and tubular membrane rearrangements wrapped by one (single membrane vesicle, SMV), two (double membrane vesicles, DMVs) and multiple membranes (multi membrane vesicles, MMVs), induced by picornaviruses, coronaviruses and hepatitis C virus (HCV). Most of these structures have no visible connection to the cytoplasm and the functional significance of the different vesicle subtypes as well as the localization of the RNA synthesis machinery is still a matter of debate. However, these structures can typically be induced simply by expression of the replicase proteins in absence of RNA replication, exemplified by picornaviruses [26–29] and HCV [30–32]. Sole expression of individual nonstructural proteins already induces distinct membrane alterations, which are less complex than those derived from the polyprotein. Still, such studies have allowed the identification of those viral proteins contributing to the morphogenesis of viral replication sites and the unraveling of some of their functions [26–28,30,31,33–35]. In the case of HCV, virus induced membrane alterations have been identified as efficient drug targets for silibinin [36], direct acting antivirals like NS5A inhibitors [37] and host targeting drugs like cyclophilin inhibitors [38].
Our understanding of the ultrastructure of huNoV replication organelles is currently scarce, mostly due to the lack of efficient cell culture models [39]. A replicon model has been established in case of GI noroviruses [40], but no ultrastructural analysis is currently available. A plasmid driven GII.3 replicon model allows moderate RNA replication levels, but it remains difficult to dissect the contribution of protein expression and bona fide RNA replication in this system [41]. Recently, tremendous progress has been achieved in cultivating the more clinically relevant GII.4 strains in both B-cells [42] and enteric organoids [43], still neither of these models has yet been proven to allow ultrastructural studies. Therefore, most of our knowledge of norovirus induced membrane alterations has been obtained using the MNV model [44]. Previous studies showed accumulations of vesicles in the cytoplasm of infected macrophages consisting of single and double membrane vesicles, which have not been further characterized [44]. In addition, it has been shown that MNV induced vesicle clusters co-localize with all MNV NS proteins and with viral replication intermediates and that these extensive rearrangements of intracellular membranes are mainly derived from the secretory pathway, including ER, Golgi and endosomal membranes [15]. Furthermore, the MNV replication organelles seem tightly associated with the cytoskeleton, probably mediated by the main capsid protein [45]. Little is known so far about the contribution of individual nonstructural proteins to virus induced replication vesicles, but it is believed that NS1-2 and NS4 are the main drivers in this process due to their membrane association and comparison to picornavirus proteins (reviewed in [8]). In addition, NS3 is associated with membranes and recently has been shown to be associated with highly motile vesicular structures [13,46].
The current study aimed to investigate membrane alterations induced by clinically highly prevalent GII.4 using a transient expression system in Huh7 cell lines. Membrane structures induced by expression of the polyprotein of three important outbreak strains (Den Haag 2006, New Orleans 2009 and Sydney 2012) comprised SMV, DMV and MMV structures. We further observed that SMVs and DMVs were reminiscent of structures found upon MNV infection. The impact of individual GII.4 NS proteins on intracellular membranes was studied by correlative light and electron microscopy, allowing the localization of each protein within the cellular ultrastructural context. GII.4 NS1-2 induced membrane proliferation of the smooth ER, which was strikingly different from MNV NS1/2. NS3 was tightly associated with lipid droplets (LDs) and induced convoluted membranes. However, only NS4 expression was sufficient to induce SMV and DMV formation, much like the ability of HCV NS5A and poliovirus (PV) 3AB to induce DMVs.
We aimed to study the determinants of membrane alterations induced by huNoV with a specific focus on clinically relevant GII.4 outbreak strains. We therefore wanted to exploit expression of ORF1 and of individual NS-proteins to assess the morphology of virus induced membrane alterations in Huh7 cells. We chose Huh7 cells for two reasons: first, Huh7 cells have been shown to support RNA replication of a human GI Norwalk replicon [40] and a plasmid based GII.3 replicon [41], suggesting that they are in principle permissive for huNoV. Second, Huh7 have been used to study membrane alterations for a variety of positive strand RNA viruses, including HCV, hepatitis A virus (HAV), coronaviruses and Dengue virus (reviewed in [22,23]), thereby facilitating the comparison of these structures among different virus groups.
We first aimed to evaluate whether structures induced by ORF1 expression indeed resembled those found in infected cells using MNV as a model, ideally using the same cell type as intended for huNoV. We therefore generated Huh7-CD300lf cells stably expressing the MNV receptor [47,48] and verified that these cells were indeed permissive for MNV infection by demonstrating the presence of NS3 24h after infection (S1A and S1B Fig). In addition, MNV infected Huh7-CD300lf cells produced similar amounts of progeny virus compared to RAW 264.7 cells, albeit with slightly delayed kinetics (S1C Fig). In contrast, Huh7 cells lacking CD300lf failed to amplify the virus inoculum (S1C Fig). Therefore, ectopic expression of CD300lf rendered Huh7 cells fully permissive for MNV infection and supported the entire MNV replication cycle. Ultrastructural analysis of MNV infected Huh7-CD300lf revealed two major phenotypes not observed in uninfected cells (Fig 1): (1) Areas containing vesicles with a variety of shapes, sizes and types (Fig 1A). In addition to previously reported SMVs, more complex structures like DMVs and MMVs were found, often in proximity to lipid droplets (LD). This phenotype likely resembled an early replication phase described previously [44]. (2) A massive rearrangement of the entire endomembrane system consisting of complex structures, often associated with virions, and lacking an organized morphology was observed during what was likely a later stage of the replication cycle [44] (Fig 1B). We found similar phenotypes in RAW 264.7 cells (S2 Fig), which have been used in previous studies to characterize the ultrastructure of the MNV replication organelle [15,44], except that these cells lacked LDs (S2 Fig). We next analyzed whether similar structures were generated by expression of MNV ORF1. MNV ORF1 was expressed in Huh7 T7 cells under transcriptional control of the T7 promoter and translational control of an encephalomyocarditis virus internal ribosomal entry site (EMCV IRES), allowing efficient cytoplasmic expression of proteins of interest in the presence of T7 RNA polymerase (Fig 2A and 2B). For ultrastructural analysis we used chemical fixation (CF) or high-pressure freezing (HPF) (Fig 2C). Regardless of the fixation technique, we identified basically the same types of membrane alterations observed in phenotype 1 of MNV infected cells: vesiculated areas with SMVs, DMVs and MMVs, again in close proximity to LDs, whereas phenotype 2 (complex membrane structures lacking organized morphology) was not found upon expression of ORF1.
We concluded that expression of ORF1 generates membrane alterations comparable to replication organelles found in MNV infection. Therefore, ORF1 expression seemed a valid model to study the morphology of huNoV induced membrane alterations.
We used ORF1 sequences of three GII.4 strains associated with pandemic outbreaks: a Den Haag 2006b variant (DH) [49], a New Orleans 2009 variant (NO) [50] and a Sydney 2012 variant (Syd) [51]. We first tested the expression of the different ORF1 proteins and their processing to assess the integrity of the polyproteins. In a coupled in vitro transcription/translation system (S3A Fig) most of NS-proteins remained buried in precursors, which according to their sizes could represent the entire ORF1 and NS4-NS7 (S3A Fig). Mature cleavage products were only found for NS1-2 and/or NS3, however the size of these proteins was almost identical. This result was in line with previous data studying ORF1 in vitro processing of GII.4 [43,52]. In addition, we assessed polyprotein expression and processing by Western blotting (WB) after transfection of the plasmids encoding the three ORF1 proteins into Huh7 T7 cells. We could detect cleavage products corresponding to NS3, (S3B Fig), NS7 (S3C Fig) and NS1-2, the latter by expressing N-terminal eGFP tagged ORF1 from the NO isolate since we lacked a specific antibody (S3G Fig). No distinct cleavage products were observed for NS4, NS5 and NS6 (S3D–S3F Fig), indicating that they might be retained in relatively stable precursor proteins, as suggested by in vitro translation.
To investigate membrane alterations resulting from huNoV nonstructural proteins, we expressed the complete ORF1 protein of the three GII.4 isolates in Huh7 T7 cells and performed EM analysis (Fig 3, S4 Fig). As for MNV, the expression of ORF1 polyproteins resulted in the formation of complex vesicular structures for all three GII.4 isolates (Fig 3, S4 Fig), similar to those found in MNV-infected cells (Fig 1, S2 Fig) and irrespective of the fixation technique used. The average diameter of DMVs was approximately 100–200 nm (Fig 3B), and resembled structures found upon HCV and picornavirus infection [28,31,53–56]. Most membrane alterations induced by ORF1 expression in Huh7 cells were again found in close association with LDs, similar to MNV. Overall, no consistent differences were found between ORF1 expression and infection regarding the ratio of SMVs, DMVs and MMVs (Fig 3C). However, high variability in vesicles size was observed among cells of the same condition, likely due to differences in protein abundance, time of infection, cell type, etc. Still, SMVs were by far the most abundant vesicle species in all conditions.
In summary, expression of ORF1 of different GII.4 isolates gave rise to a complex set of membrane alterations independent of RNA replication, but similar to structures found in MNV-infected cells. Altogether our data suggested that norovirus replication organelles might belong to the DMV-morphotype, comparable to those observed for enteroviruses and HCV.
We focused our subsequent analyses on one of the three GII.4 strains (NO), since neither polyprotein processing nor the ultrastructural analysis revealed distinct differences upon the expression of ORF1 among the three strains.
To gain deeper insights into the morphology and biogenesis of ORF1-induced membrane structures we further analyzed tomograms of cells expressing ORF1 of the NO strain fixed by high pressure freezing (S1–S5 Movies, Fig 4A and 4B and S5 Fig), allowing a better preservation of the cell membranes. Areas appearing as simple accumulations of SMVs, DMVs and MMVs revealed complex structures in close proximity to ER sheets, including clusters of interwoven vesicles delimited by one or several lipid bilayers (S1 and S2 Movies and Fig 4A and 4B). In addition, we found double membrane vesicles (DMVs) connecting to multivesicular bodies (MVBs) or late endosomes (S1 and S3 Movies and S5A and S5B Fig) and autophagosome-like structures (ALS, S1 and S3 Movies and S5C Fig).
Since clusters of SMVs, DMVs and MMVs most closely resembled the organization of the MNV replication sites in infected cells, we rendered these areas to address their 3D organization (Fig 4A and 4B, S4 and S5 Movies). SMVs (white), DMVs (yellow) and MMVs (blue) all appeared rather vesicular than tubular, were tightly attached to each other and mostly found adjacent to ER cisternae. In some serial slices, the membrane of a DMV was found still in continuation with the ER (Fig 4A, panel 2). MMVs were mainly generated by enwrapping of SMVs with tubular structures, most likely elongated or collapsed SMVs, appearing as multilamellar vesicles in cross-sections (Fig 4B, panel 2). Alternatively MMVs were originated as SMVs or DMVs engulfing pre-existing DMVs (Fig 4B, panel 3, 4). Overall, the morphology and complexity of the membrane alterations, as well as their biogenesis, appeared very reminiscent of later stages of the enterovirus replication sites [53–56].
Altogether, the ET analysis revealed complex interwoven vesicular structures adjacent to the ER, with one or several membrane bilayers, appearing as SMVs, DMVs and MMVs. MMVs were likely generated by enwrapping of SMVs with elongated SMVs, very similar to the mechanism proposed for enteroviruses.
Little is known so far about the contribution of individual norovirus NS proteins to the biogenesis of the viral replication complex. We therefore fused NS1-2, NS3 and NS4, known to be associated with membranes [13], N-terminally with eGFP, to study their propensity to induce membrane alterations by correlative light and electron microscopy (CLEM). We first confirmed by WB the expression of stable fusion proteins and the absence of free eGFP (S6A and S6B Fig).
The N-terminal protein of norovirus ORF1 is thought to induce membrane rearrangements and is considered to be involved in replication complex formation (reviewed in [8]). First, we examined the localization of eGFP-NS1-2 with respect to subcellular markers by immunofluorescence (Fig 5A). Interestingly, eGFP-NS1-2 had a very peculiar filamentous subcellular distribution in most cells (Fig 5A). A minority of cells showed a focal distribution of NS1-2 or an intermediate phenotype (Fig 5A, white and yellow asterisk, respectively). Since the filamentous localization was observed for N-terminally HA-tagged NS1-2, an artifact caused by eGFP fusion can be excluded (Fig 5C). We found a significant co-localization of NS1-2 with a marker of the ER, judged by Pearson correlation values above 0.5 (Fig 5A and 5B). Next, we used CLEM to determine the ultrastructural morphology of NS1-2-GFP positive structures (Fig 5D). Here, regions with strong eGFP fluorescence, indicating high NS1-2 expression, were represented by a network of tubular membrane protrusions. The absence of ribosomes associated with these structures, and the general co-localization of NS1-2 with an ER marker indicated that these membrane protrusions originated from the smooth ER. It is interesting to note that similar tubular structures can be induced by overexpression of ER-shaping proteins such as REEP1 and CLIMP-63 through direct interaction with microtubules [57]. However, we found no indication for a co-localization of NS1-2 induced ER-tubules with microtubules or intermediate filaments (S6C and S6D Fig). In addition we analyzed co-localization of eGFP-NS1-2 and NS3 upon expression of eGFP-ORF1 (S6E and S6F Fig) to assess the impact of the polyprotein on eGFP-NS1-2 localization. Interestingly, we found a variety of phenotypes in various cells, ranging from a focal distribution of both proteins (upper panel) to the filamentous localization of NS1-2 observed upon individual expression (lowest panel). This result indicated a mutual impact of the NS proteins on their subcellular localization, retaining significant co-localization in all cases (S6F Fig), as reported for MNV [15]. Since the distribution of GII.4 NS1-2 was very different from the pattern reported for NS1/2 of MNV [13], we further analyzed eGFP-NS1/2 of MNV by CLEM (S7 Fig). Interestingly and in concordance with literature, NS1/2 of MNV was widely distributed throughout the cell and co-localized with the ER (S7B Fig). However, we found no indications for specific membrane structures induced by MNV NS1/2 (S7C Fig) comparable to those found for GII.4 NO. The MNV NS1/2 signal could be often correlated with membranes surrounding LDs (S7C Fig), which may represent the ring like structures described in a previous study [13].
Taken together, eGFP-NS1-2 of GII.4 induced tubular protrusions of membranes likely derived from smooth ER in a focused, mainly perinuclear area. This is in contrast to MNV NS1/2, which is widely distributed on the ER. This suggests that NS1-2 of GII.4 probably does not directly induce the vesicular membrane rearrangements observed upon ORF1 expression, but rather may contribute to the proliferation of membranes engaged in replication complex formation. Our results further illustrate differences in the subcellular localization and possible functions of NS1-2 proteins from different norovirus genogroups.
We next analyzed the co-localization of eGFP-NS3 of GII.4 NO with different markers of subcellular compartments (Fig 6A). We observed a distinct, rather dot-like staining pattern for NS3 (Fig 6A and 6B), which was similar to the pattern described for MNV NS3 [13]. NS3 significantly co-localized with markers of the Golgi apparatus, rough ER and LDs (Fig 6A and 6C). We also characterized the subcellular localization of NS3 upon expression of ORF1 (S8A–S8C Fig). Again we found some co-localization of NS3 with ER and LDs (S8A and S8C Fig), albeit to a lesser extent as in case of individually expressed NS3. In contrast, very little co-localization with several other markers of membranous intracellular organelles was observed, including Golgi apparatus (S8C Fig), suggesting again a mutual impact of the NS proteins on their subcellular localization. The tight association of GII.4 NS3 with LDs was also validated in CLEM experiments (Fig 6B). Donut like structures of NS3 observed in IF were indeed NS3-studded membrane layers surrounding LDs (area 1). Interestingly, large foci with strong NS3 but weak LD signal were found to represent highly ordered membrane proliferations and were often observed in close proximity to one or more LDs (area 2, 3). Such convoluted membranes were similar to previously described OSER (organized smooth ER) membranes with cubic symmetry (reviewed in [58,59]). However, we did not observe these structures upon ORF1 expression (Fig 3). Overall, our findings indicate that GII.4 NS3 was found on different membrane compartments of the secretory pathway and was closely associated to intracellular lipid storage compartments.
The fluorescence pattern of eGFP-NS3 and -NS4 was quite similar, revealing a dot like pattern with donut like and filled structures (Fig 7A) located mainly in the perinuclear area. Still, the eGFP-NS4 signal tended to accumulate in larger clusters compared to the majority of eGFP-NS3 (Fig 7A and 7B), but also co-localized with markers of ER, Golgi apparatus and LDs (Fig 7A and 7B). CLEM analysis identified several interesting types of membrane alterations in areas with strong eGFP signal, in agreement with the idea that NS4 is a key driver in the formation of the norovirus replication compartment. Donut like structures were found to be membranes tightly associated with LDs (Fig 7B, area 2), similar to those found for NS3 (Fig 6B). In addition, large eGFP-NS4 positive foci found in close proximity to LDs consisted of vesicle clusters composed of DMVs and SMVs (Fig 7B, area 3, 4; 7C, area 5). The size and morphology of NS4 induced DMVs was very similar to those observed upon expression of the polyprotein (Fig 7D compared to Fig 3B), but their abundance was apparently lower. In contrast, SMVs were much more abundant and heterogeneous in size, with a diameter ranging from 50 nm to 300 nm, although more than ~80% had a diameter <100 nm (Fig 7E). These data suggested that NS4 on its own was capable of inducing vesicle accumulations reminiscent of vesicle clusters of the GII.4 replication compartment. Finally, similarly to NS3, highly fluorescent clusters of NS4 within the cells corresponded to regions forming regularly shaped membrane lattices (Fig 7C, areas 6 and 7). However, NS4 expression resulted in membranes aligned predominantly in tubules with hexagonal symmetry (Fig 7C, area 6). This hexagonal symmetry appeared very similar to the arrangement observed upon overexpression of the hydroxy-methylglutaryl (HMG)-CoA reductase [60], although regions with cubic symmetry were also observed (Fig 7C, area 7). Furthermore, NS4 induced crystalline membrane structures were found in proximity to vesicle clusters (Fig 7C, area 1), suggesting that their formation might be concentration dependent and a consequence of very high local concentrations of NS4.
In essence, our results indicated that NS4 was a key factor in the biogenesis of GII.4 induced membrane alterations. Specifically, the sole expression of NS4 was sufficient to induce several different types of membrane structures, including SMVs, DMVs, as well as geometric membrane lattices not found in infected cells.
Our results obtained from individual expression of eGFP-NS1-2, NS3 and NS4 of GII.4 strain NO revealed that they were indeed associated with membranes. Since the functions of all three proteins, in particular NS1-2 and NS4, are widely enigmatic, we next aimed to generate structural models based on secondary structure analysis and homology searches, allowing for the development of hypotheses accessible to experimental validation. There are no close homologs of known structures for these three proteins, but advanced search methods unambiguously detect distant homologs (20% sequence identity) for parts of both NS1-2 and NS3. These together with secondary structure predictions show that the 334-residue NS1-2 can be described as a three-partite protein with an unstructured N-terminus (residues 1–110) followed by a papain-like thiol hydrolase domain (residues ca 120–230) that is related to a family of phospholipases and acyltransferases [61], and finally a hydrophobic domain (residues ca 250–310) with one or possibly two transmembrane helices. We can thus draw two possible topologies for membrane association of NS1-2 (Fig 8A). Interestingly, the catalytic cysteine and histidine of the putative thiol hydrolase are both present in NS1-2 as C205 and H139. The 366-residue NS3 is a distant homolog of AAA ATPases with an extra 50 residues at the N-terminus comprising a hydrophobic helix that could be transmembrane. Again, we have two possibilities for NS3 membrane association. Finally, we could find no homolog of known structure for the 179-residue NS4, but secondary structure predictions show that its approximately 140 N-terminal residues are highly structured and end in an amphipathic helix connecting to a natively unfolded C-terminus (Fig 8A).
Based on these predictions we finally aimed to gain some evidence for the putative hydrolase domain and its importance for membrane protrusions induced by NS1-2 of GII.4. To this end we generated mutants of the highly conserved proposed catalytic residues H139 and C205 in the context of eGFP-NS1-2. Since these residues are invariant in all norovirus genogroups, we used MNV as a surrogate model to study their functional importance for norovirus replication. Interestingly, both NS1-2 mutants affected the abundance of the filamentous phenotype (Fig 8B–8D), arguing for a contribution of the putative hydrolase domain to the generation of tubular ER protrusions. Albeit mutations H139A and C205A both resulted in an increased abundance of the intermediate phenotype, or a focal distribution of eGFP-NS1-2, all variants still remained localized to the ER (Fig 8E). Importantly, both residues were indeed essential for norovirus replication, as mutations at the corresponding positions H150A and C216A of MNV NS1-2 abrogated the production of infectious virus (Fig 8F).
In summary, our analysis of individually expressed NS1-2, NS3 and NS4 suggest that these proteins, in particular NS4, were the main drivers of replication complex formation for GII.4. The structural models proposing their membrane topology will allow more in depth studies of their precise functions. Our data further suggest a role of the predicted hydrolase domain in the membrane shaping activity of NS1-2, but not to the general localization to the ER.
In this study, we analyzed membrane alterations induced by ORF1 expression of clinically relevant GII.4 isolates and by the individually expressed NS-proteins. ORF1 expression generated vesicle accumulations comparable to those observed in MNV infected cells, mainly consisting of SMVs, DMVs and MMVs. Therefore, norovirus-induced membrane alterations can be generated in the absence of active RNA synthesis, and are reminiscent of picornaviruses and HCV [28,31,53–56]. Our data indicate that NS1-2, NS3 and NS4 are the main drivers in the formation of GII.4 replication organelles. However, only NS4 generated SMVs and DMVs similar to those observed upon ORF1 expression and MNV infection. These data provide the first experimental evidence for the hypothesis that NS4 might be the key factor in the morphogenesis of norovirus replication organelles (reviewed in [8]).
Previous studies on the ultrastructure of MNV replication organelles in RAW 264.7 cells reported on the appearance of vesiculated areas consisting of SMVs and DMVs at 12h post infection, progressing to a complete destruction of ER and Golgi at later time points, coinciding with the accumulation of virions [44]. We found in principle the same two phenotypes upon MNV infection of Huh7-CD300lf cells and in RAW 264.7 cells. ORF1 expression induced structures very similar to the phenotype lacking virions, likely corresponding to an early/intermediate infection stage, but did not progress into the complex endomembrane system observed concomitantly to intracellular virions appearance. Progression to the late stages found in infection might require the presence of the structural proteins or resemble cytopathic effects induced by infection, but not by ORF1 expression. Interestingly, we found no obvious difference in the membrane rearrangements induced by ORF1 or MNV compared to any of the three GII.4 isolates included in this study, arguing for similar mechanisms driving these processes in various norovirus genogroups and validating MNV as a useful surrogate model to study particular aspects of general norovirus biology. Still, our data obtained upon GII.4 ORF1 expression will ultimately require validation in a cell culture model with bona fide RNA replication, such as the recently established enteric organoid cultures [43] or infection of B cells [42]. The same holds true for the functional significance of the LD association we found for GII.4 NS3 and NS4. While LDs are clearly not essential for MNV replication, since they are not detectable in RAW 264.7 cells, we found them close to all virus induced membrane alterations in Huh7 cells, both, upon expression of ORF1 and upon MNV infection. Whether this is co-incidence due to high LD abundance in Huh7 cells or whether LDs have a functional significance in GII.4 replication remains to be determined. At this point we have not been able to detect membrane alterations that could unequivocally be assigned to norovirus infection by EM in norovirus infected enteroid cultures (S. Boulant, personal communication), likely due to the yet limited efficiency of this model and the high percentage of infected cells required for a thorough EM analysis. However, further optimization of culture conditions will hopefully allow such studies in the future, as well as the establishment of a reverse genetics model for GII.4 isolates.
Our ultrastructural analysis does not allow drawing firm conclusions on the origin and biogenesis of GII.4 induced SMVs, since we did not observe such vesicles directly connected to cellular membranes. In contrast to PV, we could not find evidence for an interconnected tubular network [55]. Most of our evidence points to membranes of the ER as the origin of SMVs, similar to MNV [15], due to the co-localization of NS1-2 and NS3 with ER membranes when expressed in the context of ORF1. Furthermore, our tomograms reveal a close proximity of the induced vesicles to ER membranes. Finally, individually expressed NS1-2, NS3 and NS4 all at least in part co-localize to and rearrange ER membranes. Therefore, huNoV, like many other viruses, may hijack this cell organelle to generate its replication organelles (reviewed in [62]). Regarding DMVs and MMVs, we were able to identify intermediate structures. Many vesicles appearing as DMVs in a single plane were in fact MMVs in stato nascendi, generated by enwrapping vesicles with collapsed, elongated SMVs or ER cisternae. DMVs might originate as well from collapsing SMVs. This indeed closely resembles the mechanisms demonstrated for picornaviruses [55,56]. Finally, our tomograms also showed a variety of complex structures (e.g. MVBs, ALS) that might be linked to the biogenesis of these vesicles, however these structures are currently difficult to interpret regarding their functional significance. Overall, our data suggest that norovirus induced membrane alterations are very closely related to picornavirus replication organelles with respect to both morphology and biogenesis. The relatively low abundance of DMVs and MMVs further argues for a non-essential role in norovirus infection, again in line with reports on picornaviruses, where replication most likely occurs on single membrane structures, whereas the appearance of DMVs is associated with later stages of infection [55,56].
Expression of the individual nonstructural proteins fused to eGFP and subsequent analysis by CLEM revealed striking phenotypes for NS1-2, NS3 and NS4, which are very poorly understood regarding specific functions in viral RNA replication and topology (Fig 8A). The sequence of NS1-2 is highly divergent between different norovirus genogroups with no specific function assigned thus far. Bioinformatic analysis proposed an unstructured N-terminal region, which was confirmed biochemically [9], a central domain with potential hydrolase function [10] and a C-terminal transmembrane domain [9,63]. Norwalk virus NS1-2 was shown to disrupt the Golgi apparatus [11], whereas MNV NS1-2 mainly localizes to the ER [13]. Our data now show a very peculiar filamentous localization of GII.4 NS1-2 which could be identified as tubular protrusions of the smooth ER when analyzed by CLEM. Similar structures have been observed upon overexpression of cellular ER remodeling proteins such as REEP1 and CLIMP-63 through direct interaction with microtubules [57], which appear to be the driving force in protruding the tubules. In the case of NS1-2, we found no co-localization with microtubules or intermediate filaments, therefore it is currently not clear how the ER-tubules are expanded. Since mutations in the active site of the putative hydrolase domain severely affected the formation of the tubular ER-protrusions, it is tempting to speculate about a role of this predicted enzymatic function in the membrane shaping activity of NS1-2. However, the functional significance of the ability of GII.4 NS1-2 to form filamentous ER tubules still remains unclear, since MNV NS1/2 is devoid of this property. Overall these data are suggestive for the presence of an enzymatic activity residing in NS1-2, but they clearly provide no formal proof, which will require the demonstration of such activity in a purified protein. Interestingly, a recent study identified an interaction of the variable, unstructured N-terminus of MNV NS1/2 with the host proteins VAPA and VAPB, critical for viral replication [64]. VAPA has already been observed to complex with Norwalk NS1-2 [65], presumably via a different, yet to be defined region. More strikingly, VAPA and VAPB have been implicated in the biogenesis of the HCV replication compartment via interaction with NS5A (reviewed in [23]). Therefore, VAPA and VAPB interactions might contribute to the function of NS1-2 in the formation of the norovirus replication organelle, or even be a common host factor of many DMV-type positive strand RNA viruses.
Overall, our data provide a first step to define a role for GII.4 NS1-2 in the biogenesis of the huNoV replication compartment in proliferating membranes of the smooth ER, thereby generating the material that could be transformed into vesicle clusters by other NS-proteins and suggesting a role of the predicted hydrolase function in this process.
Our study demonstrates that NS3 and NS4 both had the ability to create organized ER structures known as convoluted membranes, tubulo-reticular structures, crystalloid ER or cubic membranes, which have been found in a variety of viral infections and upon overexpression of ER membrane shaping proteins (reviewed in [58,59]). Interestingly, similar structures have been found upon expression of HAV 2C and 2BC, with the former regarded as the functional counterpart of norovirus NS3 in picornaviruses [34]. These structures clearly resemble OSER membranes with cubic (NS3) and hexagonal (NS4) symmetry. It has been reported that OSER structures can be generated through weak protein-protein interactions, which can even be triggered by overexpression of GFP tagged to cytochrome b(5), with the GFP moiety providing the homotypic interactions [66]. Therefore, we currently cannot exclude that the fusion of NS3 and NS4 with eGFP, required for CLEM, also contributed to this phenotype. However, it seems unlikely that these structures are essential for viral replication, since they have not been observed so far in either MNV infected cells or upon ORF1 overexpression [15,44]. Therefore, the ability of NS3 and NS4 to induce cubic membranes might point to an intrinsic property, probably weak homotypic interactions, and seems to be linked to strong overexpression of single NS proteins. Upon ORF1 expression and virus infection, such structures might be prevented due to the interactions among NS-proteins and by lower local concentrations of individual NS-proteins. Also in the case of HAV, cubic membranes were only found upon 2C/2BC overexpression [34] but not in HAV infected cells [28].
The strong association of NS3 with ER membranes surrounding LDs as well as its potential to induce convoluted membranes suggests an active function in the formation of GII.4 replication organelles, which clearly requires more detailed studies beyond the scope of this manuscript. A similar localization pattern was previously found for MNV NS3 in Vero cells, but not tested for LD localization at that point [13]. Interestingly, NS1-2 co-localized to NS3 on LDs in most cells upon expression of ORF1, whereas this LD association was never found upon sole expression of GII.4 NS1-2. This result argues for an additional role of NS3 in recruiting NS1-2 upon formation of the viral replication organelle, probably involving direct interactions. The recently observed localization of MNV NS3 with microtubules and cholesterol rich lipids when transiently expressed in Vero cells [45,46] further argues for a function of NS3 in shaping the replication organelle. Whether or not the NTPase activity, which has been demonstrated for the NS3 protein of GI.1 [14], is important for localization and membrane activity of NS3 remains to be determined. However, the most obvious putative function of NS3 in RNA replication is a supposed helicase activity, suggested by conserved SF3 helicase motifs [67]. Still to date, only the related 2C protein from the picornavirus Enterovirus 71 has been shown to function as an ATP dependent helicase [68].
NS4 is by far the most enigmatic among the norovirus NS proteins and little is currently known about its structure or function. Here, we provide the first direct experimental evidence suggesting that NS4 indeed might be the central organizer of the norovirus replication complex. In contrast to HCV, where all NS proteins can induce the formation of vesicular structures [31], NS4 was, in our hands, the only GII.4 protein capable of vesicle formation upon individual expression. In addition, SMVs and DMVs were found, similar to MNV infection and ORF1 expression, albeit with differing abundance. Altogether, these data suggest that NS4 may be the key driver in the formation of norovirus induced replication vesicles, requiring auxiliary functions of NS1-2 and NS3 to finally shape the replication organelles. This is reminiscent of HCV, where the complexity of the so called membranous web is only found upon expression of a polyprotein precursor encoding NS3 to NS5B, arguing for a concerted action engaging several NS-proteins [31,33]. In the case of picornaviruses, 2BC [27], 2BC/3A [26], 2C [34,69] and 3AB [35] have been found to generate distinct membrane alterations upon individual expression, with 2BC/3A and 3AB generating DMVs. However the ultrastructure of the replication organelles is far more complex upon infection also in case of picornaviruses [55]. Still, it is interesting to note that the unrelated proteins NS5A of HCV and 3A of PV are capable of inducing DMVs and share a similar structural organization. Specifically, both contain a unique structured region lacking enzymatic functions, which has been resolved for HCV NS5A [70] and PV 3A [71], an intrinsically unfolded region engaged in recruitment of host factors [72–74] and a membrane attachment region. Our current prediction for the organization of NS4, albeit highly speculative at this point, is strikingly similar regarding subdomain organization and functions (Fig 8A). This model will provide a valuable starting point for further in depth studies regarding the function of norovirus NS4.
Another important aspect that needs to be addressed in future studies is the function of NS4 as part of various stable polyprotein cleavage intermediates. Our study, including three different GII.4 strains suggests a delayed cleavage of the NS4-NS7 precursor, based on in vitro translation. However, a far more detailed analysis of polyprotein cleavage kinetics using a different GII.4 strain suggests a number of cleavage intermediates with NS4, including NS4-NS7, NS4-NS6 and NS4-NS5 [52], which might serve specific functions in the norovirus replication cycle. It is particularly tempting to speculate that delayed cleavage of NS4-NS7 might avoid diffusion of the replicase components, since NS4 seems the only protein associated with membranes in this precursor protein.
Our model of the putative membrane organization of NS1-2, NS3 and NS4 (Fig 8A) is still highly speculative and contains several uncertainties regarding transmembrane topology. While our predictions are in favor of one transmembrane domain for both, the NS1-2 and NS3 proteins, this topology would require a post-cleavage membrane insertion to keep the cleavage site accessible to NS6. Alternatively, and still in line with the predictions, NS1-2 could span the membrane twice and NS3 may harbor an amphipathic alpha helix tethering the protein to membranes, thereby keeping the NS6 cleavage site in the cytoplasm. Regardless, the transmembrane topology of NS1-2 and NS3 can be experimentally addressed in future studies using our expression model and, for example, testing the accessibility of N- and C-termini to proteases in cell lysates.
In summary, our study reveals a first insight into the organization of the putative GII.4 replication organelle and the contribution of individual NS proteins to its biogenesis using a protein expression model. In the case of HCV, comparable expression models have been invaluable to study formation and structure of the viral replication organelles as they allow mechanistic studies using replication deficient mutants or inhibitors interfering with replication. Thereby, the contribution of viral NS-proteins could be clearly defined [31,33] in addition to the importance of host factors like PI4KA [75,76]. Only expression models allowed us to identify viral membrane alterations as targets of direct antiviral agents (NS5A inhibitors [37]) and host targeting drugs including PI4KA inhibitors [75] and cyclophilin inhibitors [38]. Similar approaches might help to identify novel strategies to develop drugs targeting norovirus replication in future studies. Furthermore, our study lays the groundwork for an in-depth analysis of the functions of NS1-2 and NS4 in replication complex formation.
All cell lines were cultured in Dulbecco’s Modified Eagle Medium (DMEM; Life Technologies, Darmstadt, Germany) supplemented with 10% fetal bovine serum, non-essential amino acids (Life Technologies, Darmstadt, Germany), 100U/ml penicillin and 100ng/ml streptomycin (Life Technologies) and cultivated at 37°C and 5% CO2. The human hepatoma cell line Huh7 (maintained in our laboratory), stably expressing T7 RNA polymerase under blasticidine selection (5 μg/ml, Invitrogen, Germany) [74], was used for transient expression of plasmids encoding GII.4 NoV proteins that were analyzed by immunofluorescence and Western blot assays. Huh7 cells expressing the MNV receptor CD300lf were generated by transduction with a lentiviral vector encoding the murine CD300lf cDNA [47] (generous gift from R.C. Orchard and H.W. Virgin). Cells were selected by puromycin to obtain a stable culture of Huh7 cells with CD300lf expression. The murine macrophage cell line RAW 264.7 was obtained from ATCC (Middlesex, UK) and used for infection with MNV. MNV-CW1 [44] was used at a multiplicity of infection of 1 and analyzed 24 h after infection, unless otherwise stated. HEK293T-cells (Birke Bartosch, Lyon) were used for production of MNV virus stocks upon transfection of plasmid pMNV-CW1 (generous gift from H.W. Virgin).
The genomes representing consensus sequences of respective patient isolates of three GII.4 strains including a Den Haag 2006b variant (DH) (GenBank accession no. AB447456), a New Orleans 2009 variant (NO) (GenBank accession no JQ613573) and a Sydney 2012 variant (Syd) (GenBank accession no JX459908) were used in this study. Coding sequences corresponding to ORF1 of the three isolates were synthesized with protein sequences identical to the GenBank entries in vector pBMH by Biomatik (Cambridge, Canada). Full length ORF1 of NO, Saga, Sydney, MNV were amplified by PCR from the pBMH construct. Restriction sites NcoI and PacI were used to insert fragments into a basic pTM1-2, AgeI and PacI was used for insertion into basic pTM 1–3.
To generate pTM vectors allowing expression of N-terminally HA- or eGFP tagged individual norovirus nonstructural proteins, respective coding sequences were amplified using primers given in Table 1 and cloned into pTM-HA or pTM-eGFP, respectively using the indicated restriction sites.
All PCR amplifications for cloning were performed with Phusion Flash High-Fidelity PCR Master Mix according to the manufacturer’s instructions (Thermo Fisher Scientific, Germany). PCR products were separated by agarose gel electrophoresis and purified with NucleoSpin Gel and PCR Clean-up kits purchased from Macherey-Nagel (Germany). Restriction digests were performed according to the instructions of the manufacturer (New England Biolabs). All parts of plasmid sequences amplified by PCR were analyzed by Sanger-sequencing to verify sequence fidelity and the correct reading frame (GATC Biotech, Konstanz, Germany).
MNV stocks were obtained by transfecting plasmid pMNV-CW1 (generous gift from Herbert W. Virgin) into 293T cells, as described [77]. 293T cells were seeded in 10 cm tissue culture dish (Corning, Durham, NC, USA) at a density of 3x105 cells/ml in 15.5 ml of complete DMEM. After 24 hours, cells were transfected with 15 μg of pMNV-CW1 plasmid DNA using TransIT-LT1 Transfection Reagent (Mirrus Bio LLC, Madison, WI, USA) following the instructions of the manufacturer. Transfections were incubated for 48 hours. Virus stocks were obtained by harvesting the cells in their culture medium and twice freezing at -80°C and then thawing at 37°C. Lysates were then centrifuged at 2500 x g for 5 minutes and clarified virus stocks were stored at -80°C. To determine virus titers, RAW 264.7 cells were seeded in 96 well tissue culture plates (Corning, Durham, NC, USA) at a density of 2x104 cells per well. After 24 hours, wells were infected in quadruplicate with serial dilutions of virus stocks diluted in DMEM medium. Assays were harvested 72 hours post infection by aspirating supernatant, washing with PBS and staining with 50 μl of a 1.25% (w/v) crystal violet solution (Merck, Darmstadt, Germany) in a 25% (v/v) ethanol solution for 10 minutes at room temperature. The wells were washed twice with distilled water and scored as positively infected or negative. The TCID50/ml was then calculated using the Spaerman-Kärber method [78,79].
For the transient expression of norovirus eGFP or HA-tagged proteins, Huh7 T7 cells were transfected with LT1 transfection agent (Mirus Bio LLC, Madison, WI, USA) according to the manufacturer's instructions. Cells were processed for IF as described in [76]. Briefly, cells were fixed in 4% paraformaldehyde (PFA) for 20 min and permeabilized with 0.5% Triton X-100 (PBS) for 15 min for co-localization analysis of eGFP-tagged proteins with subcellular markers. Primary antibodies were incubated in 3% bovine serum albumin (BSA) for 1 h at room temperature (RT). NS3 was detected with an in-house created rabbit polyclonal serum, animals were immunized with NS3 expressed and purified from E. coli by Davids Biotechnology, Regensburg, Germany. Antibodies detecting NS4/p20 and NS6 of GII.4 were a generous gift from Stefan Taube [80]. Anti NS5/VPg (strain SAB60) and anti NS7 (NO-strain) sera were raised in rabbits (Eurogentech) based on purified proteins expressed in E. coli. αNS3MNV PAB was kindly provided by Prof. Dr. Ian Goodfellow, Cambridge University, UK, and was obtained by immunization of rabbits with GV MNV and GI sequences [81]. Subcellular compartments and dsRNA were labeled by the following commercially available antibodies: SEC31A/ COPII Vesicles: BD Bioscience / 612351 (Becton Dickinson GmbH, NJ USA); Golgi Apparatus: Anti-Golgin 97 (ab84340, Abcam); autophagosome: p62/ SQSTM1 (M162-3, MBL Life science). Mitochondria were labeled with MitoTracker Deep Red (M22426) and LDs were stained with HCS LipidTox Red Neutral Lipid Stain (Thermo Fisher Scientific Waltham, Massachusetts, USA). ER was stained with a polyclonal anti-PDI antibody (ab31811; Abcam, Cambridge, UK) unless otherwise stated. In case co-staining did not allow the use of this antibody, a mouse monoclonal antibody against Climp-63 (mainly rough ER, ALX-804-604, Enzo) or reticulon 3 (RTN3 sc-374599, Santa Cruz Biotechnology) was used. The lysosome was marked with anti LBPA1 antibody (Clone 6C4, Sigma-Aldrich, Germany). All primary antibodies were utilized in a 1:50 dilution. Alexa 488 or 647-conjugated secondary antibodies (Invitrogen, Molecular Probes) were incubated in 3% BSA for 45 min at RT with a dilution of 1:1000. Nuclei were stained using 4,6-diamidino-2-phenylindole (DAPI) for 1 min at a dilution of 1:4000 after incubation with secondary antibodies. Cells were mounted with Fluoromount G (Southern Biotechnology Associates, Birmingham, AL, USA). Confocal microscopy was conducted on a Leica SP5 AOBS Point Scanning Confocal Microscope (Leica Microsystems). Confocal microscopy was conducted on a Leica SP5 and on Leica SP8 AOBS Point Scanning Confocal Microscopes (Leica Microsystems). Image analysis was performed using the ImageJ software package Fiji (http://fiji.sc/wiki/index.php/Fiji) [82] and the Coloc 2 plugin was used to calculate the Pearson's correlation coefficient. A Pearson's correlation coefficient higher than 0.5 indicates a strong colocalization.
Cells grown on glass coverslips were subjected to chemical fixation and subsequent epon embedding. For overexpression of norovirus proteins, Huh7-Lunet T7 cells were transfected with TransIT-LT1 (Mirus, Madison USA) transfection reagent according to manufacturer’s instructions and fixed 24 hours post transfection. RAW 264.7 (ATCC, UK) were used for infection with MNV-CW1 [83]. For chemical fixation, cells were washed 3 times with 1x PBS and fixed for 30 min with pre-warmed 2.5% glutaraldehyde in 50 mM sodium cacodylate buffer (pH 7.2) containing 1 M KCl, 0.1 M MgCl2, 0.1 M CaCl2 and 2% sucrose. Cells were washed thoroughly 5 times with 50 mM cacodylate buffer and post-fixed on ice in the dark with 2% OsO4 in 50 mM cacodylate buffer for 40 min. Cells were washed with H2O overnight, treated with 0.5% uranyl acetate in H2O for 30 min, rinsed thoroughly with H2O and dehydrated in a graded ethanol series at RT (40%, 50%, 60%, 70% and 80%) for 5 min each and 95% and 100% for 20 min each. Cells were immersed in 100% propilene oxid and immediately embedded in an Araldite-Epon mixture (Araldite 502/Embed 812 Kit; Electron Microscopy Sciences). After polymerization at 60°C for 2 days, coverslips were removed and the embedded cell monolayers were sectioned using a Leica Ultracut UCT microtome and a diamond knife. Sections with a thickness of 70 nm were counter-stained with 3% uranyl acetate in 70% methanol for 5 min and 2% lead citrate in H2O for 2 min, and examined with the transmission electron microscope Philips CM120 TEM (Biotwin, 120 kV).
Huh7-T7 cells seeded on glass-bottom dishes containing a photoetched gridded coverslip (MatTek) were transiently transfected with expression plasmids coding for GFP-fused norovirus proteins. After 24 h cells were washed twice with PBS, fixed for 30 min at room temperature with PBS containing 4% paraformaldehyde and 0.2% glutaraldehyde and stained with DAPI and far-red LipidTOX neutral lipid stain (Thermofisher) according to the manufacturer's instructions. Samples were analyzed on a Nikon TE2000 Ultraview ERS spinning disc (PerkinElmer). Z-stacks of GFP-positive cells were collected and the positions of the cells of interest were recorded using transmitted light with a differential interference contrast configuration. Cells were then fixed in EM fixative and embedded in Epon/araldite resin, as described above.
Seventy nm ultrathin sections were prepared and examined with a Jeol JEM-1400 transmission electron microscope (Jeol Ltd., Tokyo, Japan). Landmark correspondence plugin from Fiji imageJ distribution was used to correlate the light microscopy and the electron microscopy datasets. Briefly, a single optical section displaying a LDs distribution that matched the one observed in the electron microscopy micrograph was extracted from the z-stack and the corresponding LDs on the two images were used as landmark to calculate the transformation.
Cells were seeded onto 3 mm sapphire discs (M. Wohlwend GmbH, Sennwald, Switzerland) that had been carbon coated to improve cell adhesion. One day after transfection or infection cells (Huh7-Lunet T7 and RAW 264.7, respectively) were frozen after immersion in 1-hexadecene (Merck, Hohenbrunn, Germany) using a high-pressure freezer (M. Wohlwend GmbH). Frozen discs were stored in liquid nitrogen until further processing. Freeze substitution was done in acetone containing 0.2% (w/v) OsO4, 0.1% (w/v) UA, and 5% (v/v) water by slowly warming the samples from −90°C to 0°C during a period of 20 h [84]. Samples were kept at 0°C and at room temperature for 30 min each, washed with acetone, and embedded in four-step epon series (Fluka, Buchs, Switzerland) using 1 h-incubation in 25%, 50% and 75% epon dissolved in acetone and overnight incubation in 100% epon. Epon was exchanged, polymerized for 3 d at 60°C and sapphire discs were removed by immersion in liquid nitrogen. Seventy or 250 nm thick sections for were examined by conventional transmission EM or electron tomography, respectively.
Sections of 250 nm thickness were collected on palladium-copper slot grids (Science Services, Munich, Germany) coated with Formvar (Plano, Wetzlar, Germany). Protein A-gold (10 nm) was added to both sides of the sections as fiducial markers. Single axis tilt series were acquired with a FEI TECNAI F30 microscope operated at 300 kV and equipped with a 4k FEI Eagle camera over a −65° to 65° tilt range (increment 1°) and at an average defocus of −0.2 μm. Reconstruction of the tomograms and rendering of their 3D surface was performed by using the IMOD software package (version 4.9)[85] (bio3d.colorado.edu/imod).
For Western blotting, the cells in a 6-well plate well were lysed and denatured in 150 μL of 6× Laemmli buffer by heating to 95°C for 5 minutes, and loaded onto an 12% polyacrylamide-SDS gel. After resolution by SDS-PAGE, proteins were transferred to a polyvinylidene difluoride (PVDF) membrane, with the exception of NS7, which was transferred to a nitrocellulose membrane (Amersham Protran 0.45 NC, GE Healthcare Life Science). NS proteins were detected using NS-protein specific polyclonal rabbit antibodies described in IF section in a 1:1000 dilution. β-actin was detected by monoclonal mouse antibody (A5441), Sigma-Aldrich. Primary antibodies were detected using αrabbit/ αmouse horseradish peroxidase (HRP)-coupled secondary antibodies (Sigma-Aldrich) and imaging was done with the ChemoCam 6.0 ECL system (INTAS Science Imaging, Goettingen, Germany).
pTM based constructs were phenol/chloroform purified and reconstituted in RNase-free double-distilled H2O to a concentration of 1 μg/μL. 0.5 μg of the plasmid preparation were then mixed with 10 μL of the L1170 T7 TNT-kit (Promega, Madison, USA) and 1μL of 35S Methionine 10 mCi/ml. The reaction mixture was then incubated at 30°C for 90 min. Afterwards, the reaction was suspended in 2x laemmli buffer, and denatured at 95°C for 5 min before being loaded onto a 12% SDS gel for electrophoresis. Radiolabeled proteins were visualized by autoradiography using a phospho-imager (BioRad, Munich, Germany).
Stable cell lines expressing the MNV receptor CD300lf [47,48], were created by lentiviral transduction. The lentiviral vectors were created by co-transfecting 293T cells with a gag-pol plasmid (pCMV∆8.31), the retroviral vector containing the CD300lf sequence (gift from Dr. Herbert Virgin, Washington University at St. Louis, St. Louis, MO, USA), and an envelope plasmid (pMD.G) mixed in a 3:3:1 ratio, respectively, as described elsewhere [86]. Huh7 cells were seeded at a density of 1x105 cells per well in 6-well plates and transduced with 1 ml of lentivirus suspension mixed with 1 ml of fresh DMEM medium for 12 hours. Media was aspirated and replaced with 1:1 lentivirus suspension and 1 ml of fresh DMEM two more times at 12 hour intervals. After 36 hours, the media was changed and 2 ml of fresh DMEM containing 2 μg/ml puromycin selection was added and selective pressure was maintained during passaging.
We used resources of the MPI bioinformatics Toolkit [87] to define domains in NS1-2, NS3 and NS4. HHPRED was used to find domains with homologs of known structure and Quick2D to identify structured and non-structured regions and putative transmembrane helices [88,89]. Transmembrane helices were further sought with Polyphobius. For putative membrane-peripheral and transmembrane helices, ideal alpha-helices were generated with the “fab” function of PyMOL, as were intervening loops. The NS1-2 central domain (residues 119 to 213) was homology modelled with SwissModel [90] from Protein Data Bank entry 4DPZ [61]. The non-structured N-terminus and helical C-terminus were then generated with PyMOL 1.8.2.0 [91]. The three parts were assembled in PyMOL with the PyMOL sculpting function. NS3 was modelled with I-Tasser [92] and NS4 was modelled with RaptorX [93], with default parameters. A membrane model was generated with the Charmm-GUI webserver [94] with a lipid composition similar to the endoplasmic reticulum (Phosphatidylcholine 60%; Phosphatidylethanolamine 25%; Phosphatidylinositol 15%)[95]. The positions of the transmembrane/peripheral helices relative to this membrane were adjusted with PyMOL. The structures were then minimized with secondary structure restraints with the Phenix geometry minimization function [96].
General statistical analyses as indicated in the corresponding figures were performed using Graphpad Prism Software.
|
10.1371/journal.ppat.1006808 | Biogenesis of the mitochondrial DNA inheritance machinery in the mitochondrial outer membrane of Trypanosoma brucei | Mitochondria cannot form de novo but require mechanisms that mediate their inheritance to daughter cells. The parasitic protozoan Trypanosoma brucei has a single mitochondrion with a single-unit genome that is physically connected across the two mitochondrial membranes with the basal body of the flagellum. This connection, termed the tripartite attachment complex (TAC), is essential for the segregation of the replicated mitochondrial genomes prior to cytokinesis. Here we identify a protein complex consisting of three integral mitochondrial outer membrane proteins—TAC60, TAC42 and TAC40—which are essential subunits of the TAC. TAC60 contains separable mitochondrial import and TAC-sorting signals and its biogenesis depends on the main outer membrane protein translocase. TAC40 is a member of the mitochondrial porin family, whereas TAC42 represents a novel class of mitochondrial outer membrane β-barrel proteins. Consequently TAC40 and TAC42 contain C-terminal β-signals. Thus in trypanosomes the highly conserved β-barrel protein assembly machinery plays a major role in the biogenesis of its unique mitochondrial genome segregation system.
| Trypanosoma brucei and its relatives are important human and animal pathogens. Unlike most other eukaryotes trypanosomes have a single mitochondrion with a single unit mitochondrial genome, termed the kinetoplast DNA (kDNA). During each cell cycle the kDNA is replicated and subsequently segregated into the two organelles that are formed during binary fission of the mitochondrion. Segregation depends on the tripartite attachment complex (TAC) which physically links the kDNA to the basal body of the flagellum. Thus, the TAC couples the segregation of the replicated kDNA to the segregation of the old and new flagella. We have characterized the outer membrane section of the TAC and shown that it contains a complex of three integral membrane proteins, TAC60, TAC42 and TAC40, each of which is essential for TAC function. Furthermore, we have identified which protein import systems are required for their biogenesis. In the case of TAC60 we demonstrate that membrane insertion and sorting to the TAC are separate processes requiring distinct cis-elements. Finally, we show that TAC42 is a novel mitochondrial beta-barrel protein whose biogenesis depends on the beta-signal in its C-terminus. Thus, TAC60, TAC42 and TAC40 are essential trypanosomatid-specific proteins that may be exploited as drug targets.
| Mitochondria are a hallmark of eukaryotic cells [1]. They derive from an endosymbiotic event between an archaeal host cell and an α-proteobacterium. The bacterial symbiont was subsequently converted into an organelle. Continued evolution since the origin of the mitochondrion, approximately 1.5–2 billion years ago, has led to a great diversification of the organelle [2, 3]. This is illustrated by the immense variation of the morphology and the behaviour of mitochondria in different species and the large variation in the organization and coding content of their genomes. However, faithful transmission of mitochondria and their genomes to their daughter cells is a problem essentially all eukaryotic cells need to solve [4, 5].
In contrast to most other eukaryotes trypanosomes and their relatives have a single mitochondrion that only contains a single unit mitochondrial genome, termed kinetoplast DNA (kDNA). The kDNA consists of two genetic elements the maxi- and the minicircles. The maxicircles are present in 25–50 copies and are 22 kb in length [6]. They contain a number of protein-coding genes expected to be present in the mitochondrial genome. Most of them are cryptogenes whose primary transcripts have to be edited by multiple uridine insertions and or deletions to become functional mRNAs. The minicircles are heterogenous in sequence, occur in several thousand copies and encode the guide RNAs that provide the information for RNA editing [7, 8]. Maxi- and minicircle are highly topologically interlocked and build a large disc-shaped network that is physically linked to the basal body of the single flagellum via a structure termed tripartite attachment complex (TAC). The three zones of the TAC include the unilateral filaments that connect the kDNA network to the inside of the inner membrane (IM), a segment containing tightly apposed detergent-resistant differentiated IM and outer membranes (OM) and finally the exclusion zone filaments that build a bridge from the OM to the basal body [9–11].
Due to the single unit nature of the kDNA network, its replication and segregation need to be tightly coordinated. kDNA replication occurs at a precise stage of the cell cycle immediately before the onset of the S phase in the nucleus [12, 13]. During replication the kDNA doubles in size forming a dumbbell-shaped network. The process has been studied in detail and involves numerous protein factors. The segregation of the replicated kDNA networks depends on an intact TAC. Thus during kDNA replication a new TAC forms that connects the basal body of the new flagella with the replicating kDNA network. Division of the replicated kDNA network finally is linked to the segregation of the old and the new basal bodies. Initially the two kDNA networks remain connected by a structure termed "nabelschnur" which becomes resolved when the distance between the two networks is larger than 1 μm. The segregation then continues and the basal bodies with the attached kDNA networks move further apart [14, 15]. Subsequently, prior to cytokinesis, the mitochondrion is divided in two, the plane of division intersecting between the two kDNA networks [16].
Presently only a few components of the TAC have been identified. The first one was p166 [17]: it contains a single predicted transmembrane domain (TMD) which however is not essential for its localization making it difficult to decide whether p166 is a TAC protein of the IM or the unilateral filaments. TAC102 is a component of the unlilateral filaments [18], whereas p196 [19] and TAC65 [20] were shown to be extramitochondrial TAC subunits that localize to the exclusion zone filaments. The same was the case for an as yet unkown antigen recognized by the monoclonal antibody Mab 22 [21]. Furthermore, TAC40 and peripheral archaic translocase of the OM 36 (pATOM36), two OM proteins that localize to the TAC, have also been described [20, 22]. TAC40 is a β-barrel protein of the mitochondrial porin family. pATOM36 is unusual as it is localized in the TAC region but also present all over the OM. The dual localization of pATOM36 reflects its dual function in kDNA inheritance and in the biogenesis of a subset of α-helically anchored OM proteins including most subunits of the archaic translocase of the OM (ATOM) [20, 23]. Complementation experiments revealed that the two functions are distinct, since the C-terminus is only essential for the biogenesis of OM proteins but not for the segregation of the kDNA. Thus, pATOM36 is an integrator of mitochondrial protein import and mitochondrial genome inheritance [20, 24].
Unlike other OM proteins, the OM subunits of the TAC not only need to be targeted to mitochondria and inserted into OM, but they also have to be sorted to the region of the OM close to the single unit kDNA, where a new TAC is being formed.
Here we have discovered two novel subunits of the TAC that are integral mitochondrial OM proteins. Moreover we provide a detailed analysis of the biogenesis pathways of the two newly discovered TAC subunits as well as of the previously characterized TAC40. We show that one of the new subunits contains separable targeting signals for import into the organelle and for sorting to the TAC, whereas the other one defines a novel class of mitochondrial β-barrel proteins.
Recently we have shown that the mitochondrial β-barrel protein TAC40 exclusively localizes to the TAC and is essential for its function [22]. In order to identify further TAC subunits we performed SILAC-based immunoprecipitation (IP) experiments using 29–13 T. brucei cells and a cell line expressing in situ HA-tagged TAC40. The two cell lines were grown in the presence of isotopically-labeled heavy or light lysine and arginine. Subsequently, identical cell numbers from both populations were mixed and whole cell lysates were prepared which were subjected to IP using anti-HA antibodies. The resulting eluates were analyzed by quantitative MS and Fig 1A shows that they contain only two proteins that were five-fold or more enriched. One was TAC40, whose tagged variant served as a bait, the other one a 60 kDa protein, termed TAC60 (Tb927.7.1400). Subsequently a cell line allowing tetracycline-inducible expression of a Myc-tagged version of TAC60 was used to do a second set of SILAC-IPs. The most highly enriched proteins recovered in these IPs were TAC60, TAC40 and a protein of 42 kDa, termed TAC42 (Fig 1B). Finally we produced a cell line expressing an HA-tagged version of TAC42 and performed a third set of SILAC-IPs which besides TAC42 itself recovered TAC40 and TAC60 as the most enriched proteins (Fig 1C). In summary these reciprocal IPs show that TAC40, TAC60 and TAC42 form a protein complex.
To find out where TAC60 is localized the cell line expressing Myc-tagged TAC60 was analyzed by immunofluorescence (IF). The cells were stained with an anti Myc-antibody, with the DNA-staining agent DAPI, which labels both the nuclei and the kDNA networks, and with the monoclonal antibody YL1/2. The latter recognizes tyrosinated α-tubulin and in T. brucei stains the basal bodies as well as the distal part of the subpellicular array of microtubules [25]. An overlay of all three signals reveals that TAC60 localizes to a dot-like structure between the basal bodies and the kDNA networks (Fig 2A, left panel), as would be expected for a TAC subunit. The TAC is to a large part detergent-resistant, which allows the isolation of a fraction containing flagella whose basal bodies are still connected to the kDNA [9]. IF analysis shows that in such fractions Myc-tagged TAC60 localizes between the kDNA and the flagellum (Fig 2A, right panel) indicating that the protein is a structural subunit of the TAC.
In order to investigate the function of TAC60 a tetracycline-inducible RNAi cell line was produced. Analysis of DAPI-stained cells in the left panel of Fig 2B shows that ablation of TAC60 in insect stage T. brucei leads to a rapid loss of the kDNA networks reaching approximately 50% after 1.5 days of induction. Growth on the other hand is affected after 3 days only (Fig 2B). The right panel in Fig 2B demonstrates that a large majority of the 30% of cells, that have retained their kDNA after two days of RNAi induction, have greatly enlarged kDNA networks. Moreover, in a minority of cells smaller kDNA networks are observed. The massive over-replication of kDNA networks is a hallmark that distinguishes cells with a deficient TAC from cells ablated in kDNA replication [17, 18, 20, 22]. As expected a transmission electron microscopy analysis shows enlarged kDNA networks, which are in part stacked on top of each other but whose ultrastructure is undisturbed (Fig 2D). Mitochondrial translation and thus the kDNA as well as the TAC are essential in both the insect and bloodstream form of T. brucei [26]. In line with this ablation of TAC60 in bloodstream form cells leads to a rapid loss of the kDNA with a subsequent growth arrest (Fig 2C, left panel). However, if the same experiment is done in a bloodstream form cell line of T. brucei that, due to a single point mutation in the nuclear-encoded γ-subunit of the mitochondrial ATPase, can grow in the absence of the kDNA [27], a different result is obtained: in such a cell line TAC function is dispensable and ablation of TAC60, while causing the loss of the kDNA, does not slow down growth (Fig 2C, right panel).
The set of experiments done for TAC60 (Fig 2) were also used to analyze TAC42 (Fig 3). The obtained results were essentially identical for both proteins. In summary, these experiments (Fig 2 and Fig 3) establish that TAC60 and TAC42 are essential novel subunits of the TAC that are not involved in any other essential functions unrelated to the kDNA.
As expected tagged TAC60 co-fractionates with the mitochondrial marker ATOM40 when cells are extracted with a low concentration of digitonin (Fig 4A, upper panel). Moreover, TAC60 is exclusively recovered in the pellet when a crude mitochondrial fraction is subjected to carbonate extraction at high pH indicating that TAC60 is an integral membrane protein (Fig 4A, middle panel).
It has previously been shown that organellar proteins that accumulate in the cytosol upon inhibition of mitochondrial protein import are rapidly degraded by the cytosolic proteasome [28, 29]. Thus to analyze whether TAC60 localizes to the outer or the inner mitochondrial membrane we followed the fate of a tagged version of the protein in inducible ATOM40- and TbTim17-RNAi cell lines. ATOM40 and TbTim17 are core subunits of the archaic protein translocase of the OM (ATOM) and the protein translocase of the IM (TIM), respectively [23, 30–34]. The top panel of Fig 4B shows that the steady state levels of tagged TAC60 in whole cells rapidly decrease during ATOM40 RNAi. The same is the case for a previously characterized β-barrel protein, the voltage dependent anion channel (VDAC), which first needs to be translocated into the intermembrane space (IMS) before it is inserted into the OM. In the TbTim17 RNAi cell line in contrast the steady state levels of TAC60 remain essentially constant. The IM protein cytochrome oxidase subunit IV (CoxIV), whose import requires TbTim17 and therefore serves as a positive control, however accumulates as unprocessed precursor protein (Fig 4B, bottom panel). Thus mitochondrial import of tagged TAC60 depends on ATOM40 but not on TbTim17 indicating that it is an OM protein. This is further supported by a normalized abundance profile of TAC60 over six subcellular fractions, produced in a previous proteomic analysis (Fig 4A, bottom panel) (However since TAC60 was detected in only one of two experiments, it was not included in the OM proteome defined in the study) [35].
In silico analysis of TAC60 of T. brucei using various prediction programs detects two high confidence TMDs (121–141 and 238–258) that are found in essentially all TAC60 orthologues of trypanosomatids. Moreover, HHPred analysis [36] indicates that the C-terminal 150 aa of TAC60 and its orthologues has some similarity to bacterial tRNA/rRNA methyltransferases.
In order to analyze the topology of TAC60 experimentally we used the split GFP approach [37]. To that end a cell line expressing N-terminally HA-tagged GFP lacking the C-terminal β-strand (HA-GFP1-10) was produced. As expected the truncated GFP fractionates with the cytosol (Fig 4C, top right panel). Subsequently, two variants of TAC60 that were N- or C-terminally fused to the last β-strand of GFP (GFP11-TAC60, TAC60-GFP11) were expressed in the same cell line. The IF analysis in Fig 4C (bottom panel) shows a GFP signal that to a large part is in close proximity of the kDNA and thus is consistent with a TAC localization. The same signal is not seen in the absence of tetracycline, which prevents the expression of the fusion proteins. A weak background signal close to but not overlapping with the TAC is visible for both TAC60-GFP11 and GFP11-TAC60 in the absence and in the presence of tetracycline. S1A Fig shows that the signal is due to autofluorescence. In summary, these results show that both the N- and the C-termini of TAC60 face the cytosol and thus are consistent with the notion that TAC60 has two TMDs.
The integral membrane subunits of the TAC not only need to be imported into mitochondria but also require sorting to the single unit TAC. In order to test whether mitochondrial targeting and subsequent sorting to the TAC require distinct signals, we expressed C-terminal tagged versions of TAC60 that were truncated either at their N- or C-termini or on both ends (Fig 5A). IF analysis showed four distinct localizations of the truncated TAC60 versions (Fig 5B, 5C, 5D and 5E, S1B Fig.):
- The variants lacking the C-terminal 153 and 283 aa (ΔC153/ΔC283) were localized to the TAC. Although some dots that do not overlap with the kDNA are also seen (Fig 5B). The same was the case if the C-terminal 283 amino acid deletion was combined with N-terminal truncations of 75 and 97 aa (ΔN75_ΔC283/ΔN97_ΔC283) (Fig 5B). Interestingly, however expression of these two variants causes some cells to show an enlarged kDNA or kDNA loss, respectively.
- The variant lacking the N-terminal 114 aa (ΔN114) was also localized to the TAC (Fig 5C), although in contrast to the variants described above (Fig 5B) the rest of the mitochondrion was also stained. Thus, only a fraction of the tagged ΔN114 variant is localized at the TAC indicating that the efficiency of TAC sorting is reduced.
- The variant lacking the N-terminal 140 aa (ΔN140) was mitochondrially localized (Fig 5D), but not sorted to the TAC demonstrating that mitochondrial targeting and TAC-sorting are distinct events.
- The variants lacking either N-terminal 233 and 257 aa (ΔN233/ΔN257) or the C-terminal 320 and 408 aa (Δ320/ΔC408) finally showed a diffuse cytosolic localization (Fig 5E).
In summary, these results define a 140 aa long segment of TAC60—aa 140–283—encompassing the IMS-exposed loop of the protein and the more C-terminal TMD as essential for mitochondrial targeting. Sorting to the TAC however requires an additional N-terminal 26 aa segment that consists essentially of the first TMD of TAC60. However, on its own the segment only confers an incomplete TAC localization which is illustrated by the fact that a fraction of the ΔN114 variant yields an overall mitochondrial staining.
The immunoblot in S2A Fig confirms that all TAC60 variants are expressed in comparable amounts. However, in many cases the anti-tag antiserum detects additional signals below or above the predicted bands. The ΔN114 and ΔN140 variants show the highest heterogeneity and besides the correctly sized protein at least four major degradation products are detected as well. These degradation products are all mitochondrially localized (S2B Fig) suggesting that TAC subunits that are imported into mitochondria but not sorted to the TAC are degraded. For many other variants closely spaced double bands or additional signals above the correctly sized protein are observed. Preliminary experiments indicate that at least in the case of full length TAC60 and the ΔC153 variant protein phosphatase treatment results in a more simplified pattern shifted towards a lower molecular weight range (S2C Fig). Thus, the observed heterogeneity within the TAC60 variants might be caused by phosphorylation and possibly other postranslational modifications.
To investigate whether the correctly localized TAC60 truncations ΔC153, ΔC283, ΔN75_ΔC283, and ΔN97_ΔC283 are functional we expressed them in a TAC60-RNAi cell line that targets part of the ORF that is absent in the truncations and therefore allows complementation experiments.
The results in Fig 6 demonstrate that both C-terminally truncated variants complemented growth to wild-type level when expressed in the corresponding RNAi cell line. Thus the C-terminal domain of TAC60 that shows similarity to bacterial tRNA/rRNA methyltransferases is dispensable for TAC function. The same experiments were also done for the ΔN75_ΔC283 and ΔN97_ΔC283 TAC60 variants. While both of them localize to the TAC (Fig 5B) they were not able to complement the growth phenotype indicating that their function is impaired (Fig 6). Thus, while the N-terminal 97 aa are dispensable for TAC60 targeting they are required for the function of the protein.
Bioinformatic analysis of TAC42 does not detect any significant sequence similarity to proteins outside the kinetoplastids. As TAC60, tagged TAC42 co-fractionates with the mitochondrial marker ATOM40 when cells are extracted with low concentration of digitonin (Fig 7A, upper panel). Moreover, the normalized abundance profile of TAC42 produced in a previous proteomic analysis suggest an OM localization [35] (as in the case of TAC60, TAC42 was only detected in one of two experiments in this study and was therefore not included in the OM proteome) (Fig 7A, bottom panel). TAC42 is exclusively recovered in the pellet fraction in a carbonate extraction suggesting it is an integral membrane protein (Fig 7A, middle panel). This is surprising since TAC42 lacks predicted TMDs and in silico analyses do not predict it to be a β-barrel membrane protein. Thus, to test whether the biogenesis of TAC42 depends on the β-barrel insertion machinery we expressed a C-terminally tagged version of TAC42 in a cell line allowing inducible ablation of Sam50, the core subunit of the sorting and assembly machinery (SAM). Interestingly, in this cell line the level of a tagged version of TAC42 decreased during RNAi (Fig 7B). As expected the same was the case for the well characterized β-barrel proteins VDAC [38] and ATOM40 [28], which serve as positive controls. The C-terminally anchored OM protein ATOM69 [23] and cytosolic translation elongation factor 1a (EF1a) on the other hand were not affected.
A pioneering study in yeast identified a moderately conserved sequence in the last β-strand of mitochondrial β-barrel proteins that serves as a sorting signal which directs the protein to the β-barrel protein insertion machinery [39]. Fig 8A shows that the three trypanosomal β-barrel proteins TAC40, ATOM40 and VDAC as well as TAC42 have C-termini corresponding to the β-signal consensus sequence. Thus, in order to test whether these sequences function as sorting signals we expressed tagged TAC42 and TAC40 variants containing mutated variants of the putative β-signals. In one variant, termed 1mut, the invariant glycine was mutated to alanine. In the other variant, termed 4mut, all four conserved positions were changed, the first two to alanines and the last two to serines.
The digitonin extraction in the top panel of Fig 8B shows that approximately 45% of the tagged wildtype version of TAC42 is recovered in the pellet corresponding to a crude mitochondrial fraction. The fact that 55% of the tagged proteins remains in the supernatant is likely due to overexpression when compared to the endogenous protein. Interestingly, however in the case of the 1mut and 4mut versions of TAC42 only 8–11% of the proteins are recovered in the pellet fractions. The same experiments were also done with the previously characterized β-barrel protein TAC40 [22] and, as in the case of TAC42, mitochondrial targeting of the 1mut and 4mut versions of TAC40 was dramatically impaired (Fig 8B). TAC40 was also ectopically tagged but likely less overexpressed compared to TAC42 which may explain why almost all of the tagged wildtype variant of the protein is mitochondrially localized. Thus, mutating the conserved glycine or all conserved amino acids in the β-signal consensus sequence of TAC42 and TAC40 progressively reduces mitochondrial targeting of the two proteins.
The β-signal mediates the interaction with the β-barrel insertion machinery [39]. Its absence is therefore expected to interfere with membrane insertion after the proteins have been translocated into the IMS. In order to test this prediction we analyzed the mitochondria-associated fractions of the tagged TAC42 and TAC40 and its corresponding 1mut and 4mut versions by carbonate extraction at high pH to determine whether the proteins have been inserted into the OM. The lower panels in Fig 8B show that the tagged wildtype TAC42 and TAC40 are recovered in the pellet fractions together with ATOM40, which serves as marker for a correctly inserted integral membrane protein. However, membrane insertion of the 1mut variant of TAC42 is reduced by 50% and in the case of 4mut variants by 70–80% for both proteins. Thus, TAC42 and TAC40 contain a C-terminal β-signal that is essential for correct targeting and membrane insertion of the proteins into the mitochondrial OM.
The β-signal is expected to be conserved in all eukaryotes [39]. In order to show direct interaction of the putative β-barrel protein TAC42 with the SAM complex we therefore performed in vitro import experiments using trypanosomal substrate proteins and isolated yeast mitochondria. Radioactive substrates, produced by in vitro translation using rabbit reticulocyte lysate, were incubated with isolated mitochondria from either wildtype S. cerevisiae or from a yeast strain carrying a deletion of the SAM complex subunit Sam37, whose function is to promote β-barrel protein insertion into the OM by linking the SAM and the TOM complex [40, 41]. The in vitro import reactions were analyzed by BN-PAGE. The results in Fig 9A show that the wild-type protein but not the 4mut variant of TAC42 accumulate in a time dependent manner in a complex of approximately 200 kDa. Moreover, when the wild-type TAC42 is imported into mitochondria lacking Sam37 a much smaller amount of the complex is observed. Furthermore, its molecular weight is slightly lower due to the absence of Sam37 [42]. In summary these results show that trypanosomal TAC42 directly interacts with the yeast SAM complex in a β-signal and Sam37-dependent manner indicating that TAC42 is indeed a novel kinetoplastid-specific mitochondrial β-barrel protein.
Thus at least two essential OM subunits of the TAC—TAC40 and TAC42—are β-barrel proteins. In line with this, ablation of Sam50, the core component of the SAM complex, leads to a rapid increase of cells lacking kDNA networks, whereas in the cells that have retained the kDNA it is massively overreplicated (Fig 9B). Ablation of Sam50 therefore essentially reproduces the phenotypes that are observed after ablation of individual TAC subunits. However, the same is not seen in cells ablated for ATOM40 the channel subunit of the main OM protein translocase [20]. Thus these results underscore the importance of the trypanosomal SAM complex for the assembly of this unique mitochondrial DNA inheritance system.
The TAC is a single unit structure that links the kDNA to the basal body. During each cell cycle a new TAC needs to be formed to guarantee that the replicated kDNA networks are correctly segregated during the binary fission of the single mitochondrion of T. brucei. Understanding TAC biogenesis is hampered by the fact that only few of its subunits are known and that their targeting pathways have not been studied. The exclusion zone filaments, that form the bridge from the basal body to the OM, consists of cytosolic proteins which may reach the TAC region by diffusion and therefore not require specific targeting. The subunits of the unilateral filaments, however, need to be imported into the mitochondrial matrix before they can assemble to link the kDNA network with the mitochondrial IM. A similar situation is found for the membrane-embedded TAC subunits in the differentiated membranes which connect the cytosolic with the matrix-localized TAC filaments. Its subunits potentially first need to be imported and inserted into the OM and IM membranes, respectively, before they are laterally sorted to the new TAC that is being assembled.
In our studies we have focused on the OM region of the differentiated membrane domain of the TAC. We have discovered two novel integral OM TAC subunits—TAC60 and TAC42, that are required for kDNA segregation—determined their topology and deciphered their biogenesis pathways.
TAC60 is essential for TAC function, contains two TMDs and its N- and C-termini both face the cytosol. In a deletion analysis we have uncoupled mitochondrial targeting from TAC sorting and identified the first TAC sorting signal. Mitochondrial targeting of TAC60 requires a 143 long region that includes the IMS-exposed loop and its more C-terminal TMD. Insertion of TAC60 into the mitochondrial OM is mediated by ATOM40, the pore-forming subunit of the master protein translocase in the OM. However, to reach its final destination, the TAC, TAC60 requires an additional 26 aa comprising the first TMD. Whether the TAC sorting signal depends on the TAC60 mitochondrial targeting signal or whether it could in principle work on its own, provided that the first TMD of TAC60 is inserted into the OM in the correct topology, remains unknown at the moment. Moreover, presently the TAC sorting signal appears to be specific for TAC60 indicating that other TAC subunits may have different sorting signals.
The essential TAC subunit TAC42 lacks predictable TMDs. Its localization to the TAC depends on both Sam50, the core component of the SAM complex, and on the presence of a β-signal consensus sequence at its C-terminus. It has been shown in yeast that this sequence mediates the interaction of β-barrel proteins with the SAM complex [39]. Moreover, TAC42 can be inserted into the OM of isolated yeast mitochondria provided that they have a functional SAM complex and that it carries a functional β-signal. In summary these result demonstrate that TAC42 is mitochondrial β-barrel protein even though in silico analysis fails to predict so.
T. brucei has six known β-barrel membrane proteins: the metabolite transporter VDAC [38], a second VDAC-like protein of unknown function [43], ATOM40 [28, 44] and Sam50 [45], the core components of the ATOM and the SAM complex, and finally TAC40 [22] and TAC42 essential subunits of the TAC. Four of them (VDAC, the VDAC-like protein, ATOM40, TAC40) belong to the mitochondrial porin protein family of whereas Sam50 is an Omp85-like protein. TAC42 is unique it is neither a mitochondrial porin nor an Omp85-like protein but defines a novel class kinetoplastid-specific β-barrel proteins essential for mitochondrial DNA inheritance.
The presence of two distinct β-barrel proteins, TAC40 and TAC42, in the TAC is striking. β-barrel proteins are exclusively found in the OMs of bacteria, mitochondria and plastids [41, 46]. It can be speculated that their prominent presence in the TAC indicates that the ancestor of this DNA inheritance system evolved very early, at a time when the integral membrane proteins present in the OM of the mitochondrial ancestor were restricted to β-barrel proteins.
Moreover, the membrane domain of the TAC shows some architectural similarity to the double membrane spanning secretion systems of gram negative bacteria [47]. Both types of structures link OM and IM of bacterial evolutionary origin. Moreover, the OM is in both cases spanned by a β-barrel type structure (generally multimeric in the case of bacteria) and thus requires a Sam50/BamA-type insertion system. However, whereas the bacterial secretion systems serve to export bacterial effector proteins, there is no evidence that the TAC is involved in transport processes.
While many subunits of the TAC are still unknown we get a progressively more detailed picture of its OM constituents. Up to now four essential, integral OM subunits of the TAC have been characterized: the β-barrel proteins TAC40 and TAC42, which form a complex with TAC60, as well as the dually localized pATOM36. This complexity is surprising since the TAC is expected to have a structural function, linking the kDNA network to the basal body of the flagellum. A single OM protein that interacts on the cytosolic side with the exclusion zone filaments and on the IMS side with an IM protein should in principle be sufficient to do this job. We would like to propose two possible explanations for this unexpected complexity.
It could be that the function of the TAC goes far beyond providing a structural linkage. Being localized between the kDNA and the flagellum, the TAC would be ideally suited to serve as a signaling platform that for example may regulate and integrate kDNA replication and segregation with flagellar growth and cytokinesis. Indeed pATOM36 has already been shown to mediate both mitochondrial protein import and mitochondrial DNA inheritance [20].
Alternatively, it might be that the TAC is the product of constructive neutral evolution. This ratched-like evolutionary process provides a non-adaptive explanation why macromolecular complexes can be comprised of more subunits than their function seem to demand [48, 49]. In the case of the TAC, a possible scenario would be that an autonomously functioning ancestral TAC subunit would fortuitously bind to another protein. Binding to this protein would not affect the function of the TAC subunit, but it would have the potential to suppress mutations, which if present in the absence of the binding partner would inactivate the TAC subunit. Should such mutations occur the TAC subunit would lose its autonomy, as its function would now depend on the other protein. Thus, constructive neutral evolution may have led to four or more OM TAC subunits even though common sense suggests a single one should be enough.
The two proposed explanations are not mutually exclusive, as both may have contributed to the complex TAC architecture. In order to disentangle the two we need a more complete picture of the TAC composition and architecture as well as a detailed functional analysis of its subunits.
Transgenic procyclic cell lines are based on T. brucei 29–13 [50] and were cultured at 27°C in SDM-79 containing 10% (v/v) fetal calf serum (FCS). Transgenic bloodstream form trypanosomes are based on the New York single marker (NYsm) strain or on a derivative thereof termed F1γL262P [27]. All bloodstream from cells were grown at 37°C in HMI-9 supplemented with 10% FCS (v/v).
Full length TAC60 (Tb927.7.1400) (Fig 1B, Fig 2A) or deletion variants of it (Fig 5, Fig 6, S1B and S2 Figs), termed ΔN114 (nt 343–1662), ΔN140 (nt 421–1662), ΔN233 (nt 697–1662), ΔN257 (nt 772–1662), ΔC153 (nt 1–1200), ΔC283 (nt 1–810), ΔN75_ΔC283 (nt 226–810), ΔN97_ΔC283 (nt 292–810), ΔC320 (nt 1–696), ΔC408 (nt 1–435) were cloned into a modified pLew100 expression vector containing a puromycin resistance gene in which a cassette had been inserted allowing C-terminal triple Myc-tagging [51].
For the experiment shown in Fig 4A, B one allele of TAC60 was tagged in situ at the C-terminus with a triple c-Myc-epitope [51] in the background of procyclic RNAi cell lines targeting ATOM40 and TbTim17, both of which have been described before [28, 52].
Full length TAC42 (Tb927.7.3060) (Fig 8) and TAC40 (Fig 8) or mutated variants of them (Fig 8), termed 1mut (TAC40: G351A; TAC42: G382A) or 4mut (TAC40: R349A, G351A, L354S and V356S; TAC42: R380A, G382A, A385S and L387S) were cloned into a modified pLew100 expression vector containing a puromycin resistance gene in which a cassette had been inserted allowing C-terminal triple Myc-tagging [51]. For the experiment shown in Fig 3A and Fig 7 one allele of TAC42 was tagged in situ at the C-terminus with a triple HA-epitope and expressed in T. brucei 29–13 cells and in a previously described RNAi cell line targeting Sam50 [52]. The in situ HA-tagged TAC40 cell line (Fig 1A) has been described before [22]. The RNAi in the procyclic and bloodstream form cell lines was targeted against the ORF of TAC60 (nt 256–772) or TAC42 (nt 238–710), respectively. For the complementation experiments of TAC60 (Fig 6), a different RNAi cell line targeting the 3'-part of the TAC60 ORF (nt 1220–1629) was established.
The Split-GFP approach was used as described [53]. For the results shown in Fig 4C, the GFP 1–10 OPT (GFP1-10) and the M3 strand 11 (GFP11) were amplified from a pET-15b-based vector (generous gift from Prof. Steven Boxer, University of Stanford). GFP 1–10 OPT was cloned into a pLew100-based expression vector containing a blasticidin resistance cassette. The resulting construct allows inducible expression of N-terminal HA-tagged cytosolic GFP1-10 (HA-GFP1-10). GFP11 was cloned into a pLew100 expression vector containing the puromycin resistance gene. The resulting construct allows C-terminal tagging of proteins with GFP11. Subsequently, the complete ORF of TAC60 was cloned into this modified pLew100 vector, yielding the construct TAC60-GFP11. Another pLew100 expression vector was established, which allows N-terminal tagging of proteins with GFP11. Subsequently, the complete ORF of TAC60 was cloned into this modified pLew100 vector, yielding the construct GFP11-TAC60. A T. brucei 29–13 cell line expressing HA-GFP1-10 was established, which subsequently was transfected with TAC60-GFP11 or GFP11-TAC60, respectively.
For visualization of tagged proteins, the respective cell lines were induced for 1 day with 1 μg/ml of tetracycline.
The following polyclonal rabbit antisera directed against the indicated antigens were produced in our lab and have been used before [23, 35]. The working dilutions for immunoblots (IB) and IF are indicated: VDAC (IB 1:1,000), ATOM40 (IB 1:10,000; IF 1:1,000), cytochrome C (IB 1:1,000), ATOM69 (IB 1:50), LipDH (IB 1:10,000) and CoxIV (IB 1:1,000). Commercially available monoclonal antibodies were used as follows: mouse c-Myc (Invitrogen, 132500; IB 1:2,000; IF 1:50), mouse HA (Enzo Life Sciences AG, CO-MMS-101 R-1000; IB 1:5,000; IF 1:1,000) and mouse EF1a (Merck Millipore, Product No. 05–235; IB 1:10,000). Monoclonal anti-tyrosinated α-tubulin antibody YL1/2 [54] (IFA 1:500) produced in rat was a generous gift from Prof. Keith Gull, University of Oxford.
Secondary antibodies for IB analysis were IRDye 680LT goat anti-mouse, IRDye 800CW goat anti-rabbit (LI-COR Biosciences, 1:20,000) and horse radish peroxidase-coupled goat anti-mouse and anti-rabbit (Sigma-Aldrich, 1:5,000). Secondary antibodies for IF were goat anti-mouse Alexa Fluor 633, goat anti-mouse Alexa Fluor 596, goat anti-rabbit Alexa Fluor 488 and goat anti-rat Alexa Fluor 488 (all from ThermoFisher Scientific, 1:1000)
Digitonin extraction was used to generate crude mitochondrial enriched fractions [55] to demonstrate mitochondrial localization of a protein of interest. For this, 5x107 or 1x108 cells were incubated for 10 min on ice in 20 mM Tris-HCl pH 7.5, 0.6 M sorbitol, 2 mM EDTA containing 0.025% (w/v) digitonin. After centrifugation (6,800 g, 4°C), the resulting mitochondria-enriched pellet was separated from the supernatant and equal cell equivalents of each fraction were subjected to SDS-PAGE and immunoblotting. Alternatively, the mitochondria-enriched pellets were used for subsequent alkaline carbonate extractions (see below).
A mitochondria-enriched pellet fraction obtained by digitonin extraction was resuspended in 100 mM Na2CO3 pH 11.5, incubated on ice for 10 min and centrifuged (100,000 g, 4°C, 10 min) to separate the membrane fraction from soluble proteins. Equal cell equivalents of all samples were analyzed by SDS-PAGE und immunoblotting.
SILAC-IP experiments were essentially done as described [20]. T. brucei 29–13 cells, their derivatives allowing expression of in situ HA-tagged TAC40 or TAC42, and a transgenic cell line allowing inducible expression of Myc-tagged TAC60 were used as indicated (Fig 1). Cells expressing or not expressing the tagged bait protein were grown in either light (unlabeled) or heavy (13C615N4-L-arginine; 13C615N2-lysine) arginine- and lysine-containing SDM-80 medium containing 10–15% (v/v) dialyzed FCS (BioConcept, Switzerland) for around 10 doubling times to establish complete labeling of proteins with light or heavy amino acids.
Equal numbers of cells grown in the presence of heavy or light arginine and lysine were mixed and harvested. The resulting pellets were solubilized in 20 mM Tris-HCl, pH 7.4, 0.1 mM EDTA, 100 mM NaCl, 10% glycerol, 1.5% (w/v) digitonin and 1X Protease Inhibitor mix (EDTA-free, Roche) for 15 min at 4°C. The extracts were centrifuged (20,000 g, 15 min, 4°C) and the resulting supernatants were incubated with anti-HA affinity beads (Roche) or anti-Myc affinity beads (EZview red, Sigma), equilibrated in the same buffer as above but containing only 0.2% (w/v) of digitonin. After 2 h of incubation at 4°C, the supernatant was discarded and the beads were washed 3 times with 0.5 ml of the same buffer. TAC40 and TAC60 protein complexes were eluted by boiling the resin for 5 min in 60 mM Tris-HCl, pH 6.8 containing 0.1% SDS whereas TAC42 protein complexes were eluted using SDS gel loading buffer without β-mercaptoethanol. SILAC-IP experiments were performed in three biological replicates including a label-switch each.
Proteins purified in TAC42 SILAC-IPs were separated on a 4–20% Mini Protean TGX SDS-PAGE gel (BioRad). Afterwards, the gel was fixed with acetic acid and methanol and stained for 3 hours with fresh ammonium sulfate-based colloidal coomassie (10% phosphoric acid, 10% ammonium sulfate, 0.12% coomassie brilliant blue G, 20% methanol). Subsequently, the gel lanes were cut into 10 pieces per replicate, which were then destained by repetitive rounds of incubation in 10 mM NH4HCO3 and 10 mM NH4HCO3 containing 50% ethanol. When completely destained, the gel pieces were dehydrated in 100% ethanol. Finally, cysteine residues of proteins were reduced with 5 mM bond breaker solution (Thermo Fisher Scientific), alkylated with 100 mM iodoacetamide in 10 mM NH4HCO3, and proteolytically digested with trypsin (37°C, incubation overnight). Proteins purified in TAC40 and TAC60 SILAC-IPs were were reduced, alkylated and tryptically digested in-solution as described previously [29].
LC-MS analyses of peptide mixtures were performed on an LTQ Oritrap XL (TAC40, TAC60) or an Orbitrap Elite (TAC42) mass spectrometer (Thermo Fisher Scientific, Bremen, Germany), each directly coupled to an UltiMate 3000 RSLCnano HPLC system (Thermo Fisher Scientific, Dreieich, Germany), as described before [32]. For quantitative MS data analysis, MaxQuant/Andromda was used (version 1.4.1.2 for TAC40, 1.5.1.0 for TAC60, and 1.5.3.30 for TAC42 data; [56, 57]). MS/MS data were searched against all entries for T. brucei TREU927 retrieved from the TriTryp database (version 8.1) applying MaxQuant default parameters with the exceptions that protein identification and quantification were based on one unique peptide and one ratio count (i.e. SILAC peptide pair). Mean log10 protein abundance ratios (TAC40/42/60 versus control) and p-value (one-sided Student's t-test) across at least two 2 biological replicates were determined. Proteins identified and quantified in TAC40, TAC42, and TAC60 SILAC IPs are listed in S1–S3 Tables.
35S-Met-labelled proteins were synthesized using the TNT T7 Quick for PCR (Promega) in vitro translation kit according to the instruction manual. Radiolabeled precursors proteins were incubated with isolated yeast mitochondria in import buffer (3% (w/v) BSA, 250 mM sucrose, 80 mM KCl, 5 mM MgCl2, 5 mM L-methionine, 2 mM KH2PO4, 10 mM MOPS-KOH, pH 7.2, 2 mM NADH, 5 mM ATP, 10 mM creatine phosphate, 0.1 mg/ml creatine kinase) at 25°C. Mitochondria were washed with SEM (250 mM sucrose, 1mM EDTA, 10mM MOPS pH 7.2) and analyzed by blue native electrophoresis (4–16% gradient gels) and autoradiography.
IF and northern blots were done as described [23]. IF images were acquired with a DFC360 FX monochrome camera (Leica Microsystrems) and a DMI6000B microscope (Leica Microsystems). Image analysis was done using LAS X software (Leica Microsystems), ImageJ, and Adobe Photoshop CS5.1 (Adobe). Relative quantification of the fluorescent intensity of kDNA networks is described in [22]. Isolation of flagella was done according to [58]. Transmission electron microscopy was exactly done as described [22].
|
10.1371/journal.ppat.1008035 | Phytoplasma SAP11 effector destabilization of TCP transcription factors differentially impact development and defence of Arabidopsis versus maize | Phytoplasmas are insect-transmitted bacterial pathogens that colonize a wide range of plant species, including vegetable and cereal crops, and herbaceous and woody ornamentals. Phytoplasma-infected plants often show dramatic symptoms, including proliferation of shoots (witch’s brooms), changes in leaf shapes and production of green sterile flowers (phyllody). Aster Yellows phytoplasma Witches’ Broom (AY-WB) infects dicots and its effector, secreted AYWB protein 11 (SAP11), was shown to be responsible for the induction of shoot proliferation and leaf shape changes of plants. SAP11 acts by destabilizing TEOSINTE BRANCHED 1-CYCLOIDEA-PROLIFERATING CELL FACTOR (TCP) transcription factors, particularly the class II TCPs of the CYCLOIDEA/TEOSINTE BRANCHED 1 (CYC/TB1) and CINCINNATA (CIN)-TCP clades. SAP11 homologs are also present in phytoplasmas that cause economic yield losses in monocot crops, such as maize, wheat and coconut. Here we show that a SAP11 homolog of Maize Bushy Stunt Phytoplasma (MBSP), which has a range primarily restricted to maize, destabilizes specifically TB1/CYC TCPs. SAP11MBSP and SAP11AYWB both induce axillary branching and SAP11AYWB also alters leaf development of Arabidopsis thaliana and maize. However, only in maize, SAP11MBSP prevents female inflorescence development, phenocopying maize tb1 lines, whereas SAP11AYWB prevents male inflorescence development and induces feminization of tassels. SAP11AYWB promotes fecundity of the AY-WB leafhopper vector on A. thaliana and modulates the expression of A. thaliana leaf defence response genes that are induced by this leafhopper, in contrast to SAP11MBSP. Neither of the SAP11 effectors promote fecundity of AY-WB and MBSP leafhopper vectors on maize. These data provide evidence that class II TCPs have overlapping but also distinct roles in regulating development and defence in a dicot and a monocot plant species that is likely to shape SAP11 effector evolution depending on the phytoplasma host range.
| Phytoplasmas are parasites of a wide range of plant species and are transmitted by sap-feeding insects, such as leafhoppers. Phytoplasma-infected plants are often easily recognized because of their dramatic symptoms, including shoot proliferations (witch’s brooms) and altered leaf shapes, leading to severe economic losses of crops, ornamentals and trees worldwide. We previously found that the virulence protein SAP11 of aster yellows witches’ broom phytoplasma (AY-WB) interferes with a specific group of plant transcription factors, named TCPs, leading to witches’ brooms and leaf shape changes of the model plant Arabidopsis thaliana. SAP11 has been characterized in a number of other phytoplasmas. However, it is not known how phytoplasmas and their SAP11 proteins modulate processes in crops, including cereals such as maize. We identified a SAP11 homolog in Maize bushy stunt phytoplasma (MBSP), a pathogen that can cause severe yield losses of maize. We found that SAP11 interactions with TCPs are conserved between maize and Arabidopsis, and that MBSP SAP11 interferes with less TCPs compared to AY-WB SAP11. This work provides new insights into how phytoplasmas change maize architecture and corn production. Moreover, we found that TCPs regulate leaf defence responses to phytoplasma leafhopper vectors in Arabidopsis, but not in maize.
| Phytoplasmas (“Candidatus (Ca.) Phytoplasma”) are economically important plant pathogens that infect a broad range of plant species. The more than 1000 phytoplasmas described so far comprise three distinct clades within a monophyletic group of the class Mollicutes that are characterized by the lack of a bacterial cell wall and small genomes (580 kb to 2200 kb) [1–3]. These fastidious pathogens are restricted to the phloem sieve cells of the plant vasculature and depend on phloem-sap-feeding insect vectors, including leafhoppers, planthoppers and psyllids, for transmission and spread in nature [4]. Many phytoplasmas induce dramatic changes in plant architecture such as increased axillary branching (often referred to as witches’ broom), formation of leaf-like flowers (phyllody), the production of green floral organs such as petals and stamens (virescence), changes of leaf shape, and premature bolting [5–10].
Phytoplasmas change plant architecture via the secretion of proteinaceous effectors that interact with and destabilize plant transcription factors with fundamental roles in regulating plant development. Effectors of Aster yellows phytoplasma strain Witches Broom (AY-WB; “Ca. Phytoplasma asteris”) are particularly well characterized. AY-WB and its predominant leafhopper vector Macrosteles quadrilineatus have broad host ranges that are mostly dicots, including Arabidopsis thaliana [6]. SAP11 destabilizes Arabidopsis TEOSINTE BRANCHED1-CYCLOIDEA-PROLIFERATING CELL FACTOR (TCP) transcription factors, and specifically class II TCPs, leading to the induction of axillary branching and changes in leaf shape of this plant [8,11], and SAP54 degrades Arabidopsis MADS-box transcription factors leading to changes in flower development that resemble phyllody and virescence symptoms [9,12]. Moreover, both effectors modulate plant defence responses leading to increased colonization of M. quadrilineatus on A. thaliana [8,9,13]. For SAP11AYWB this involves the inhibition of jasmonate (JA) synthesis [8]. SAP11 and SAP54 homologs of other phytoplasmas also target TCPs and MADS, respectively, leading to corresponding changes in plant development and architecture [10,14–16]. The majority of phytoplasma effector genes lie within composite-transposon-like pathogenicity islands named potential mobile units (PMUs) that are prone to recombination and horizontal gene transfer [17–20].
Maize bushy stunt phytoplasma (MBSP) belongs to the Aster yellows (AY) group (16SrI) “Ca. P. asteris” [21] and is the only known member of this group to be largely restricted to maize (Z. mays L.), whereas the majority, including AY-WB, are transmitted by polyphagous insects and infect dicotyledonous plants [13,22]. MBSP is transmitted by the maize-specialist insects Dalbulus maidis and D. elimatus; both MBSP and insect vectors are thought to have co-evolved with maize since its domestication from teosinte [23]. Symptoms of MBSP-infected maize plants include the formation of long lateral branches, decline in ear development and emergence of leaves that are often twisted with ripped edges and that display chlorosis and reddening [24]. We previously identified a SAP11 homolog in the MBSP genome [22] and SAP11MBSP is identical in sequence among multiple MBSP isolates collected from Mexico and Brazil [24]. SAP11AYWB and SAP11MBSP lie on microsyntenic regions within the phytoplasma genomes, indicating that these effectors are likely to have common ancestry [22]. However, D. maidis does not produce more progeny on MBSP-infected plants that show advanced disease symptoms; the insects prefer infected plants that are non-symptomatic [25]. In this study we wished to compare the roles of SAP11AYWB and SAP11MBSP in symptom induction and plant defence to insect vectors of A. thaliana and maize.
TCP transcription factors comprise an ancient plant-specific family [26] that are distinguished from other transcription factors by a conserved ± 60 amino acid TCP domain [27]. The TCP domain consists of a helix-loop-helix region that form TCP homo or heterodimers and a basic region that mediates interactions of TCP dimers with DNA motifs [28] and is required for SAP11 binding to TCPs [11]. TCP transcription factors are grouped into three clades based on TCP domain sequences: (i) class I PROLIFERATING CELL FACTOR-type TCPs (PCF clade); (ii) class II CINCINNATA-type TCPs (CIN clade); and (iii) class II CYCLOIDEA/TEOSINTE BRANCHED 1-type TCPs (CYC/TB1-clade) [29]. The latter is also known as the glutamic acid-cysteine-glutamic acid (ECE) clade [30]. PCFs promote cell proliferation, whereas CIN clade TCPs promote leaf and petal cell maturation and differentiation and have antagonistic roles to PCFs [31–34]. The ECE clade includes maize TEOSINTE BRANCHED 1 (TB1) and TB1 homologs of A. thaliana BRANCHED 1 (BRC1 (AtTCP18)) and BRC2 (AtTCP12), that repress the development of axillary branches in plants [35–38], and CYCLOIDEA (CYC) that control flower symmetry [39]. TB1 and genes in the TB1 network have been targeted for selection during maize domestication from a teosinte ancestor [40–41].
Here we show that SAP11AYWB and SAP11MBSP have overlapping but distinct specificities for destabilizing class II TCP transcription factors. The SAP11 effectors induce unique phenotypes in Arabidopsis and maize that indicate divergent roles of class II TCP transcription factors in regulating development and defence in the two plant species. We argue that SAP11MBSP evolution may be constrained due to the specific functionalities of class II TCPs in maize.
SAP11AYWB and SAP11MBSP interaction specificities for Arabidopsis TCPs (AtTCPs) were investigated via yeast two-hybrid (Y2H) assays and protein destabilization assays in A. thaliana mesophyll protoplasts. In the protoplast experiments, SAP11AYWB destabilized the majority of AtCIN-TCPs and none of the class I AtTCPs tested (Fig 1A, S1 and S2 Figs), confirming previous results [8]. In addition, SAP11AYWB also destabilized CYC/TB1-TCPs BRC1 (AtTCP18) and BRC2 (AtTCP12) (Fig 1A). In contrast, SAP11MBSP destabilized the CYC/TB1 TCPs BRC1 (AtTCP18) and BRC2 (AtTCP12), whereas 7 out of 8 class II AtCIN-TCPs and all tested class I AtTCPs remained stable (Fig 1A). The Y2H assays showed that SAP11AYWB interacts with Arabidopsis CIN-TCPs (Fig 1B), confirming previous data [8,11], whereas SAP11MBSP did not (Fig 1B). However, both SAP11AYWB and SAP11MBSP interacted with CYC/TB1 BRC1 (AtTCP18) and BRC2 (AtTCP12) (Fig 1B). Therefore, SAP11MBSP binds and destabilizes a narrower set of class II TCPs compared to SAP11AYWB.
Alignments of the SAP11AYWB and SAP11MBSP amino acid sequences revealed conservation of the signal peptide and C-terminal sequences, while the central region that includes the domains required for nuclear localization and TCP-binding of SAP11AYWB [11, 17] are more variable (Fig 2A). SAP11AYWB has a bipartite NLS that is required for nuclear localization of this effector [11]. However, the NLS sequence is not conserved in SAP11MBSP; instead NLStradamus [42] predicted the NLS to locate in the C-terminal part of the MBSP effector (Fig 2A). Localization studies with GFP-tagged SAP11 proteins in protoplasts, in the presence of BRC2 (AtTCP12), showed that both SAP11 proteins localize to plant cell nuclei, in contrast to GFP alone, which is distributed throughout the cells (S3 Fig). Therefore, the two SAP11 proteins target cell nuclei in the presence of BRC2 (AtTCP12). We previously demonstrated that the MEILKQKAEEETKNL of SAP11AYWB is required for TCP-binding, whereas deletion of the C-terminal KEEGSSSKQPDDSKK sequence did not affect the TCP binding of the effector [11]. Therefore, we assigned the MEILKQKAEEETKNL sequence as the TCP-binding domain (Fig 2A). To dissect what sequences within the SAP11 protein determine binding specificity to TCPs, we generated chimeras between SAP11AYWB and SAP11MBSP and studied their binding to CYC/TB1 BRC1 (AtTCP18) and CIN TCP2 (Fig 2B, S4A Fig). Therefore, the TCP-interaction domain plays a role in determining SAP11 binding specificity to the two TCPs (Fig 2A and 2B, S4A Fig).
To investigate which region of the TCP domain determine SAP11 binding specificity, chimeras of the basic region and helix loop helix regions of the TCP domains of CIN-TCP AtTCP2 and CYC/TB1-TCP BRC1 (AtTCP18) were constructed (Fig 2C and 2D) and tested for interactions with the two SAP11 proteins. SAP11AYWB and SAP11MBSP interacted with the TCP domains of AtTCP2 and BRC1 (AtTCP18) (Fig 2D, S4B Fig), as observed for full-length TCPs (Fig 1B), confirming that the TCP domain itself is sufficient for SAP11 interaction and specificity. Furthermore, SAP11AYWB interacted with all AtTCP2-BRC1 (AtTCP18) chimeras used in the assay (Fig 2D, S4B Fig), whereas SAP11MBSP interacted with chimeras containing BRC1 (AtTCP18) helix-loop-helix and AtTCP2 basic regions, but not with those composed of AtTCP2 helix-loop-helix and BRC1 (AtTCP18) basic region or with mixed helix, loop and helix sequences (Fig 2D, S4B Fig). Therefore, the entire helix-loop-helix region of the TCP domain is required for the specific binding of SAP11MBSP to CYC/TB1 TCPs. Taken together, multiple amino acids are likely to determine the specificity of SAP11-TCP interactions.
To investigate if the SAP11 binding specificity to TCPs aligns with in planta interactions, phenotypes of A. thaliana Col-0 stable transgenic lines that produce SAP11AYWB and SAP11MBSP under control of the 35S promoter (Fig 1C) were compared to those of the A. thaliana brc1-2 brc2-1 double mutant, hereafter referred to as the brc1 brc2 mutant, which is a null mutant for both CYC/TB1-TCPs BRC1 (AtTCP18) and BRC2 (AtTCP12) [35] and the 35S::miR319a x 35S::miR3TCP line in which CIN-TCPs are knocked down [31]. Whereas the crinkled leaves of 35S::SAP11AYWB lines phenocopied those of 35S::miR319a x 35S::miR3TCP (Fig 1D) [8], leaves of 35S::SAP11MBSP lines were not crinkled and more similar to WT Col-0 leaves (Fig 1D). Rosette diameters of the 35S::SAP11AYWB and 35S::miR319a x 35S::miR3TCP lines were smaller than WT Col-0 plants, unlike the rosettes of 35S::SAP11MBSP and A. thaliana brc1 brc2 mutant lines that looked similar to those of WT plants (Fig 1F). Both 35S::SAP11AYWB and 35S::SAP11MBSP lines produced significantly more primary rosette-leaf branches (RI) [35] than WT plants. With exception of the 35S::SAP11MBSP line 3 that had a lower number of RIs, the production of RI was similar to the A. thaliana brc1 brc2 mutant. In contrast, 35S::miR319a x 35S::miR3TCP plants produced a reduced number of RI compared to WT Col-0 (Fig 1E and 1G, S5E Fig). Therefore, 35S::SAP11MBSP lines phenocopied the A. thaliana brc1 brc2 mutant and the 35S::SAP11AYWB lines both the A. thaliana brc1 brc2 and 35S::miR319a x 35S::miR3TCP mutant lines, indicating that SAP11AYWB destabilizes Arabidopsis CIN and CYC/TB1 TCPs and SAP11MBSP the CYC/TB1-TCPs BRC1 (AtTCP18) and BRC2 (AtTCP12), in agreement with the results of protoplast-based destabilization and Y2H binding assays.
Beyond phenotypes described above, we found that the 35S::miR319a x 35S::miR3TCP and 35S::SAP11AYWB lines produced less rosette leaves compared to WT plants, unlike the A. thaliana brc1 brc2 and 35S::SAP11MBSP lines (S5A Fig). Bolting time, plant height and numbers of primary cauline-leaf branches (CI) [35] were variable among the 35S::SAP11AYWB and 35S::SAP11MBSP lines (S5B–S5E Fig). Roots of 35S::miR319a x 35S::miR3TCP and 35S::SAP11AYWB lines were consistently shorter compared to WT plants as described by Lu et al. [43]. In contrast, the root length of A. thaliana brc1 brc2 and 35S::SAP11MBSP lines did not show obvious differences compared to those of WT plants (S6 Fig).
We previously showed that the AY-WB insect vector M. quadrilineatus produces 20–30% more progeny on 35S::SAP11AYWB A. thaliana [8]. By repeating this experiment and including 35S::SAP11MBSP A. thaliana, we confirmed the previous result for 35S::SAP11AYWB A. thaliana but not for 35S::SAP11MBSP A. thaliana (Fig 3A). Therefore, SAP11AYWB appears to modulate plant defences in response to M. quadrilineatus, whereas SAP11MBSP does not. To test this further, the transcriptomes of wild type, 35S::SAP11AYWB and 35S::SAP11MBSP A. thaliana with and without exposure to M. quadrilineatus were compared via RNA-seq (S1 Table, GEO accession GSE118427). PCA showed that, in samples exposed to M. quadrilineatus, 35S::SAP11MBSP and WT Col-0 group together, whereas the 35S::SAP11AYWB samples form a separate group (Fig 3B). Therefore, SAP11AYWB has a measurable impact on the transcriptome of A. thaliana, unlike SAP11MBSP.
Analyses of differentially expressed genes (DEGs) of Col-0 and transgenic plants exposed to M. quadrilineatus identified 96 DEGs for 35S::SAP11AYWB versus Col-0 and only one DEG for 35S::SAP11MBSP versus Col-0 (Fig 3C and 3D). Hierarchical cluster of the DEGs expression levels was in agreement with the PCA results demonstrating that the M. quadrilineatus-exposed 35S::SAP11AYWB treatments cluster separately from those of Col-0 and 35S::SAP11MBSP (Fig 3E, S2 Table). Moreover, M. quadrilineatus-exposed 35S::SAP11AYWB treatments cluster together with non-exposed samples. Of the 96 DEGs 30 have a role in regulating plant defence responses, including hormone and secondary metabolism, such as Myb, AP2/EREBP and bZIP transcription factors, receptor kinases, cytochrome P450 enzymes, proteases, oxidases and transferases (highlighted in yellow, S3 Table). The 96 genes also included 11 natural anti-sense genes and at least 30 genes with unknown functions. Taken together, these data indicate that defence responses to M. quadrilineatus are suppressed in 35S::SAP11AYWB plants.
To investigate SAP11 interactions with maize TCPs we first identified maize TCP sequences. The CDS of 44 Z. mays (Zm) TCPs available on maize TFome collection [44] were extracted from the Grass Regulatory Information Server (GRASSIUS) (http://grassius.org/grasstfdb.html) [45]. We identified two class II CYC/TB1-TCPs, including TB1 (ZmTCP02) and ZmTCP18, 10 class II CIN-TCPs and 17 class I PCF-like TCPs. The ZmTCPs were assigned to groups based on characteristic TCP domain amino acids conserved in each of the groups, highlighted in yellow, red and green (Fig 4) [29]. In contrast to A. thaliana, maize appears to have an additional group of class II TCPs that share amino acids conserved in the TCP domains of both CIN and TB1/CYC TCPs (Fig 4). One of these is BRANCHED ANGLE DEFECTIVE1 (BAD1), which is expressed in the pulvinus to regulate branch angle emergence of inflorescences, particularly the tassel [46]. BAD1 was placed in a subclade of CYC-TB1 TCPs named as TCP CII. Hence, we assigned all members in this additional group to TCP CII. TCPs similar to TCP CII appear to be absent in the monocots sorghum (S. bicolor) and rice (O. sativa) (S7 and S8 Figs, S4 Table). Seven CIN-TCPs of maize, four CIN-TCPs of rice and five CIN-TCPs of sorghum are potentially regulated by miR319a (Fig 4, S7–S9 Figs). While this study was ongoing, Chai et al. [47] reported the expression characteristics of 29 maize TCPs. To promote consistency, we adopted their nomenclature for these TCPs as ZmTCP01 to ZmTCP29, and continued the numbering of the additional 15 maize TCP genes extracted from GRASSIUS as ZmTCP30 to ZmTCP44 (Fig 4, S4 Table).
Y2H assays revealed that SAP11MBSP interacts with the CYC/TB1-TCPs TB1 (ZmTCP02) and ZmTCP18, but not with ZmTCP members of the CIN and CII subgroups (Fig 5A). In contrast, SAP11AYWB interacted also with CIN and CII ZmTCPs (Fig 5A). GFP-SAP11MBSP and GFP-SAP11AYWB destabilized HA-tagged TB1 (ZmTCP02) and ZmTCP18 in maize protoplasts in contrast to GFP controls (Fig 5B, S10 Fig), indicating that the SAP11 homologs also destabilize maize TCPs in maize cells.
SAP11AYWB and SAP11MBSP were cloned as N-terminal 3XFLAG tag fusions downstream of the maize Ubiquitin promoter, and transformed into HiIIAXHiIIB hybrid Z. mays. Ubi::FLAG-SAP11MBSP primary transformants (T0) were female sterile, but produced pollen, which were used for fertilizing flowers of a wild type HiIIA plant. In contrast, Ubi::FLAG-SAP11AYWB primary transformants were male sterile, but produced flowers, which were successfully fertilized with pollen from a HiIIA plant. The T1 progenies of both crosses had similar production of SAP11 proteins (Fig 5C) and were further phenotyped.
Unlike WT HiIIA, Ubi::FLAG-SAP11MBSP T1 plants produced multiple tillers arising from the base of the main culm (Figs 5D (a, c) and 6). Both main culm and tillers produced apical male inflorescences with tassels that carried anthers with pollen (Figs 5D (j, l, insets 7, 10, 11) and 6). These pollen were fertile, as they were used to pollinate HiIIA female inflorescence for seed reproduction. At the upper nodes of the main culm where in WT plants short primary lateral branches with apical ears would develop from the leaf sheath (Figs 5D (g) and 6), long primary lateral branches emerged that also had apical tassels (Figs 5D (i, inset 3) and 6). Hence, Ubi::FLAG-SAP11MBSP plants were female sterile. These phenotypes of Ubi::FLAG-SAP11MBSP plants are similar to those of the Z. mays tb1 mutant (Fig 6) [40,48]. Essentially, Ubi::FLAG-SAP11MBSP and Z. mays tb1 mutant lines resemble teosinte, though the latter produces small ears located at multiple lateral positions of the primary lateral branches (Fig 6) [49]. Therefore, Ubi::FLAG-SAP11MBSP plants phenocopy the maize tb1 mutant, in agreement with the results of yeast two-hybrid and protoplast destabilization assays showing that SAP11MBSP destabilizes CYC/TB1 TCPs.
MBSP-infected maize plants show multiple tillers developing from the base of the main culm and primary lateral branches with apical tassels [24], like Ubi::FLAG-SAP11MBSP and tb1 maize plants (Fig 6). The MBSP-infected maize plants produce ears at the same position where the elongated lateral branches appear, though the ears are fewer in number, substantially smaller and produce less seed than WT non-infected plants. The latter may occur because of tillering, which distributes energy/carbon to the many tillers rather than the development of kernels leading to reduced fertility, and possibly because SAP11MBSP inhibits development of female reproductive organs, like in Ubi::FLAG-SAP11MBSP and tb1 maize plants. In the case of the transgenic plants, SAP11MBSP is being expressed from the start, when the plants grow up, whereas for the MBSP infection, the plants are exposed to the effector later when already partly developed (maize plants are infected with MBSP when they are 3 weeks old, as the leafhopper vectors tend to kill the maize seedlings when they are younger). Phytoplasmas are phloem-limited, and SAP11 and other effectors secreted by the phytoplasmas can unload from the phloem and migrate to distant tissues, including the apical meristem [17,50]. Therefore, it is highly likely that SAP11 interact with and destabilize TCP transcription factors during infection, in agreement with the symptoms of MBSP-infected plants.
Ubi::FLAG-SAP11AYWB T1 plants also produced more tillers from the base of the main culm, but were shorter than WT HiIIA and Ubi::FLAG-SAP11MBSP (Fig 5D (a, b, c)). The majority of leaves of Ubi::FLAG-SAP11AYWB plants had curly edges, unlike Ubi::FLAG-SAP11MBSP and HiIIA plants (Fig 5D (d, e, f, h, inset 2)). Ubi::FLAG-SAP11AYWB plants produced red-coloured silks emerging directly from the leaf sheath without prior ear formation (Figs 5D (h, inset 2) and 6). Upon pollination of the red-coloured silks, ears with reduced husk leaves and exposed corn emerged (Fig 5E (o)). As well, the tip of the main culm and tillers carried tassel-like structures with female flowers and emerging silks (Figs 5D (k, insets 8, 9) and 6). Pollination of these silks with HIIA pollen induced the formation of a few corns (Fig 5E (m,n)). Thus, SAP11AYWB induces tassel feminization and interferes with leaf development, including the modified leaves that generate the husk of the ear.
We investigated if SAP11AYWB and SAP11MBSP modulate maize processes in response to the AY-WB and MBSP insect vectors M. quadrilineatus and D. maidis, respectively. We did not observe any differences in fecundity of both insect vectors on HiIIA, Ubi::FLAG-SAP11AYWB and Ubi::FLAG-SAP11MBSP plants (Fig 7A and 7B). PCA of RNA-seq data from WT and transgenic maize plants indicate that SAP11AYWB and SAP11MBSP modulate maize transcriptomes with SAP11AYWB having a larger effect than SAP11MBSP (Fig 7C and 7D, S5 and S6 Tables, GEO: GSE118427), in agreement with morphological data of the maize lines (Figs 5 and 6). However, M. quadrilineatus-exposed HiIIA Ubi::FLAG-SAP11AYWB and Ubi::FLAG-SAP11MBSP maize clustered together and separately from non-exposed maize in PCA (Fig 7C). D. maidis exposed maize samples grouped with the non-exposed ones (Fig 7D), suggesting that the SAP11 homologs do not have obvious effects on transcriptome responses of maize to the insects. Moreover, M. quadrilineatus has a larger impact and D. maidis a minor impact on maize gene expression (Fig 7C and 7D). Together, these data indicate that SAP11AYWB and SAP11MBSP do not alter maize susceptibility to M. quadrilineatus and D. maidis.
We found that SAP11AYWB and SAP11MBSP have overlapping, but distinct, binding specificities for class II TCP transcription factors. We identified the TCP-interaction domain that is involved in determining the specificity of SAP11AYWB and SAP11MBSP binding to CYC/TB1 and CIN-TCPs and found that the two effectors bind to the helix-loop-helix region of the TCP domain of the TCP transcription factors. The helix-loop-helix region of the TCP domain is required for TCP-TCP dimerization and configuration of the TCP domain beta sheets of both TCP transcription factors in a way that allows binding of the beta sheets to promoters [28]. We also found that SAP11-TCP binding specificities are correlated with the ability of the SAP11 homologs to destabilize these TCPs in leaves [8] and protoplasts (this study) and the induction of specific phenotypes in plants [8, this study]. Whereas it remains to be resolved how SAP11 destabilizes TCPs, it is clear that SAP11 is highly effective at destabilizing TCPs in plants as evidenced by the specific SAP11-induced changes in A. thaliana and maize architectures that phenocopy TCP mutants and knock-down lines of these plants.
TCP domains of each TCP (sub)class have characteristic amino acid sequences that have remained conserved after the divergence of monocots and eudicots [51]. The helix-loop-helix regions are characteristic for each TCP (sub)class and are conserved among plants species, including dicots and monocots. We found that SAP11 binding specificity is determined by TCP (sub)class rather than plant species, as SAP11MBSP specifically interacts with class II CYC/TB1-TCPs of both A. thaliana and maize, and not class II CIN-TCP and class I TCPs of these divergent plant species. Similarly, SAP11AYWB interacts with all class II TCPs and not the class I TCPs of A. thaliana and maize. Therefore, SAP11AYWB and SAP11MBSP binding specificity is likely to involve multiple amino acids within the TCP-interaction domain of the SAP11 proteins and the helix-loop-helix region of the TCP domain.
We found that SAP11MBSP interacts with and destabilizes TCPs of the TB1 clade, including A. thaliana BRC1 (AtTCP18) and BRC2 (AtTCP12) and maize TCP02 and TCP18. These binding specificities are supported by plant phenotypes; A. thaliana 35S::SAP11MBSP and maize Ubi::FLAG-SAP11MBSP lines phenocopy A. thaliana brc1 brc2 lines and maize tb1 lines, respectively. The A. thaliana 35S::SAP11MBSP lines show stem proliferations, in agreement with A. thaliana BRC1 (AtTCP18) and BRC2 (AtTCP12) and maize TB1 (ZmTCP02) being suppressors of axillary bud growth [38,52–54]. We also show that A. thaliana 35S::SAP11MBSP and brc1 brc2 lines produce fully fertile flowers, whereas maize Ubi::FLAG-SAP11MBSP plants produced only male tassels and no female inflorescences like maize tb1 plants [40,48]. This is in agreement with BRC1 (AtTCP18) not directly affecting A. thaliana flower architecture [55,56], and maize TB1 (ZmTCP02) being a direct positive regulator of MADS-box transcription factors that control maize female inflorescence architecture [41]. Interestingly, many phytoplasmas have SAP54 effectors, which degrade MADS-box transcription factors leading to the formation of leaf-like sterile flowers [9,10,57,58] whereas no effector gene with sequence similarity to SAP54 was identified in MBSP [24]. It is possible that the maize-specialist phytoplasma strain does not require an additional effector (such as SAP54) to modulate floral development of its host, as SAP11MBSP indirectly targets flowering via TB1 (ZmTCP02).
AtTCP10, which is a CIN-TCP, appears to be destabilized by both SAP11 effectors. This is unexpected given that TCP domains are extremely conserved among TCPs in which those of the CIN-TCPs and CYC/TB1 TCPs are distinct (S11 Fig). We demonstrate that SAP11MBSP binding to CYC/TB1 TCPs requires the entire helix-loop-helix region of CYC/TB1 TCPs, as replacement of the loop or helices with that of a CIN-TCP prevents binding of SAP11MBSP. Based on this, SAP11MBSP is unlikely to bind AtTCP10 directly. TCPs are known to regulate the expression of each other and may also form complexes and, therefore, the expression and abundance of (some) CIN-TCPs may be indirectly affected by deregulation of (SAP11MBSP-mediated) CYC/TB1 TCPs.
Whereas SAP11MBSP interacts and destabilizes TB1 TCPs, SAP11AYWB interacts with all class II TCPs of A. thaliana and maize, in agreement with A. thaliana 35S::SAP11AYWB lines phenocopying both A. thaliana brc1 brc2 and A. thaliana 35S::miR319a x 35S::miR3TCP lines. Information about the role of TCPs in maize development are limited, potentially due to redundant functions of TCPs belonging to the same subgroup and the challenges of obtaining multiple knockdown lines. Therefore, at this time we do not know if maize Ubi::FLAG-SAP11AYWB lines phenocopy maize mutant lines for all CIN and CII TCPs. Nonetheless the leaf crinkling phenotypes of Ubi::FLAG-SAP11AYWB maize plants are in agreement with what is known about the functions of CIN TCPs in Arabidopsis where CIN TCPs play a role in leaf development [8,33,59]. The CII subgroup member BAD1 regulates branch angle emergence of the maize tassel [46] indicating that CII TCPs regulate male inflorescence development in maize. Such a role of these TCPs in maize is consistent with the phenotype of Ubi::FLAG-SAP11AYWB maize plants, given that these produce only female reproductive organs; that is, male developmental organs may not be produced due to absence of CII (and CIN) TCPs in Ubi::FLAG-SAP11AYWB maize plants. Therefore, our finding that Ubi::FLAG-SAP11AYWB maize plants solely producing female inflorescences and no tassels expands the current knowledge about maize CII and CIN-TCPs to a potential role in plant sex determination. We cannot fully exclude the possibility that SAP11AYWB destabilizes other proteins in maize, though we think this is unlikely given our finding that SAP11-TCP interactions are specific involving conserved TCP helix-loop-helix sequences and that SAP11AYWB induces changes in A. thaliana development that are entirely consistent with destabilization of class II TCPs in this plant.
We previously demonstrated that 35S::SAP11AYWB A. thaliana plants are affected in jasmonate production and LOX2 expression upon wounding and that the AY-WB insect vectors produce more progeny on LOX2-silenced plants [8]. A number of TCPs have roles in plant JA production regulation [32,60–65]. Here, we show a clear role of SAP11AYWB suppression of plant defence response genes to M. quadrilineatus, including those involved in phytohormone responses. These genes were not differentially regulated in SAP11MBSP plants response to M. quadrilineatus, indicating that destabilization of CIN-TCPs alone or in combination with Arabidopsis BRC1 (AtTCP18) and BRC2 (AtTCP12) alters plant defence responses to M. quadrilineatus. SAP11AYWB does not promote M. quadrilineatus and D. maidis fecundity on maize suggesting that maize class II TCPs do not play a major role in regulating defence responses of maize leaves. Therefore, class II TCPs appear to regulate plant defence responses in leaves of Arabidopsis but not in maize.
MBSP and the insect vectors D. maidis and D. elimatus are thought to have co-evolved with maize since its domestication from teosinte [23]. We previously sequenced the genomes of MBSP isolates from geographically distant locations and found single nucleotide polymorphisms (SNPs) throughout the genomes of these isolates but that SAP11MBSP remained conserved [24]. The effector may be subject to purifying selection because the destabilization of maize TB1 TCPs and subsequent induction of axillary branching and inhibition of female flower production promote MBSP fitness in maize in a manner that is so far unknown. As well, SAP11MBSP evolution may be constrained by possibly negative effects of maize CIN and ECE TCP destabilization on MBSP fitness or because SAP11MBSP alleles that destabilize other maize TCPs may not be selected in MBSP populations because maize TCPs do not impact D. maidis fitness. Finally, both D. maidis and MBSP predominantly colonize maize, whereas M. quadrilineatus and AYWB colonize a wide range of plants species presenting the possibility that a positive effect of SAP11 on insect fecundity may have more benefit for a generalist phytoplasma and insect vector than for more specialized ones.
In conclusion, we found that SAP11 effectors of AY-WB and MBS phytoplasmas have evolved to target overlapping but distinct class II TCPs of their plant hosts and that these transcription factors also have overlapping but distinct roles in regulating development in these plant species. In addition, TCPs may or may not impact plant defence responses to phytoplasma leafhopper vectors. The distinct roles of TCPs in regulating plant developmental and defence networks are likely to shape SAP11 effector evolution of phytoplasma.
We generated Gateway compatible entry clones for all experiments, except for the constructs to transform maize. The cloning of the codon-optimized version of SAP11AYWB without the sequence corresponding to the signal peptide into pDONR207 is described previously [8]. The cloning of sequences corresponding to the open reading frames (ORFs) of AtTCP2, AtTCP3, AtTCP4, AtTCP5, AtTCP7, AtTCP10, AtTCP13 and AtTCP17 (S4 Table) into pDONR207 was also done previously [7]. The full-length ORF of AtTCP6, AtTCP8, AtTCP9, AtTCP12, AtTCP14 and AtTCP18 (S4 Table) were PCR amplified from complementary DNA (cDNA) with gene-specific primers that contain partial sequences of the attB1 and attB2 Gateway recombination sites (S7 Table). The fragments were further amplified with attB1 and attB2 adapter primers and cloned into pDONR207 with Gateway BP Clonase II Enzyme Mix (Invitrogen, Carlsbad, USA). Gateway compatible pENTR/SD/D/TOPO vectors containing the full length ORFs of ZmTCP01 (clone UT5707), ZmTCP02 (clone UT5978), ZmTCP05 (clone UT1680), ZmTCP12 (clone UT6182), ZmTCP13 (clone UT3439) and ZmTCP18 (clone UT4097) were ordered from The Arabidopsis Information Resource (TAIR) (S4 Table). A codon-optimized version of SAP11MBSP without the sequence corresponding to the signal peptide and DNA sequences corresponding to the TCP domains of ZmTCP9, AtTCP12, AtTCP18 and the AtTCP2 and SAP11 chimeras were gene synthesized by Genscript (New Jersey, USA) with Gateway compatible attL1 and attL2 attachment sites (S4 and S8 Tables) and provided in pMS (Genscript).
All genes were transferred from the Gateway compatible entry clones into the respective expression vectors with the Gateway LR Clonase II enzyme mix (Invitrogen). Full-length ORFs of all TCPs were cloned into pUGW15 [66] to produce N‐terminally HA‐tagged proteins. The codon-optimized versions of SAP11AYWB and SAP11MBSP without signal peptide sequences were cloned into pUBN-GFP-DEST [67] to produce N‐terminally GFP‐tagged SAP11AYWB and SAP11MBSP. To generate a plasmid for expression of GFP alone, the ccdB cassette of pUBN-GFP-DEST was replaced with a GFP sequence that carries two translational stop codons instead of the translational start codon. The GFP-sequence was amplified from pUBN-GFP-DEST with the gene-specific primers STOP-GFP forward and reverse (S7 Table), cloned into pDONR207 with the Gateway BP Clonase II Enzyme Mix (Invitrogen) and transferred to pUBN-GFP-DEST using the Gateway LR Clonase II Enzyme Mix (Invitrogen).
Isolation and transformation of Arabidopsis and maize protoplasts were performed as described by [68]. Protoplasts were generated from 6-week-old Arabidopsis and four-leaf stage maize plants grown in controlled environmental conditions with a 14h, 22 C°/ 10h, 20°C light / dark period. The maize plants were transferred into dark for five days before protoplast isolation. 600-μl-protoplast-suspensions were transformed with the indicated constructs and placed in the dark for 12h for gene expression. Protoplasts were harvested by mild centrifugation (1 min, 200 x g) and mixed with 20μl 2X sodium dodecyl sulfate (SDS)- polyacrylamide gel electrophorese (PAGE) sample buffer (50 mM Tris/HCl, 10% (w:v) SDS, 50% (v:v) glycerol, 0.02% bromophenolblue, 10% ß-mercaptoethanol, pH = 6.8). Samples were separated in an SDS-PAGE using 15% SDS-polyacrylamide gels and blotted on 0.45μm BA85 Whatman Protran nitrocellulose membranes (Sigma-Aldrich) with the BioRad (Life Science, Hemel Hempstead, UK) minigel and blotting system. Proteins were detected via western blot hybridization with specific antibodies. For detection of GFP-fusion proteins, anti-GFP polyclonal primary antibody (Santa Cruz Biotechnology, Dalla, USA, catalog number: sc-8334, diluted 1:1000) and anti-rabbit-HRP secondary antibody (Sigma-Aldrich, diluted 1:10000) were used. After the anti GFP-antibodies were removed by treatment of the membrane with 0.2 M glycine, 0.1% SDS, 100 mM ß-mercaptoethanol, pH = 2, the HA-fusion proteins were detected on the same blot with anti-HA11 monoclonal primary antibody (Covance, New Jersey, USA, order number: MMS-101P, diluted 1:1000) and anti-mouse-HRP secondary antibody (Sigma-Aldrich, diluted 1:10000).
All genes were transferred from the above generated Gateway compatible entry clones into the respective Yeast Two-Hybrid vectors with the Gateway LR Clonase II enzyme mix (Invitrogen). The codon-optimized sequences corresponding to mature proteins (without signal peptides) of SAP11AYWB, SAP11MBSP and SAP11 chimeras were transferred into pDEST-GAD-T7 [69]. The TCP sequences encoding for full length TCPs or TCP domains were transferred into the pDEST-GBK-T7 [69]. Saccharomyces cerevisiae strain AH109 (Matchmaker III; Clonetech Laboratories, Mountain View, CA, USA) was transformed using a 96-well transformation protocol [70] and interaction studies were carried out on media depleted of leucine, tryptophan and histidine with addition of 20 mM 3-Amino-1,2,4-triazole (3AT) to suppress auto activation.
The generation and analysis of the 35S::SAP11AYWB Arabidopsis Col-0 lines, was described previously [8]. Idan Efroni (Weizmann Institute of Science, Rehovot, Israel) provided seeds of the 35S::miR319a x 35S::miR3TCP Arabidopsis Col-0 lines described in Efroni et al. [31] and Pilar Cubas (Centro Nacional de Biotecnologia, Madrid, Spain) provided seeds of the brc1 brc2 Arabidopsis Col-0 line described in Aguilar-Martinez et al. [35]. For generation of the 35S::SAP11MBSP Arabidopsis Col-0 lines the codon optimized version of the SAP11MBSP sequence without the sequence corresponding to the signal peptide was transferred from the Gateway compatible entry clone (described above) into the pB7WG2 binary vector using the Gateway LR Clonase II Enzyme Mix (Invitrogen) and Arabidopsis Col-0 plants were transformed using the floral dipping method [71].
SAP11 transcript levels in 35S::SAP11AYWB and 35S::SAP11MBSP A. thaliana plants were quantified in mature leaves of three independent, 5-week-old plants. Total RNAs were extracted from 100 mg snap frozen A. thaliana leaves with TRI-reagent (Sigma Aldrich) and cDNA synthesis was performed from 0.5 μg total RNA using the M-MLV-reverse transcriptase (Invitrogen). cDNA was subjected to qRT-PCR using SYBR Green JumpStart Taq ReadyMix (Sigma-Aldrich) in a CFX96 Touch Real-Time PCR Detection System (Biorad) using gene-specific primers for the SAP11-homologs and Actin 2 (AT3G18780) (S9 Table).
A. thaliana seeds were sterilized in 5% sodium hypochlorite for 8 minutes and washed five times with sterile water. Seeds were germinated on ½ x MS medium with 0.8% (w/v) agar. Three days after germination, seedlings were transferred to ½ x Hoagland medium [72] with 0.25 mM KH2PO4 containing 1% (w/v) sucrose and 1% (w/v) agar [43]. Plates were placed vertical to allow root growth on the agar surface. After an additional growth period of 10 days seedlings were removed from the plates individually and their root length measured using a ruler.
Codon optimized versions of the SAP11AYWB and the SAP11MBSP sequences without sequences corresponding to the signal peptide including a sequence encoding an N-terminal 3xFLAG-tag were synthesized with flanking BamH1 and EcoRI restriction sites (S10 Table) that were used for cloning into the multiple cloning site of the p1u Vector (DNA Cloning Service, Hamburg, Germany). The resulting Ubi::FLAG-SAP11-nos cassette was transferred from p1U into the binary Vector p7i (DNA Cloning Service, Hamburg, Germany) via SfiI restriction sites. Agrobacterium-mediated transformation of maize HiIIAxHiIIB embryos and BASTA (Bayer CropScience, Monheim, Germany) selection of T0 transgenic HiIIAxHiIIB plants was performed by Crop Genetic Systems (CGS) UG (Hamburg, Germany). This resulted in the three independent, transgenic, heterozygous lines of UBI::FLAG-SAP11AYWB 1–3 and the two independent, transgenic, heterozygous lines of UBI::FLAG-SAP11MBSP 1–2. For seed reproduction T0 transgenic plants were crossed with HiIIA plants because the described defects in sexual organs development (Fig 5) impeded self-pollination. Plants were analyzed for production of proteins from transgenes via western blot hybridizations (explained above) with anti-FLAG M2 monoclonal primary antibody (Sigma-Aldrich, order number: F3165, diluted 1:1000) and anti-mouse-HRP secondary antibody (Sigma-Aldrich, diluted 1:10000) and then used for experiments.
Plants were grown under controlled environmental conditions with a 14h, 22 C°/ 10h, 20°C light / dark period for Arabidopsis and 16h, 26°C/ 8h, 20°C light/dark period for maize. Seven-week-old Arabidopsis and three-week-old maize plants were individually exposed to 10–15 adult M. quadrilineatus or D. maidis insects (7–10 females and 3–5 males) for 3 days. The insects were removed and progeny (nymphs or adults) were counted four weeks later.
Fully expanded leaves of seven-week-old A. thaliana Col-0 wt and transgenic plants were exposed to five adult M. quadrilineatus (2 males and 3 females) in a single clip cage with one clip-cage per plant. For the generation of non-treated samples, clip-cages were applied without insects. After 48h the areas covered by the clip-cages were harvested, snap frozen in liquid nitrogen and stored at -80°C until further processing for RNA extraction. For maize, complete three-week-old maize HiIIA wild type (WT) or transgenic plants were exposed to 50 adult M. quadrilineatus or D. maidis insects (20 males and 30 females) for 48 hours and the complete above soil plant material was harvested, snap frozen in liquid nitrogen and stored at -80°C until further processing for RNA extraction.
Total RNA was extracted from ground Arabidopsis leaf tissue and from 200 mg ground maize material using the RNeasy plant mini kit with on-column DNase digestion (Qiagen). The RNA-seq data of the A. thaliana experiments were generated at Academia Sinica (Taipei, Taiwan) and at the Earlham Institute (EI, Norwich, UK). The RNA-seq data of all maize experiments were generated at EI. At Academia Sinica, libraries were generated with the llumina Truseq strand-specific mRNA library preparation without size selection, and sequenced on the Illumina HiSeq2500, 125-bp paired-end reads (YOURGENE Bioscience, New Taipei City, Taiwan). Libraries at EI were generated using NEXTflex directional RNA library (HT) preparation (Perkin Elmer, Austin, Texas, USA) and sequencing was done on the Illumina HiSeq4000, 75-bp paired-end reads (EI). To assess if the RNA-seq data for the A. thaliana experiments received from EI and Academia Sinica are comparable, four samples were sequenced at both facilities. Principal Component Analysis (PCA) showed that the samples generated by these two facilities cluster together demonstrating that batch effects are negligible (S12 Fig).
The adapter sequences of the raw RNAseq reads were removed using Trim Galore, version 0.4.4 (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). The paired-end reads were aligned to the reference genome (A. thaliana/TAIR 10.23 and Z. mays/AGPv4) with the software TopHat, version 2.1.1 [73]. The number of aligned reads per gene was calculated using HTSeq, version 0.6.1 [74], and data were initially analysed via PCA, using the R/Bioconductor package DESeq2 [75]. Obvious outliers were excluded from the analysis; this amounted to one sample per experiment, as follows: one wild type (WT) Col-0 + M. quadrilineatus sample from the A. thaliana experiment; one Ubi::FLAG-SAP11AYWB + M. quadrilineatus sample from one of the maize experiments; one Ubi::FLAG-SAP11AYWB + D. maidis sample from the other maize experiment; and one Ubi::FLAG-SAP11MBSP sample in common with both experiments (S13 Fig, S1, S5 and S6 Tables). Differential expression analysis was conducted with DESeq2, using the function -contrast- to make specific comparisons. For further analyses we selected genes that satisfy 3 criteria: p value <0.05 after accounting for a 5% false discovery rate (FDR) (Benjamini-Hochberg corrected), mean gene expression value >10 and fold change in expression >2. Cluster analysis was performed on z-score normalized data using the hierarchical method [76].
RNA-seq data of M. quadrilineatus and D. maidis males and females (~25 million reads each) were downloaded from NCBI, accession number SRP093182 and SRP093180 respectively. The reads were used for de novo assemblies of male and female transcriptomes separately. Reads were trimmed to remove adaptor sequence and low-quality reads using Trim Galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). Reads over 20-bp in length were retained for downstream analysis. Trimmed reads were de novo assembled using Trinity r20140717 [77] allowing a minimum contig length of 200 bp and minimum k-mer coverage of 2 with default parameters. Assembled contigs were made non-redundant and lowly expressed contigs were filtered with FPKM cut-off 1 using build-in Perl script provided by Trinity. This resulted in 48474 transcripts for male M. quadrilineatus, 44409 transcripts for female M. quadrilineatus, 42815 transcripts for male D. maidis and 59131 transcripts for female D. maidis. These assemblies were used to validate the origin of RNA-seq data by assessing if reads aligning to leafhopper transcripts were present in RNA-seq data derived from plants exposed to the leafhoppers as opposed to those of plants that were not exposed to the leafhoppers.
|
10.1371/journal.ppat.1001238 | Interaction of c-Cbl with Myosin IIA Regulates Bleb Associated Macropinocytosis of Kaposi's Sarcoma-Associated Herpesvirus | KSHV is etiologically associated with Kaposi's sarcoma (KS), an angioproliferative endothelial cell malignancy. Macropinocytosis is the predominant mode of in vitro entry of KSHV into its natural target cells, human dermal microvascular endothelial (HMVEC-d) cells. Although macropinocytosis is known to be a major route of entry for many viruses, the molecule(s) involved in the recruitment and integration of signaling early during macropinosome formation is less well studied. Here we demonstrate that tyrosine phosphorylation of the adaptor protein c-Cbl is required for KSHV induced membrane blebbing and macropinocytosis. KSHV induced the tyrosine phosphorylation of c-Cbl as early as 1 min post-infection and was recruited to the sites of bleb formation. Infection also led to an increase in the interaction of c-Cbl with PI3-K p85 in a time dependent manner. c-Cbl shRNA decreased the formation of KSHV induced membrane blebs and macropinocytosis as well as virus entry. Immunoprecipitation of c-Cbl followed by mass spectrometry identified the interaction of c-Cbl with a novel molecular partner, non-muscle myosin heavy chain IIA (myosin IIA), in bleb associated macropinocytosis. Phosphorylated c-Cbl colocalized with phospho-myosin light chain II in the interior of blebs of infected cells and this interaction was abolished by c-Cbl shRNA. Studies with the myosin II inhibitor blebbistatin demonstrated that myosin IIA is a biologically significant component of the c-Cbl signaling pathway and c-Cbl plays a new role in the recruitment of myosin IIA to the blebs during KSHV infection. Myosin II associates with actin in KSHV induced blebs and the absence of actin and myosin ubiquitination in c-Cbl ShRNA cells suggested that c-Cbl is also responsible for the ubiquitination of these proteins in the infected cells. This is the first study demonstrating the role of c-Cbl in viral entry as well as macropinocytosis, and provides the evidence that a signaling complex containing c-Cbl and myosin IIA plays a crucial role in blebbing and macropinocytosis during viral infection and suggests that targeting c-Cbl could lead to a block in KSHV infection.
| KSHV is etiologically associated with Kaposi's sarcoma (KS), the most common AIDS related neoplasm. The first key step in KSHV infection is its initial contact with target cells and entry. While it is known that KSHV uses macropinocytosis for its infectious entry into its natural target cells, HMVEC-d cells, we know little about the molecule(s) involved in this event. Here, we show that the adaptor protein c-Cbl plays a major role in regulating bleb associated macropinocytosis of KSHV. The results demonstrate that c-Cbl protein functions as an adaptor for the myosin II hexameric complex in macropinocytic events. Knocking down c-Cbl by shRNA induces defects in myosin II dependent blebbing and KSHV entry, indicating that c-Cbl uses myosin II to coordinate signaling pathways, resulting in bleb formation and bleb retraction. This work provides a clear understanding of the role of c-Cbl in the recruitment and integration of signaling molecules around the macropinosome during virus infection, and identifies potential targets to intervene in KSHV infection.
| KSHV is etiologically associated with Kaposi's sarcoma (KS), the most common AIDS related malignancy, as well as with two lymphoproliferative diseases, primary effusion lymphoma (PEL) and multicentric Castleman's disease [1], [2]. KSHV infects a variety of target cells both in vivo and in vitro. Entry into the target cells is the most crucial step in the establishment of a successful infection for all viruses. KSHV utilizes different modes of endocytosis to enter different target cells in vitro [3]. For example, KSHV enters human foreskin fibroblasts (HFF) via clathrin mediated endocytosis and enters HMVEC-d cells via macropinocytosis [3], [4], [5]. During the early stages of infection of HMVEC-d cells, KSHV forms a multi-molecular complex with host cell heparan sulfate, integrins (α3β1, αVβ3 and αVβ5) and transporter protein xCT with the subsequent induction of overlapping signal cascades [3]. Our studies show that KSHV induces a complex set of signaling molecules that are involved in diverse biological functions to regulate the various aspects of KSHV endocytosis including internalization, trafficking in the cytoplasm and nuclear delivery [3]. KSHV activates FAK, Src, PI3-K, Rho-GTPases and cytoskeleton rearrangement which are all critical for entry of virus [6], [7], [8], [9]. KSHV also activates other downstream molecules such as PKC-ζ, MEK, ERK1/2 and NFkB which are essential for viral gene expression [6], [7], [8], [9].
The Cbl family of adaptor proteins include three mammalian isoforms, c-Cbl, Cbl-b and Cbl-c or Cbl-3 [10], [11]. Cbl proteins play important roles in signal transduction as negative regulators by mediating the ubiquitinilation and down-regulation of proteins while it acts as a positive regulator through their scaffold function in assembling signaling complexes [10], [11]. c-Cbl has been shown to bind to several molecules critical in signal transduction [10], [11]. Tyrosine phosphorylation of c-Cbl has been shown to be crucial for c-Cbl mediated adaptor functions in most circumstances [11], [12], [13]. However, the adaptor functions of c-Cbl and a c-Cbl mediated signaling pathway during virus infection has not been demonstrated.
Macropinocytosis provides a major route for the productive infection of many viruses including KSHV. Macropinocytosis is an actin dependent membrane associated process which involves recruitment and integration of several signaling molecules necessary for cytoskeletal rearrangement and membrane remodeling. However, there is little information about the molecules involved in the recruitment and integration of signaling during macropinosome formation. Even though c-Cbl has been shown to recruit and link different signaling molecules in a signaling pathway, a direct role for c-Cbl in the process of macropinocytosis has not been established yet.
Here we identified that c-Cbl is involved in KSHV entry and critical for triggering the macropinocytic event. Our data provide evidence that the interaction between c-Cbl and myosin IIA, a motor protein that binds to the proline rich domain of c-Cbl, regulates macropinocytosis of KSHV. This study on the functional organization of the c-Cbl and myosin IIA complex and its effect on viral entry provide an important insight into understanding the role of c-Cbl in virus infection.
HMVEC-d cells (CC-2543; Clonetics, Walkersville, Md) were grown in endothelial cell medium (EBM2; Cambrex, Walkersville, MD). Induction of the KSHV lytic cycle in BCBL-1 cells, supernatant collection, and virus purification procedures were described previously [14]. KSHV DNA was extracted from the virus, and the copies were quantitated by real-time DNA PCR using primers amplifying the KSHV ORF 73 gene as described previously [15].
A pool of lentivirus shRNA specific for human c-Cbl and non-specific control shRNA were purchased from Santa Cruz Biotechnology (Santa Cruz, CA). HMVEC-d cells were transduced with control lentivirus shRNA and c-Cbl lentivirus shRNA according to the manufacturer's instructions and selected by puromycin hydrochloride (10 µg ml−1; Santa Cruz Biotechnology).
The following antibodies were used: mouse anti-c-Cbl, mouse anti-phospho Cbl 700 (phosphorylated at Tyr700), and mouse anti-p85 (PI-3K) antibodies (BD Transduction Laboratories, San Diego, CA); anti-phospho MLC II, anti-phospho Cbl 731, anti-phospho Cbl 774 (phosphorylated at Tyr731 and Tyr774), isoform specific anti-myosin II heavy chain antibodies myosin IIA, IIB and IIC (Cell Signaling Technology, Danvers, MA); mouse anti-phospho tyrosine (4G10 clone; Millipore, Temecula, CA); mouse anti-tubulin, mouse anti-beta actin antibodies (Sigma, St Louis, MO); rabbit anti-lamin B (Abcam, Cambridge, MA); rabbit anti-HA (Zymed, Invitrogen, Carlsbad, CA); mouse anti-ubiquitin (P4D1), mouse ant-GFP, mouse anti-GST (Santa Cruz, CA); rabbit anti-gB and mouse anti-gpK8.1A antibodies were created in our laboratory[16], [17]; anti-goat, anti-rabbit and anti-mouse antibodies linked to horseradish peroxidase (KPL Inc., Gaithersburg, Md.); DAPI, rhodamine conjugated dextran, Alexa 594 or Alexa 488 conjugated phalloidin and anti-rabbit and anti-mouse secondary antibodies conjugated to Alexa 488, Alexa 594 (Invitrogen); protein A and G–Sepharose CL-4B beads (Amersham Pharmacia Biotech, Piscataway, NJ); blebbistatin, U0126 (Calbiochem, La Jolla, CA); TPA, LY294002 (Sigma).
Unless stated otherwise, cells were infected with KSHV at 10 DNA copies (multiplicity of infection [MOI]) per cell at 37°C. Entry was measured by infecting the cells with KSHV for 30 min. The cells were washed with HBSS to remove the unbound virus, treated with 0.25% trypsin-EDTA for 5 min at 37°C to remove the bound but non-internalized virus, and washed. Cells were recovered by centrifugation and total DNA was isolated from infected or uninfected cells using a DNeasy kit (QIAGEN, Valencia, CA) as described previously [15]. To calculate percent of inhibition of KSHV entry, internalized KSHV DNA was quantitated by amplification of the ORF73 gene by real-time DNA PCR [15]. The KSHV ORF73 gene cloned in the pGEM-T vector (Promega) was used for the external standard. The cycle threshold (Ct) values were used to generate the standard curve and to calculate the relative copy numbers of viral DNA in the samples. Percentage inhibition was calculated by considering the ORF73 copy numbers in untransduced cells as 100%.
Total RNA was prepared from infected or uninfected cells using an RNeasy kit (QIAGEN) as described previously [15]. To quantitate viral gene expression, isolated RNA was subjected to ORF73 and ORF50 RNA expression by real-time reverse transcription (RT)-PCR using gene specific real-time primers and specific TaqMan probes [15]. The relative copy numbers of the transcripts were calculated from the standard curve plotted using the Ct values for different dilutions of in vitro-transcribed transcripts. These values were normalized to each other using the values of the GAPDH control reactions. Percentage inhibition was calculated by considering ORF73 and ORF50 gene expression in untransduced cells as 100%.
Cells were lysed in RIPA buffer (15 mM NaCl, 1 mM MgCl2, 1 mM MnCl2, 2 mM CaCl2, 2 mM phenylmethylsulfonyl fluoride, and protease inhibitor mixture (Sigma)) and centrifuged at 12,000 rpm at 4°C for 15 min. Lysates were normalized to equal amounts of protein and the proteins were separated by 7.5–12.5% gradient SDS-PAGE, transferred to nitrocellulose and probed with the indicated primary antibodies. Detection was by incubation with species-specific HRP-conjugated secondary antibodies. Immunoreactive bands were visualized by enhanced chemiluminescence (Pierce, Rockford, IL) according to the manufacturer's instructions. The bands were scanned and quantitated using the FluorChem FC2 and Alpha-Imager Systems (Alpha Innotech Corporation, San Leonardo, CA).
Two hundred micrograms of cell lysates prepared as described in the above section were incubated for 2 h with immunoprecipitating antibody at 4°C, and the immune complexes were captured by protein A or G-Sepharose. The samples were tested by Western blot with specific primary and secondary antibodies.
HMVEC-d cells were infected with KSHV for different time points. The samples were resolved on an SDS-PAGE gel and the gel was stained with Coomassie blue. The bands of interest were excised, digested with trypsin, separated by reverse phase nano-chromatography and analyzed by mass spectrometry.
Immunofluorescence assay was performed using HMVEC-d cells seeded on 8 well chamber slides (Nalge Nunc International). Infected and uninfected cells were fixed with 3% paraformaldehyde for 15 min, permeabilized with 0.2% Triton X-100, and blocked with Image-iTFX signal enhancer (Invitrogen). The cells were then immunostained with primary antibodies against the specific proteins, followed by fluorescent dye-conjugated secondary antibodies. For colocalization with dextran and transferrin, cells were incubated with the fluid-phase marker dextran Texas Red (40 kD, 0.5 mg ml−1; Invitrogen) or Alexa 594 transferrin (35 µg ml−1; Invitrogen) at 37°C in the presence or absence of KSHV followed by immunostaining with the appropriate antibodies. Cells were imaged with a Nikon fluorescence microscope equipped with a Metamorph digital imaging system. DIC (Differential Interference Contrast) images were acquired with objectives equipped with DIC optics. For confocal analysis, the Olympus Fluoview 300 fluorescence confocal microscope was used for imaging, and analysis was performed using Fluoview software (Olympus, Melville, NY). All experiments were performed at least three times.
HMVEC-d cells, incubated with dextran Texas Red (0.5 mg ml−1, 40 kD; Invitrogen) and KSHV for 30 min, were washed twice in HBSS. To remove surface bound dextran, cells were treated with 0.25% trypsin-EDTA and the cells were harvested. Quantitative analysis of dextran uptake was determined by counting the number of cells stained positive for dextran under immunofluorescence microscope. At least 10 different microscopic fields of 50 cells each were counted for each experiment and the results displayed as percentage of dextran positive cells.
Flow cytometry analysis was used to quantify the uptake of dextran during KSHV internalization in control shRNA and c-Cbl shRNA transduced cells. Cells were incubated with 500 µg ml−1 FITC-dextran in the presence or absence of virus at 37°C for 30 min. The cells were washed, harvested using trypsin EDTA, fixed and analyzed by flow cytometry. Mean fluorescence intensity was determined using a Becton Dickinson FACS system and CellQuest software. Cells incubated with dextran alone were used as controls.
HeLa cells (ATCC CCL-2) were cultured in DMEM containing 10% fetal bovine serum. Wild type, mutants and deletion constructs of c-Cbl, c-Cbl C-terminal domain encompassing PRD (Cbl-C) and c-Cbl N-terminal domain (Cbl-N) constructs were generously provided by Dr. Hamid Band [18] (Eppley Institute for Cancer and Allied Diseases, University of Nebraska Medical Center). Cells were transiently transfected with wild type, mutants and deletion constructs. Transfection was performed using 5 µg of plasmid DNA, lipofectamine 2000 (Invitrogen), and Opti-MEM medium (Invitrogen) according to the manufacturer's instructions. After transfection, cells were cultured for 48 h. Cells were then serum starved for 4 h and stimulated with TPA (100 ng ml−1) at 37°C for 5 min. Lysis was performed in RIPA buffer plus protease inhibitors. The cell lysate was used for immunoprecipitation and immunoblotting.
E. coli BL21 (DE3) cells were transformed with pGEX4T.1 GST-Cbl C (Cbl residues 358–906) plasmids which encode regions encompassing the C-terminal PRD domain of c-Cbl and pGEX4T.1 GST-Cbl N which encodes the N-terminal region of c-Cbl. Expression of the GST-Cbl fusion proteins was induced with IPTG (isopropyl-D-1-thiogalactopyranoside) 1 mM for 3 h at 37°C. The bacterial lysates (500 µg) were incubated with glutathione-sepharose beads (GE Healthcare, U.K.) for 2 h at 4°C. The beads were washed with lysis buffer three times. 293T cells were transiently transfected with 2 µg pEGFP C3 myosin IIA plasmids (Addgene). After 48 h of transfection, cells were lysed in RIPA buffer and 500 µg of the lysates were incubated with the glutathione-Sepharose beads bound with the GST-Cbl fusion proteins. The beads and the bound proteins were collected by centrifugation, washed and the interaction of GST-Cbl with myosin IIA was analyzed by SDS-PAGE and Western blotting using anti-GFP antibody.
HMVEC-d cells infected with KSHV for 5 min were fixed and stained with DAPI. DIC images were acquired and the cells presenting blebs or no blebs were counted visually. At least 10 random microscopic fields per experiment were counted and expressed as a proportion of the total number of DAPI stained cells.
Infected and uninfected cells were washed three times with HBSS and lysed in homogenization buffer (250 mM sucrose, 20 mM HEPES, 10 mM KCL, 1 mM EDTA, 1 mM EGTA and protease inhibitors). The homogenate was subjected to centrifugation at 3,000 rpm for 5 min. Post-nuclear supernatant was centrifuged at 8,000 rpm for 5 min at 4°C. The supernatant was again centrifuged at 40,000 rpm for 1 h at 4°C, and the supernatant and the pellet were considered the cytosolic and the membrane fractions, respectively. The membrane pellet was solubilized using RIPA buffer and used for Western blot.
To determine whether c-Cbl and the c-Cbl mediated signaling pathway play roles in KSHV infection, we first examined the early tyrosine phosphorylation kinetics of c-Cbl in KSHV infected cells. HMVEC-d cells infected with KSHV induced rapid tyrosine phosphorylation of c-Cbl, which was detectable as early as 1 min post-infection (p.i.), reaching maximum levels at 5 min (4.1-fold), followed by a decrease which was constent for as much as 30 min p.i. (Figure 1a). To determine whether the phosphorylation of c-Cbl is specifically induced by KSHV, cells were infected with KSHV pre-incubated with heparin. Heparin is known to block the binding of KSHV to the target cells [19]. Compared to the untreated virus, heparin treated virus considerably reduced the phosphorylation of c-Cbl (Figure 1a) which demonstrated the specificity of KSHV induced c-Cbl phosphorylation. The efficient tyrosine phosphorylation of c-Cbl is suggestive of the possible involvement of a c-Cbl mediated signaling pathway in KSHV infection.
We next investigated the link between c-Cbl phosphorylation and other signaling molecules activated during KSHV infection. It is well documented that the interaction of KSHV glycoproteins with integrins and other cellular receptors activate FAK and the downstream molecules Src and PI3-K [7], [9], [14], [20], [21]. c-Cbl has been shown to form a complex with PI3-K p85 in the integrin mediated signaling pathway [12]. We therefore examined whether the association of PI3-K with c-Cbl occurred during KSHV infection. KSHV infection led to an increase in the interaction of c-Cbl with PI3-K p85 in a time dependent manner (Figure 1b). To verify that the c-Cbl-PI3-K interaction is specifically induced by virus, cells were infected with heparin treated virus which notably decreased the association of c-Cbl with PI3-K (Figure 1b). The c-Cbl-PI3-K association was further confirmed by confocal analysis (Figure 1c). Consistent with previous studies [13], our results demonstrated that activated c-Cbl leads to the association of c-Cbl with PI3-K.
Previous studies have shown that ERK1/2 is activated during KSHV infection and is a key signaling molecule implicated in viral gene expression [22]. To examine whether an ERK1/2 associated pathway is involved in c-Cbl mediated signaling, we investigated the association of ERK1/2 with c-Cbl in KSHV infected cells. No colocalization was observed between ERK1/2 and c-Cbl (Figure 1c) which suggested that the ERK associated pathway is not involved in c-Cbl mediated signaling in KSHV infected cells. Taken together, our data suggests that a signaling complex which contains c-Cbl and PI3-K but not ERK1/2 is involved in the integrin mediated signaling pathway of KSHV infection.
To further demonstrate the relationship between the interaction of c-Cb1 with PI3-K but not with ERK1/2, we studied the effect of PI3-K and ERK1/2 inhibitors in KSHV induced c-Cbl phosphorylation. HMVEC-d cells pretreated with the PI3-K inhibitor, LY294002, and the ERK1/2 inhibitor, U0126 were infected with KSHV for 10 min and the lysates were analyzed for c-Cbl phosphorylation. KSHV induced c-Cbl phosphorylation was abolished by the PI3-K inhibitor LY294002, whereas the ERK1/2 inhibitor U0126 did not show any inhibition on c-Cbl phosphorylation (Figure 1d). The failure of ERK1/2 inhibitor to abolish c-Cbl phosphorylation confirmed that ERK1/2 is not associated with c-Cbl induction in KSHV infected cells.
We next used c-Cbl lentivirus encoding shRNAs to knockdown c-Cbl activity in HMVEC-d cells (c-Cbl shRNA cells) to analyze the functions of c-Cbl in KSHV infected cells. The c-Cbl specific shRNA inhibited 90% of c-Cbl expression as detected by Western blotting with antibodies to c-Cbl (Figure S1a). Untransduced, control shRNA and c-Cbl shRNA transduced cells were infected with KSHV for 2 and 24 h, and viral gene expression was determined by real-time RT-PCR analysis. Compared with control cells, c-Cbl shRNA transduced cells showed about 60–70% inhibition of the latency associated ORF 73 gene (Figure 2a) and 70–80% inhibition of the lytic switch ORF 50 gene (Figure 2b) expression.
We next determined whether the inhibition of viral-gene expression by c-Cbl shRNA was due to a blockage at the entry stage of the virus. To determine c-Cbl's role in KSHV entry, internalization of viral DNA was determined by measuring viral ORF 73 DNA copy numbers by real-time DNA PCR. We observed ∼65% inhibition of KSHV entry in c-Cbl shRNA cells compared to control cells (Figure 2c). Internalized KSHV ORF 73 DNA copy numbers and ORF 50 and ORF 73 RNA copy numbers are shown as histograms in supplementary Figure S1. Taken together, these studies demonstrated that the decreased viral gene expression observed in c-Cbl shRNA cells was due to a decrease in the entry of KSHV. These results further suggested that a c-Cbl containing signaling complex may be crucial for the initiation of entry and for a productive infection.
In our earlier studies we have demonstrated that macropinocytosis is the major pathway of KSHV entry leading to a productive infection in HMVEC-d cells [4]. Since c-Cbl inhibited KSHV's entry, we theorized that c-Cbl might be playing a role in macropinocytosis and associated signaling events. Viruses such as vaccinia virus that use macropinocytosis as a mode of entry induce signaling molecules and cytoskeletal rearrangements in the form of blebs which ultimately retract and ingest viral particles [23], [24]. To determine whether blebs were involved in KSHV infection, we used DIC image analysis and observed the association of KSHV with blebs. As shown in Figure 3a, bleb formation and the association of individual blebs with KSHV was observed by 5 min p.i.. The DIC microscopic analysis of a single bleb for viral particles confirmed that the blebs, formed during viral infection, were associated with viral particles (Figure 3b). To further investigate whether KSHV infection induces blebbing, HMVEC-d cells were infected with KSHV for 2 and 5 min and then actin, a well known determinant of cell shape and blebbing, was stained with phalloidin [25]. Within 2 min of infection, membrane protrusions appeared along the cell surface which rapidly enlarged into well-formed blebs at 5 min (Figure S2). We also observed the association of viral particles at the bleb forming site as well as with retracting blebs (Figure S2). Unlike the well formed blebs, the retracting blebs were characterized by a thick actin cortex [26]. These results demonstrated that early during infection, KSHV induces actin reorganization and the subsequent formation of blebs that may be involved in its entry.
Since we observed that c-Cbl shRNA inhibited KSHV entry, which involves bleb formation, we hypothesized that c-Cbl and its phosphorylation might be involved in the dynamics of virus induced blebbing. To understand the function of c-Cbl in blebbing, we examined the localization of phosphorylated c-Cbl (p-Cbl) in KSHV infected cells. Confocal microscopy analysis showed that KSHV induced p-Cbl localized to blebs as early as 2 min p.i. (data not shown). Bleb formation, and its association with p-Cbl, was maximal at 5 min, and by 10 min blebs containing p-Cbl started internalizing and p-Cbl was mostly observed at the nuclear periphery by 15 min p.i. (Figure 3c). A similar pattern of localization was exhibited by p-Cbl and virus at the blebs as early as 5 min p.i., and accumulation of p-Cbl and virus around the nuclear periphery was observed at 15 min p.i. (Figure 3d). These results suggested that the recruitment of phosphorylated c-Cbl to the sites of bleb formation was involved in bleb associated entry of the virus.
To further explore the role of c-Cbl in bleb associated macropinocytosis, we performed a confocal immunofluorescence colocalization study between p-Cbl and the macropinocytosis marker dextran in infected cells. This analysis showed that dextran colocalized with p-Cbl at 10 min p.i. (Figure 4a) in the infected cells. Next, we performed a dextran uptake study since the uptake of dextran has been used as a biochemical marker of macropinocytosis. We incubated the cells with dextran in the presence or absence of virus for 30 min and then quantitated the level of uptake. As shown in Figure 4b and supplementary information Figure S3a, c-Cbl shRNA cells showed a drastic inhibition of dextran uptake compared to control shRNA cells infected with KSHV. This indicated that the uptake of dextran or macropinocytosis in KSHV infected cells was a c-Cbl dependent process.
The uptake of dextran and colocalization with KSHV in non-specific control shRNA and c-Cbl shRNA cells were confirmed by immunofluorescence colocalization and DIC analysis. In control shRNA cells infected with KSHV, intracellular KSHV was highly colocalized with dextran, whereas in c-Cbl shRNA cells infected with KSHV, most of the viral particles remained at the membrane periphery although minimal colocalization of KSHV with dextran was observed in some cells (Figure 4c and Figure S3b). Control shRNA cells incubated with KSHV and Alexa 594 transferrin, a marker for clathrin-mediated endocytosis, did not show any significant colocalization with KSHV (Figure 4c and Figure S3b) which demonstrated the specificity of macropinocytosis mediated entry in HMVEC-d cells [4]. These results were consistent with the results of the dextran uptake study, confirming that c-Cbl was critical for inducing the macropinocytic process that promoted the internalization of KSHV.
The uptake of dextran in control shRNA and c-Cbl shRNA cells was further quantified by FACS analysis. Cells were incubated with dextran in the presence or absence of virus for 30 min and the uptake was measured using flow cytometry. As shown in Figure 4d, compared to the control shRNA cells, c-Cbl shRNA cells showed a notable decrease in mean fluorescence intensity. These results are consistent with the results of the immunofluorescence analysis and thus confirmed the role of c-Cbl in KSHV induced macropinocytosis.
c-Cbl is a multi-domain protein that interacts with a number of signaling molecules and performs multiple functions [10], [11]. To decipher the molecular partners interacting with c-Cbl during KSHV infection, we used mass spectrometric analysis. HMVEC-d cells were infected with KSHV for 1, 5 and 10 min, lysed and the lysates were immunoprecipitated with anti-c-Cbl antibodies. Samples were separated by SDS-PAGE, followed by Coomassie blue staining and mass spectrometry analysis. Mass spectrometry identified several novel c-Cbl interacting proteins in the infected samples (Supplementary Table S1). The most prominent protein identified in the infected samples was myosin IIA which is one of three isoforms of the non-muscle myosin II family of proteins [27], [28]. The other novel interacting partners of c-Cbl in the infected cells included vimentin, HSP70, BiP protein, Rho GEF and a solute carrier anion exchanger (Table S1).
To confirm the mass spectrometry data, uninfected and KSHV infected cell lysates were immunoprecipitated with anti-c-Cbl antibody and blotted for the three isoforms of the non-muscle myosin II family, IIA, IIB and IIC, with isoform specific antibodies. Our results confirmed that c-Cbl interacts with myosin IIA in the lysates of infected cells, whereas the other isoforms did not show any interaction with c-Cbl (Figure 5a).
To elucidate the functional domain of c-Cbl involved in myosin IIA interaction, a series of truncated and mutant constructs of c-Cbl with HA epitope tags were used (Figure S4). Since HMVEC-d cells are not easily transfectable, we used HeLa cells for this study. HeLa cells were transfected with vector alone, Cbl wild-type, Cbl-tyrosine kinase binding domain (TKB) mutant, RING domain mutant, and two truncation mutants (Cbl-Δ357 and Cbl-Δ421). As TPA (phorbol ester) has been shown to induce membrane blebbing [29], we used TPA induced HeLa cells to analyze the interaction of over-expressed c-Cbl with endogenous myosin IIA. Transfection of the Cbl-TKB mutant and RING mutant induced the interaction of c-Cbl with myosin IIA similar to full length wild-type Cbl. The truncated versions Cbl-Δ357 and Cbl-Δ421 lacking a C-terminal proline rich domain (PRD) decreased the interaction with myosin IIA considerably (Figure 5b). Expression of all constructs determined by Western blotting with HA revealed comparable levels of protein (Figure 5b). Taken together, these results indicated that the C-terminal region encompassing the PRD of c-Cbl was sufficient for association with myosin IIA.
To further confirm that the C-terminal PRD of c-Cbl interacts with myosin, an in vitro binding assay was performed using bacterially expressed GST fusion proteins of c-Cbl C-terminal (Cbl-C, encompassing PRD) and N-terminal (Cbl-N) domains. GST Cbl-C and Cbl-N proteins adsorbed on glutathione sepharose beads were incubated with 293T cell lysates expressing GFP-tagged myosin IIA. The interaction between GFP-myosin IIA and GST-Cbl was analyzed by Western blotting with anti-GFP antibody. Our results demonstrated that myosin IIA predominantly interacted with Cbl-C domains compared to Cbl-N (Figure 5c).
The interaction of myosin IIA with c-Cbl suggested that their association could be playing a role in blebbing and macropinocytosis of KSHV. To investigate this, we used blebbistatin, a specific inhibitor of myosin II ATPase activity that has been shown to inhibit myosin II induced blebbing [30], [31] and macropinocytosis [23]. As shown in Figure 6a, we observed a dose dependent inhibition of KSHV internalization in 25 µM (∼35%) and 50 µM (∼60%) concentrations of blebbistatin indicating that the entry process was dependent on myosin II activity. As reported previously [4], chlorpromazine, an inhibitor of clathrin dependent endocytosis, did not show any notable decrease in entry of KSHV (Figure 6a). These findings suggested that c-Cbl associated myosin IIA was involved in bleb mediated macropinocytosis of KSHV.
To determine whether blebbistatin treatment affects other internalization pathways, we investigated the effect of blebbistatin on clathrin-mediated internalization, a major and well characterized endocytic pathway of eukaryotic cells. To study this, untreated or blebbistatin treated HMVEC-d cells were induced with FBS in the presence of Alexa 594 labelled transferrin or Texas Red labelled dextran. The endocytic uptake of transferrin and dextran were then analyzed using immunofluorescence. As indicated in Figure 6b and d, blebbistatin strongly inhibited the uptake of dextran, whereas the uptake of transferrin was unaffected (Figure 6c and e). This demonstrated that blebbistatin specifically inhibits macropinocytosis but not clathrin mediated endocytosis pathways.
The above studies demonstrated that the c-Cbl interacting partner myosin IIA is a biologically significant component of the c-Cbl signaling pathway. We then explored the role of c-Cbl in myosin II induced blebbing in KSHV infected cells. If c-Cbl is an upstream molecule of myosin IIA, the loss of function of c-Cbl should prevent the formation of myosin II mediated blebs in c-Cbl shRNA cells. To test this hypothesis, control shRNA and c-Cbl shRNA transduced HMVEC-d cells were infected with KSHV and the percentage of cells with blebs was quantitated. As expected, in c-Cbl shRNA transduced cells, the blebs were considerably reduced compared to control shRNA cells (Figure 7a and b). This suggested that c-Cbl and associated myosin IIA molecules were linked to induce membrane blebbing in HMVEC-d cells. Our results indicated that c-Cbl plays an upstream role in the regulation of bleb formation which occurs as a result of myosin II induced cortical contractility [25], [26].
To further demonstrate that c-Cbl is upstream to myosin IIA, we infected blebbistatin treated cells with KSHV and the membrane localization of c-Cbl was observed by immunofluorescence. As shown in Figure 7c, blebbistatin did not inhibit the localization of c-Cbl to the plasma membrane, whereas it prevented the formation of blebs in the infected cells. This suggested that myosin IIA was downstream to c-Cbl and was not involved in the localization of c-Cbl to the plasma membrane.
A subclass of myosins, the class II myosins are hexameric motor proteins composed of two identical heavy chains (MHC), and two pairs of light chains (MLC). It has been well accepted that phosphorylation of the myosin light chain II is a major determinant of force generation and actomyosin dynamics during apoptotic membrane blebbing [32], [33]. Hence, we examined phosphorylation of myosin light chain II (p-MLC II) during KSHV infection. Compared to the uninfected cells, KSHV infection results in rapid and strong phosphorylation of MLC II with maximal phosphorylation at 10 min p.i. (5.8-fold increase) and decreased thereafter (Figure 7d). The specificity of virus induced MLC II phosphorylation was shown using heparin treated virus which did not induce MLC II phosphorylation (Figure 7d). Since light chain phosphorylation has been shown to regulate blebbing [32], [33], our results suggested that KSHV induced MLC II may be participating in the induction of blebbing during infection.
During virus induced and apoptotic membrane blebbing, the signaling molecules associated with cytoskeletal function are recruited to the blebs [23], [26]. It has been demonstrated that the recruitment of functional myosin II heavy and light chain complexes drive the process of bleb retraction [26]; however, it is not clear how individual myosin II molecules are recruited to the blebs. It is possible that c-Cbl interaction with myosin IIA leads to recruitment of the complex to the blebs. Therefore, we infected control shRNA and c-Cbl shRNA cells with KSHV and tested the association of c-Cbl with myosin II in the blebs.
Punctate staining of p-Cbl and p-MLC II was observed in the interior of blebs with a predominant colocalization between p-Cbl and p-MLC II in control shRNA infected cells (Figure 8a) suggesting that phosphorylated Cbl recruits individual myosin molecules to the blebs. In contrast, c-Cbl shRNA cells infected with KSHV did not show bleb formation and the recruitment and localization of myosin II to the bleb membrane (Figure 8a).
To further confirm the membrane localization of myosin IIA and c-Cbl to the blebs, membrane fractions from control shRNA infected cells and c-Cbl shRNA infected cells were isolated and analyzed by Western blotting. Compared to the uninfected cells, control shRNA cells infected with KSHV showed a 3.2 and 2.9-fold increase in membrane localization of myosin IIA and c-Cbl, respectively, whereas in c-Cbl shRNA-KSHV cells, membrane localization of myosin IIA and c-Cbl was almost completely absent (Figure 8b). This suggested that a decrease in membrane localization of myosin IIA in c-Cbl shRNA cells may be caused by a deficiency in the association of c-Cbl with myosin IIA.
Myosin II molecules recruited to the membrane blebs form contractile foci in association with actin under the bleb membrane which is critical for bleb retraction [26]. Therefore, we asked whether a similar kind of association occurs between p-MLC II and actin in control shRNA-KSHV and c-Cbl shRNA-KSHV cells. As has been previously reported [26], we observed the association of actin and p-MLC II in the membrane blebs in control shRNA KSHV infected cells (Figure 9a). Bleb formation and the association of actin with p-MLC II in the blebs were not seen in c-Cbl shRNA-KSHV cells (Figure 9a). The association of p-MLC II with actin, which is a known interacting partner of c-Cbl [10], coupled with the detection of an association with c-Cbl in the infected cells (Table S1) suggested that actin, myosin IIA and c-Cbl are part of a signaling complex which might be essential for the formation of blebs and bleb retraction.
To examine whether the interaction of c-Cbl with myosin IIA is actin dependent, we investigated the interaction of c-Cbl with myosin IIA in cells treated with actin inhibitor cytochalasin D. The interaction between c-Cbl and myosin IIA was then examined by coimmunoprecipitation and Western blot analysis. As shown in Figure 9b, cytochalasin D did not inhibit the interaction of myosin IIA with c-Cbl. The interaction of myosin IIA with c-Cbl in cytochalasin D treated cells suggested that actin is not essential for the initial association of myosin IIA with c-Cbl in KSHV infected cellular environment.
c-Cbl is an E3 ubiquitin ligase, which has been shown to be involved in poly or monoubiquitination of a number of proteins [34], [35], [36]. To further analyze the functional significance of the interaction between c-Cbl, myosin and actin, the role of c-Cbl in ubiquitination of actin and myosin was analyzed. Both myosin and actin ubiquitination were determined in control shRNA and c-Cbl shRNA cells infected with KSHV. We observed multiple bands of ubiquitinated myosin and actin probably indicating monoubiquitination on multiple sites and not polyubiquitination in the infected cells (Figure 9c and d). c-Cbl shRNA abolished the KSHV induced ubiquitination of actin and myosin suggesting that it was mediated by c-Cbl. This indicated that upon KSHV infection, c-Cbl binds to actin and myosin which in turn is responsible for ubiquitination.
Several viruses utilize macropinocytosis to gain access to target cells [24]. Macropinocytosis is strictly an actin driven process which includes the formation of membrane ruffles, lamellipodia and blebs [24]. Blebbing is a phenomenon which is mainly observed during cellular processes such as embryogenesis, cytokinesis and apoptotic cell death [25]. Bleb associated macropinocytosis provides a mechanism of virus entry into host target cells and has been shown to be an efficient tactic of the virus to enter the target cells by mimicking the apoptotic cells [23].
Macropinocytosis has been observed as a major route of entry of KSHV into HMVEC-d cells, the natural target cells of infection [3], [4]. During KSHV infection, the interaction of KSHV glycoproteins with integrins and other cellular receptors activate host cell signaling molecules FAK, Src, PI3-K, RhoA GTPase and are all recruited to the entry site [3]. The inhibition of any of these proteins significantly reduces the entry of virus suggesting that KSHV exploits preexisting host cell signaling machinery for a successful infection [3]. Our current study shows that c-Cbl is also required for efficient macropinocytic uptake, suggesting that the previously observed signaling molecules are linked via c-Cbl to perform their downstream functions. Since bleb associated macropinocytosis is an actomyosin dependent process, c-Cbl and its interaction with myosin playing a role in macropinocytosis further strengthens the possibility that c-Cbl is a critical molecule involved in linking the signaling molecules in the virus induced macropinocytic process.
Based on the strong evidences presented here, we propose that efficient bleb mediated macropinocytic uptake enables a productive infection of KSHV to occur in cells supporting a signaling cascade that contains the c-Cbl-myosin IIA complex, and that a defect in c-Cbl-myosin IIA association results in lowered macropinocytic uptake, entry and infection. The simultaneous decrease in macropinocytic uptake and blebbing by blebbistatin treatment and c-Cbl silencing with shRNA strongly suggests that bleb associated macropinocytosis is the predominant pathway of KSHV infection in HMVEC-d cells. The defect in macropinocytosis in c-Cbl shRNA cells could be due to a defect in linking myosin II molecules to the membrane associated events which is necessary for blebbing and bleb mediated macropinocytosis [24], [25], [26].
Although the molecular details of bleb associated macropinosome formation are yet to be uncovered, our study demonstrates that during KSHV infection, the interaction of c-Cbl with myosin IIA leads to bleb formation and the recruitment of myosin IIA into the blebs, where the myosin IIA molecules could be interacting with actin to accelerate actomyosin contraction and bleb retraction [26]. Retracting blebs form macropinosomes along with the viral particles close to the blebs (Figure 10) [24]. Myosins provide the ATP-dependent force to generate the movement required for the process of bleb retraction [26]. Myosin IIA is the major isoform of myosin II implicated in membrane associated functions such as the maintenance of cell shape and movement [37], [38]. The functional ending of ubiquitination is related to the type of ubiquitin chains added to the substrate protein [39]. Monoubiquitination promotes internalization of cell surface receptors and subsequent lysosomal degradation [40], whereas polyubiquitinated proteins are targeted for proteasomal degradation [41]. Whether the complex events associated with the monoubiquitination of actin and myosin by c-Cbl could be related to an increase in the activity of actomyosin contraction or directing subcellular compartmentalization during KSHV infection and macropinocytosis remains to be studied in detail. Further studies are required to understand the molecular aspects of the interaction between c-Cbl and myosin IIA and their role in subsequent stages of bleb mediated macropinocytosis. Further studies are also required to validate the specificity and the functional significance of other identified c-Cbl interacting proteins, their association with c-Cbl during infection and whether c-Cbl induces the ubiquitination of cell surface molecules recognized by KSHV.
In conclusion, our results provide for the first time clear evidence demonstrating that c-Cbl, and the interaction between c-Cbl and myosin IIA, is critical for triggering bleb mediated macropinocytic events during KSHV entry into target cells (Figure 10). This study also provides the first evidence that c-Cbl and a c-Cbl mediated signaling pathway as well as ubiquitination play roles in viral infection and that c-Cbl function as an adaptor protein for PI3-K and other KSHV induced signaling events. This also identifies c-Cbl as a potential target to intervene in KSHV infection.
|
10.1371/journal.pntd.0004123 | Spatial Analysis of Anthropogenic Landscape Disturbance and Buruli Ulcer Disease in Benin | Land use and land cover (LULC) change is one anthropogenic disturbance linked to infectious disease emergence. Current research has focused largely on wildlife and vector-borne zoonotic diseases, neglecting to investigate landscape disturbance and environmental bacterial infections. One example is Buruli ulcer (BU) disease, a necrotizing skin disease caused by the environmental pathogen Mycobacterium ulcerans (MU). Empirical and anecdotal observations have linked BU incidence to landscape disturbance, but potential relationships have not been quantified as they relate to land cover configurations.
A landscape ecological approach utilizing Bayesian hierarchical models with spatial random effects was used to test study hypotheses that land cover configurations indicative of anthropogenic disturbance were related to Buruli ulcer (BU) disease in southern Benin, and that a spatial structure existed for drivers of BU case distribution in the region. A final objective was to generate a continuous, risk map across the study region. Results suggested that villages surrounded by naturally shaped, or undisturbed rather than disturbed, wetland patches at a distance within 1200m were at a higher risk for BU, and study outcomes supported the hypothesis that a spatial structure exists for the drivers behind BU risk in the region. The risk surface corresponded to known BU endemicity in Benin and identified moderate risk areas within the boundary of Togo.
This study was a first attempt to link land cover configurations representative of anthropogenic disturbances to BU prevalence. Study results identified several significant variables, including the presence of natural wetland areas, warranting future investigations into these factors at additional spatial and temporal scales. A major contribution of this study included the incorporation of a spatial modeling component that predicted BU rates to new locations without strong knowledge of environmental factors contributing to disease distribution.
| Changes in land and use and land cover can impact ecosystems in unexpected ways, including changes in habitat suitability for environmental pathogens. Several studies have investigated the impacts of human disturbance to the landscape and changes in the composition of vector, host, and reservoir species in an altered area, but few studies have linked these disturbances to environmental pathogens. Buruli ulcer disease is a neglected tropical disease caused by the environmental pathogen Mycobacterium ulcerans. This study investigated land cover patterns surrounding villages in southern Benin to identify relationships between disturbed landscapes and disease prevalence. The authors were also interested in whether the drivers of disease prevalence had a spatial structure that could provide clues to environmental characteristics important to disease presence. Results suggested that villages surrounded by natural, or undisturbed wetlands had higher BU rates, and there was, in fact, a spatial structure to the pattern of disease prevalence. The authors used these outcomes to create the first continuous, BU risk map across southern Benin and Togo, even though the exact drivers of disease transmission remain unknown. Predicting potential risk adequately provides valuable information toward prevention, while helping to target public health resources more efficiently.
| Land use and land cover (LULC) change at multiple spatial and temporal scales is one anthropogenic disturbance linked to infectious disease emergence [1]. Anthropogenic activities with major impacts on LULC are land degradation, including agriculture intensification and water projects, urbanization, and deforestation [2]. These activities can lead to ecological edge effects that promote disease emergence [3]. Further, these activities generate new pathways through which humans can interact with previously undisturbed environments, resulting in closer proximities to potential vectors, reservoirs, and isolated pathogens [3–6].
Advances in geographic information systems (GIS), remote sensing technologies, spatial statistical methods and computational capacities facilitate observation and quantification of anthropogenic landscape disturbances, providing the tools necessary to link landscape characteristics to disease incidence and to predict disease transmission risk across landscapes and through time [7]. While current research has focused largely on wildlife and vector-borne zoonotic disease emergence [8,9], exploring linkages between anthropogenically-disturbed landscapes and human bacterial infections has been given less attention, even though recent findings suggest that disturbances may contribute to the spatial distribution of environmental bacteria that pose a human health risk [10]. Quantification of landscape patterns related to bacterial disease emergence is a central component to mapping transmission risk because ecological drivers behind these pathogens are often poorly understood.
One example of this phenomenon is Buruli ulcer (BU) disease, a necrotizing skin disease caused by the environmental pathogen Mycobacterium ulcerans (MU) [11,12]. MU produces mycolactone, an immunosuppressive agent responsible for ulcer formation. Although the ecological drivers behind MU growth in the environment remain a mystery, empirical and anecdotal linkages exist between dramatic increases in BU cases since the 1980s and anthropogenic landscape changes [13]. These disturbances include, but are not limited to, deforestation, habitat fragmentation, aquatic ecosystem disturbances from dam construction and agriculture irrigation, changing farming practices, and mining activities [12]. Although BU is not transferred between persons, the mode or modes of transmission has not been determined, and no vaccine exists [14]. Therefore, identifying landscape patterns linked to BU incidence will provide a powerful tool for surveillance and prevention while affording opportunities to learn more about the ecology of the disease system.
Several past BU studies investigated landscape features related to BU disease. Research in the Amansie West District of Ghana identified a correlation between disease incidence and proximity to soils enriched with arsenic in low-lying farmlands [15]. An additional study in the Amansie West district identified a relationship between mean arsenic levels in soil and the spatial distribution of BU cases and between increased proximities to gold mining sites and the spatial distribution of cases within the district [16]. Mantey et al. [17] investigated linkages between potential surface runoff and BU incidence in the Amansie West and Upper Denkyira West Districts of Ghana, finding that BU cases correlated positively with a low to moderate potential for surface runoff and that higher numbers of cases correlated with lower potential maximum soil water retention values. A country-wide study in Côte d’Ivoire found an association between BU incidence and closer proximities to irrigated rice fields and to artificial dams [18], while a study by Marion et al. [19] postulated an association between the construction of a large dam in the Bankim district of Cameroon and the geographic expansion of BU cases.
An additional study in the Ankonolinga health district of Cameroon investigated fine-scale patterns of BU incidence within the area, finding closer proximities to the Nyong River, disturbed forest area, and cultivated wetlands to be significant risk factors for the disease [20]. A country-wide study in Benin determined that villages at low elevations within drainage basins, with variable wetness patterns, and surrounded by forest had higher BU risk [21]. A recent study in Benin suggested an inverse relationship between BU incidence and elevation [22], while Williamson et al. (2012) found incidences to be lowest at elevations <25m or at higher elevations of 90-100m. Wagner et al. [23] determined that BU incidence increased for villages surrounded by agriculture land and decreased with surrounding urban land use at broad scales (e.g., 20 km radius around a village).
While offering important insight into linkages between landscape features and BU incidence, these studies focused on landscape composition rather than its configuration, neglecting landscape patterns as overall indicators of disturbance. Past studies using landscape metrics suggested that land cover patches with more uniform shapes (e.g. corresponding to roadways or managed forest patches) are indicators of potential anthropogenic disturbance [24]. In contrast, land cover patches with more complex shapes, such as natural wetlands, often represent undisturbed landscapes [25]. In addition, higher numbers of patches can suggest potential habitat fragmentation [26].
Although several previous BU studies have utilized spatially-referenced data, only a fraction of studies investigating landscape and environmental variables incorporated spatial statistical modeling approaches into their analyses [16,20,21,23]. Tobler’s first law of geography states that “everything is related to everything else, but near things are more related than distant things” [27], a verbal statement of the concept of spatial autocorrelation. If a non-spatial regression model is used to investigate BU-environmental relationships, which is often the case, there is a high propensity to violate basic model assumptions, such as independent and identically distributed model residuals [28]. Models that account explicitly for spatial autocorrelation offer several advantages over non-spatial models, including avoidance of incorrect inferences regarding regression coefficient significance, or Type I errors [29–31]. Further, addition of components that explicitly accommodate residual spatial dependence to the class of models considered, often improves model fit and predictive performance [32].
The purpose of this study was to quantify effects of potential anthropogenic landscape disturbances on BU incidence in Benin, West Africa, using landscape metric calculations and spatial modeling approaches. We hypothesize (1) that land cover patches with configurations indicative of anthropogenic landscape disturbance surround villages with higher BU rates in southern Benin and (2) that a spatial pattern exists for drivers of BU incidence in Benin. Our final objective was to create a BU risk surface across southern Benin and southern Togo.
The southern portions of Benin and Togo, West Africa comprise our study area, 6.30°N—8.17°N and 0.84°E to 2.48°E (Fig 1). Four major rivers flow through the study area, including the Couffu, Ouémé and Zou rivers in Benin, and the Mono River that delineates the southern border between Togo and Benin. BU is endemic in Benin and Togo, but incidence data from Togo were incomplete. Therefore, this study used BU case observations and corresponding environmental data in Benin to identify significant drivers, and then predicts BU risk across southern Benin and Togo.
Benin’s landscape consists primarily of woodland and shrub savannas, intermixed with cultivated areas and fallow fields, with semi-deciduous forests present in the southern region of the country [33]. Deforestation and environmental degradation are ongoing problems in the country, and the landscape continues to change rapidly. Food-crop and cotton cultivation expanded by 265% and 79% between 1986 and 1997, while firewood extraction and charcoal production contributed to a 30,000 ha/yr deforestation rate [33].
The Programme National de Lutte contre la Lèpre et l’ulcère de Buruli (PNLLUB) in Benin provided a subset of BU positive and BU negative villages in 2004 and 2005 for this analysis. A village was identified as BU positive if at least one case occurred there in 2004 or 2005. These data were obtained from World Health Organization (WHO) BU02 standardized forms, created using a community-based reporting system developed by the WHO to facilitate case reporting across geographic regions [34]. BU case counts, population, and latitude and longitude coordinates were provided for each village in the data set. A total of 292 villages, 183 positive and 109 negative, fell within the study area; 558 individual cases occurred, ranging between 1–29 cases per village (Fig 2). Data deposited in the Dryad repository [http://dx.doi.org/10.5061/dryad.j512f21p][35].
A land use and land cover (LULC) classification for southern Benin and Togo was performed using 30m resolution Landsat ETM+ imagery from 13 December 2000. The imagery was geometrically rectified and projected (UTM Zone 31N) before performing an unsupervised classification [36]; assignment of land cover classes followed Anderson’s Level I classification scheme [37], with the exception of four mixed classes. The final classification consisted of 10 classes total, including three mixed agriculture classes and a general mixed-use class that included pixels classified as belonging to >2 categories. Execution of a 5x5 statistical majority filter helped to eliminate classification noise [38]. Visual interpretation methods were used to delineate land cover classes [39–41] across the study area for a lack of ground truth data. Aggregated land cover classes were intended to help mitigate classification error. Data deposited in the Dryad repository [http://dx.doi.org/10.5061/dryad.j512f21p][35].
Forest, wetland, and mixed agriculture/forest classes were selected for this analysis based on potential linkages identified from empirical and anecdotal scenarios. Forest and wetland classes exhibited unique spectral signatures, with wetland signatures showing a lower spectral reflectance curve, particularly in the Near Infra-Red (NIR) range of the electromagnetic spectrum, while forest signatures demonstrated a sharper increase in the NIR range of the electromagnetic spectrum [42]. We noted larger uncertainty in separating the mixed agriculture and forest class from other non-forested classes because of the spatial resolution of our data.
Quantifying landscape patterns within concentric polygons is a common approach in multiscalar landscape analyses [43,44]. Concentric polygons with radii of 800m, 1200m, 1600m, and 2000m intervals from village centers acted as buffers within which land cover data were obtained for pattern analysis. These radii were chosen to characterize the landscape within distances traversed regularly by village residents, while maintaining extents large enough to quantify landscape patterns.
Initial calculation of a suite of landscape metrics quantified study class configurations within the concentric polygons using an 8-neighbor rule [45]. Of particular interest were landscape metrics that identify landscape disturbances linked to anthropogenic activities, for example fragmentation or uniform land cover patch shapes (refer to FragStats manual for greater detail [46]).
Collinearity among landscape metrics is common (46). In a regression context, collinearity, or correlation, among predictor variables can cause problems with inference. Hence, applied regression texts suggest avoiding correlation among predictors beyond 0.7, see, e.g., [47]. Here, we took a conservative cut-off of 0.6 to reduce issues arising from collinearity. Potential predictor pairs were assessed using Pearson's correlation coefficient and Spearman's rank order correlation. Both correlation metrics yielded comparable results. Using this criterion and exploratory analysis using the models detailed in subsequent sections, the predictor variables we consider in the subsequent analyses are: 1) shape index mean, 2) percent land cover adjacency, and 3) landscape shape index. Fig 3 provides an illustration of land cover configurations which give rise to high and low values of each metric. The following metric calculation descriptions were derived from the FragStats documentation directly. Shape index mean (SHAPE_MN) characterized patch shape complexity and was calculated as
SHAPE=pijminpij
where pij = perimeter of patch ij in terms of number of cell surfaces and min pij = minimum perimeter of patch ij in terms of number of cell surfaces [46]. Values closer to 1.0 indicate more uniformly-shaped land cover patches with complexity increasing as values increase.
Percent land cover adjacency (PLADJ) measured aggregation between patches of a similar class and is calculated as
PLADJ=(gij∑k=1mgik)
where gij = the number of like adjacencies (joins) between pixels of patch type (class) i based on a double-count method, and gik = the number of adjacencies (joins) between pixels of patch types (classes) i and k, also based on a double-count method [46]. High PLADJ values represent a more aggregated land cover class, while low PLADJ values represent a more fragmented land cover class.
The landscape shape index (LSI) is another land cover patch aggregation measurement. LSI was calculated using
LSI=eiminei
where ei = perimeter of class i in terms of number of cell surfaces, which includes all landscape boundary and background edge segments involving class i, and min ei = minimum total length of edge (or perimeter) of class i in terms of number of cell surfaces [46].
The initial non-spatial generalized linear model (GLM) considered the n = 292 villages and coinciding predictor variables indexed by location, S = {si,…, s}, where each s is a vector recording the longitude and latitude in UTM Zone 31N projection. At generic location s, the response variable y(s) was taken as the number of BU cases reported there. At generic location s, the response variable y(s) was taken as the number of BU cases reported in each village of population size N(s). We assumed y(s) followed a Binomial distribution with N(s) observations in each village and probability p(η(s)) of BU incidence, i.e., y(s) ∼ Binomial(N(s),p(η(s))). A logistic link function was used to model the probability of BU, i.e, p(η(s)) = exp(η(s)) / (1+η(s)). For the non-spatial models, the regression term η(s) = x(s)’β where the vector x(s) comprises an intercept, village-specific predictor variables (i.e., SHAPE ME, PLADJ, and LSI), and associated regression parameters β. We used a stepwise approach to explore combinations of predictor variables within candidate models over the 800m, 1200m, 1600m, and 2000m distances from village centers and each of the three land cover classes. Candidate models consisted of variables with a correlation coefficient < 0.6. Model outcomes were compared using a Deviance Information Criterion approach (DIC) [48], and individual variables within candidate models were eliminated in a stepwise process. Final models consisted of those with the lowest DIC values.
The non-spatial binomial GLMs assumed that model residuals were independent and identically distributed across the study domain [28], i.e., the included predictor variables capture any spatial patterns in the probability of BU cases. While this approach is adequate in the absence of spatial autocorrelation in model residuals, this assumption will often be unrealistic given the spatial structure of the observations [49]. As noted above, violations of this model assumption can result in misleading inference regarding the importance of predictors and subsequently erroneous predictions at locations where BU rates were not observed. To mitigate this issue, the non-spatial model was modified to allow for a spatially-varying intercept, via the addition of spatially structured random effects. Mapping of the random effects are often useful for identifying missing or unobserved predictor variables [50], in addition to accommodating any lurking residual spatial dependence, improving inference about the importance of the predictor variables, and increasing predictive performance.
The spatially varying intercept is included in the model by adding a spatial random effect, w(s), to η(s), i.e. η(s) = x(s)’β +w(s) where x(s) and β were defined previously. The spatial random effect was specified as a mean zero Gaussian Process with covariance function C(s1, s2;θ) = σ2ρ(s1, s2; ϕ), with s1 and s2 representing any two arbitrary locations and σ2 is the process variance. We assumed an exponential spatial correlation function, ρ(·,·; ϕ) = exp(-ϕ||s1 − s2 ||), where ϕ controls the rate of spatial decay and ||s1 − s2 || is the Euclidean distance between the locations [50]. See [51–53] for examples of spatial random effects applications in disease ecology research.
The model specification is completed by assigning prior distributions to all parameters. As customary, the regression coefficients β were assigned a multivariate Gaussian prior, β~N(μ, Σ), with μ set to a vector of zeros, and the matrix Σ specified with diagonal elements equal to 100 and off-diagonal elements to zero. Exploratory data analysis (EDA) using other diagonal variance values, i.e., 1000 and 10000, were assessed for the non-spatial and spatial models with results showing negligible impact of the prior choice on width or centering of posterior distributions.
The spatial variance component σ2 was assigned an inverse gamma prior IG(a, b), with the shape hyperparameter, a, set to 2 and the scale, b, parameter varied from 0.1 to 5 to assess the influence of the prior specification. Note, that following the IG definition in [54], with α = 2 the distribution mean is b and has infinite variance. EDA using the various specifications of b showed little influence on the posterior inference. Results presented in subsequent sections used an IG(2,1) prior for σ2. The process correlation parameter, ϕ, was assigned an informative prior (e.g., uniform over a finite range) with support across the maximum intersite distance among any two locations.
Model parameter distributions were estimated using Markov chain Monte Carlo (MCMC) methods employing an adaptive Metropolis (AM) algorithm with a 43% acceptance rate [55]. All models were generated using the spGLM function in spBayes R package [56]. MCMC chain starting values were obtained from the non-spatial models and subsequent posterior inference was based on 3 chains at 100,000 iterations measured over the four spatial extents (e.g. 800m, 1200m, 1600m, and 2000m).
Chain convergence was diagnosed using the Gelman-Rubin potential scale reduction factor [57]. Convergence for the non-spatial and spatial model parameters was typically achieved within 5,000–10,000 MCMC iterations. Parameter inference and subsequent predictions were based on post-convergence MCMC samples (i.e., burn-in was set at 10,000 iterations for all model parameters).
Models were compared using the popular Deviance Information Criterion (DIC). Like Akaike Information Criterion and similar model fit criteria used to compare fit among non-Bayesian models, lower DIC values indicate better performance [48]. To more fully assess predictive performance among the candidate models, the data sets were divided randomly into two subsets– 90% of the observations were used to fit the candidate models and 10% used to verify subsequent BU rate predictions. For each model, root mean squared prediction error (RMSPE) was calculated based on the 10% holdout set’s predicted and observed values, with lower values indicating improved expected performance [55].
We generated a total of 1029 new prediction locations in a gridded pattern at 5 km intervals, where BU rates were unobserved (Fig 4). Gaps among these location points represent areas where data were lacking and derivation of predictor variables was not possible for the “best” model class and distance interval. The predictive BU risk surface was created following Finley et al. (2008) and implemented in the spBayes R package. Here, the use of a Bayesian modeling paradigm is particularly advantageous because it provides access to the entire posterior predictive distribution at each new location [50], and hence, we can map any quantity of interest. For our purposes, these quantities include a map of grid cell specific posterior predictive distribution mean and variance.
The majority of BU positive communities in Benin were located in the southern region of the country (Fig 2). The departments of Zou and Couffo had the highest number of BU positive communities in our data set, with 51 and 59 locations. Ouémé, Mono, and Atlantique followed with 38, 19, and 14 BU positive communities, with Plateau and Collines having 8 and 4 positive locations from our sample data set.
The spatial GLMs produced lower DIC values than the non-spatial GLMs, suggesting that residual spatial dependence was present and that the spatial models achieved a better fit for every combination of predictor variables (Table 1). Comparison of regression coefficients between the spatial and non-spatial models shows that the non-spatial GLMs tend to overestimate the importance of the predictor variables—likely resulting from violation of the model assumptions related to non-correlated residuals (Table 2). For example, Forest was found to be significant at every spatial scale in the non-spatial model, while only 1200m Forest was significant in the spatial model. In addition, the percent land cover adjacency variable corresponding to the 2000m wetland class model reversed signs, again attributed to a violation of the basic model assumptions.
Results from the assessment of candidate model predictive performance using RMSPE are provided in Tables 3 and 4 illustrates actual versus predicted BU rates at 18 sample site locations. Here, the spatial 1200m wetland model, which included SHAPE_MN and LSI as predictor variables achieved the lowest RMSPE. Parameter estimates for this model are given in Table 5. Here, and as noted in Table 2, the regression coefficients associated with SHAPE_MN and LSI were both positive and statistically different from zero, which suggests that more complexly-shaped and aggregated wetland land cover patches surround villages with higher BU rates. The small value of the spatial decay parameter and hence the long effective spatial range (defined as the distance at which spatial correlation drops to below 0.05) suggests relatively strong residual dependence, even after accounting for the predictor variables.
We examined plots of actual BU rates versus the 1200m wetland model predictions to assess spatial patterns in the model fit (Fig 5). There appeared to be relatively strong agreement in most areas, but BU prevalence rates were under-predicted along the Ouémé River. A plot of the spatial random effects highlighted areas where spatially structured predictor variables were missing from the model, influencing the spatial distribution of the rates and promoted model fit, with higher values near locations with case presence (Fig 6). An interpolated surface of the mean of the posterior predictive distribution at new locations based on the 1200m wetland model represents BU risk across the study region (Fig 7).
This study is the first to investigate land cover configurations indicative of anthropogenic disturbances in relation to BU incidence. Model results provided insight into BU incidence in West Africa, while demonstrating the value of spatial modeling approaches in disease ecology investigations. Lower DIC values corresponding to spatial models support the hypothesis that spatial structure exists for drivers of BU incidence in the region. Importantly, comparisons between non-spatial and spatial variable significance demonstrated the potential for inaccurate estimates to occur when using non-spatial models to address ecological problems. The inclusion of the spatial random effects accounted for missing predictor variables and provided substantial improvements in predictions of BU rates at unsampled locations (Table 2).
Our model results did not support the hypothesis that more uniformly-shaped and fragmented land-cover patches, indicative of potential anthropogenic disturbance, surround villages with higher BU rates. Spatial model results suggested general trends toward more aggregated and complexly-shaped, or natural, patches surrounding villages with higher BU rates, although inconsistencies in variable significance occurred across distance intervals. Fig 8 illustrates landscapes from our classification that represent high BU prevalence rates and no BU prevalence rates that correspond with our model results.
Mixed agriculture/forest class results suggested that more aggregated and complexly-shaped patches surrounded villages with higher BU rates at 1200m–2000m distances. While the configuration did not correspond to those recognized as representing anthropogenically-disturbed environments, the presence of a mixed agriculture and forest class suggests an inherent intrusion of agriculture into undisturbed forested landscapes. Persons tending fields in these mixed landscapes could have elevated opportunities to encounter edge effects and additional disturbances known to contribute to pathogen proliferation in environmental diseases [3]. Investigation into the role of ecological succession and specific disturbance stages in BU incidence may yield important information regarding high-risk areas.
Several plausible scenarios may explain model outcomes. The first considers the regions in which most landscape ecology studies identifying shapes indicative of anthropogenic disturbance took place. These studies occurred largely in developed regions of Europe and the United States, rather than in developing areas, such as Benin. Further investigation into quantification of land cover attributes in developing versus developed regions may provide useful information for future studies. Additionally, the models constructed in this study may not have been inclusive of land cover classes that could play an important role in BU ecology. Further investigation into additional land cover classes may yield insightful results.
Another consideration is the scale at which the investigation took place. This study utilized 30m resolution satellite imagery with a 5x5 statistical majority filter to characterize the landscape for statistical analyses. Drivers promoting BU proliferation may operate at coarser or finer scales. Additionally, several landscape disturbances associated with anthropogenic activities may not have been identified using medium-resolution imagery. One example is the presence of cultivated rice paddies within wetland land cover. Uniformly-shaped plots indicative of cultivation were easily discernible in higher-resolution 4m Ikonos imagery, but these disturbances were not visible in the Landsat imagery; thus, some areas that we treated as natural wetlands may have been anthropogenically impacted as well. Therefore, additional studies investigating land cover configurations at multiple scales could benefit BU research in the region.
Identification of spatial autocorrelation effects in model residuals supported the hypothesis that a spatial structure exists for drivers of BU incidence. The effective range of the “best” model suggested that spatial autocorrelation exists within 28.6km—41.2km of village centers and contributes to shaping the distribution of BU cases in the region. Although the spatial random effects do not indicate what is driving the presence of BU, the spatial patterns provide clues about missing variables and improved the predictive accuracy of the BU risk surface. Visual inspection of the spatial random effects reveals a pattern that corresponds roughly with a geologic and soil type transition from sedimentary basins with ferralitic and hydromorphic soils in southern Benin to a crystalline basement with fersialitic soils in the central region, warranting further exploration of these phenomena in relation to the spatial distribution of BU incidence in this region (IMPETUS Atlas Benin).
The purpose of the risk surface was to act as a preliminary guide to identifying regions where BU cases would be most likely to occur. Model results identified high-risk areas in known endemic regions near the town of Tandji and between the Zou and Ouémé Rivers in Benin, demonstrating consistency with case incidence reports [58]. Results also suggested high rates along the Ouémé River in the east and along the Couffu River in the western area of the country. Lower predicted rates along the southern coast could be the result of brackish waters that negatively impact environmental conditions suitable to the MU pathogen, or may be attributed to the hypothesis that residents in this region have better access to pumped water due to greater urbanization, which may lower BU risk [21]. Further, sandy soil, which is quick to drain and does not promote the accumulation of standing water, may hinder opportunities for MU growth in this region. Additionally, model results suggested lower rates between the Couffu and the Ouémé and Zou rivers, where higher elevations separate two known endemic regions, reiterating results from previous studies [21,22,59].
More interesting were predicted BU risk areas within the boundary of Togo. Model results suggested moderate risk east of the Mono River, but north of where the river delineates the international border. This region houses the Nangbeto Dam, behind which a large reservoir resides. Although this region is located within a wetland system, few BU cases were reported following construction of the Nangbeto dam in 1987 (R. Christian Johnson, personal communication, September 20, 2015). One hypothesis is that a reduction in cases occurred because of controlled fluctuations in water levels, reducing seasonal flooding impacts in the region, similar to the situation in Ghana, where few BU cases are known to occur in the Lake Volta region [11]. While the environmental conditions may be comparable to those identified as high risk environments in Benin, the model could not account for anthropogenic interference with river flow and therefore, predicted this area as having a moderate BU risk.
Generally, predicted BU rates in Togo exhibited moderate values with less spatial variability than those in Benin. This trend may result from the increased distance from observed locations. As distance from locations with known values exceeds the effective range, predicted rates have a tendency to move toward a mean predictive value. Although this phenomenon may have impacted predicted rates within Togo, identification of regions at moderate risk for BU occurrence is a first step in bridging knowledge gaps stemming from data disparities in the region.
The role of anthropogenic ecosystem disturbances in the emergence of environmental bacterial infections is poorly understood. This study was a first attempt to link land cover configurations representative of anthropogenic disturbances to the environmental bacterial infection Buruli ulcer disease. Although results did not suggest a positive relationship between land cover patch configurations representative of anthropogenic disturbances and BU rates, study results identified several significant variables, including the presence of natural wetland areas, warranting future investigations into these factors at additional spatial and temporal scales.
Beyond the novel exploratory analysis outlined above, a major contribution of this study included the incorporation of a modeling component that partitioned the spatial structure of missing variables, providing a structure from which to predict BU rates to new locations without strong knowledge of environmental factors contributing to disease distribution at the beginning of this analysis. The resulting continuous BU risk surface demonstrates the potential to develop and to target surveillance efforts using spatial modeling approaches. The ability to predict potential risk adequately to locations where few data are available provided a first step toward prevention, while creating a tool from which a more systematic and controlled site selection process may be used to target future environmental sampling research. Continued acquisition of accurate, georeferenced case data along with georeferenced pathogen data will provide the best opportunity for robust empirical studies of relationships between ecological factors, anthropogenic activities, and BU transmission.
|
10.1371/journal.pmed.1002674 | Prediction of myopia development among Chinese school-aged children using refraction data from electronic medical records: A retrospective, multicentre machine learning study | Electronic medical records provide large-scale real-world clinical data for use in developing clinical decision systems. However, sophisticated methodology and analytical skills are required to handle the large-scale datasets necessary for the optimisation of prediction accuracy. Myopia is a common cause of vision loss. Current approaches to control myopia progression are effective but have significant side effects. Therefore, identifying those at greatest risk who should undergo targeted therapy is of great clinical importance. The objective of this study was to apply big data and machine learning technology to develop an algorithm that can predict the onset of high myopia, at specific future time points, among Chinese school-aged children.
Real-world clinical refraction data were derived from electronic medical record systems in 8 ophthalmic centres from January 1, 2005, to December 30, 2015. The variables of age, spherical equivalent (SE), and annual progression rate were used to develop an algorithm to predict SE and onset of high myopia (SE ≤ −6.0 dioptres) up to 10 years in the future. Random forest machine learning was used for algorithm training and validation. Electronic medical records from the Zhongshan Ophthalmic Centre (a major tertiary ophthalmic centre in China) were used as the training set. Ten-fold cross-validation and out-of-bag (OOB) methods were applied for internal validation. The remaining 7 independent datasets were used for external validation. Two population-based datasets, which had no participant overlap with the ophthalmic-centre-based datasets, were used for multi-resource validation testing. The main outcomes and measures were the area under the curve (AUC) values for predicting the onset of high myopia over 10 years and the presence of high myopia at 18 years of age. In total, 687,063 multiple visit records (≥3 records) of 129,242 individuals in the ophthalmic-centre-based electronic medical record databases and 17,113 follow-up records of 3,215 participants in population-based cohorts were included in the analysis. Our algorithm accurately predicted the presence of high myopia in internal validation (the AUC ranged from 0.903 to 0.986 for 3 years, 0.875 to 0.901 for 5 years, and 0.852 to 0.888 for 8 years), external validation (the AUC ranged from 0.874 to 0.976 for 3 years, 0.847 to 0.921 for 5 years, and 0.802 to 0.886 for 8 years), and multi-resource testing (the AUC ranged from 0.752 to 0.869 for 4 years). With respect to the prediction of high myopia development by 18 years of age, as a surrogate of high myopia in adulthood, the algorithm provided clinically acceptable accuracy over 3 years (the AUC ranged from 0.940 to 0.985), 5 years (the AUC ranged from 0.856 to 0.901), and even 8 years (the AUC ranged from 0.801 to 0.837). Meanwhile, our algorithm achieved clinically acceptable prediction of the actual refraction values at future time points, which is supported by the regressive performance and calibration curves. Although the algorithm achieved balanced and robust performance, concerns about the compromised quality of real-world clinical data and over-fitting issues should be cautiously considered.
To our knowledge, this study, for the first time, used large-scale data collected from electronic health records to demonstrate the contribution of big data and machine learning approaches to improved prediction of myopia prognosis in Chinese school-aged children. This work provides evidence for transforming clinical practice, health policy-making, and precise individualised interventions regarding the practical control of school-aged myopia.
| Myopia has reached epidemic levels among young adults in East and Southeast Asia, affecting an estimated 80%–90% of high school graduates, with approximately 20% of them having high myopia. Various interventions, including atropine eyedrops and orthokeratology, have been proposed to control myopia progression; however, these approaches confer significant side effects. Identifying those at greatest risk who should undergo targeted therapy is the most important clinical challenge faced by ophthalmologists and optometrists.
Electronic medical records provide large-scale real-world clinical data for use in developing clinical decision systems. Taking school-aged myopia, the most prevalent eye disease, as an example, it would be of great value to use ophthalmic-centre-based electronic medical records to develop a big-data-driven clinical prediction algorithm based on machine learning algorithms.
This study analysed 687,063 longitudinal electronic medical records from the largest ophthalmic centres in China and developed and validated individualised prediction models for myopia prediction based on machine learning techniques.
Our model predicted spherical equivalent and onset of high myopia at 18 years of age at a clinically acceptable accuracy and as early as 8 years in advance.
The algorithm, which was trained and validated using a large real-world dataset, was able to predict the presence of high myopia with clinically acceptable accuracy among Chinese school-aged populations.
Large-scale, long-term electronic medical records and machine learning algorithms provide unique opportunities for the development of prediction models for progressive diseases, such as myopia in school-aged children.
Our findings have great potential to change current approaches used to manage school myopia by paediatric and general ophthalmologists as well as general practitioners and optometrists, who are often the first point of care.
| Myopia, the most common visual impairment in children, has increased markedly in Chinese school-aged children in recent years. This “myopia boom” is a significant international public concern, impacting study performance and daily life [1]. The risk of children developing high myopia has become a great concern among parents [2], with thousands of students seeking care at optometric and ophthalmic clinics annually in China. This creates an enormous burden for the healthcare system but provides an unprecedented opportunity to collect large-scale real-world clinical data that are unified and reliable.
Big data available from service providers contain valuable “signals” for authentic disease progression and prognosis; however, the analysis of these data is challenging because such data are often contaminated by various types of “noise”, given that the data are not collected in a controlled research setting [3]. Machine learning offers a ubiquitous and indispensable method to solve these complexities of data noise and heterogeneity, having the capacity to combine enormous numbers of predictors in a non-linear and highly interactive way [4].
This study is a data-and-algorithm-driven analysis of more than half a million optometry records and data derived from long-term population-based cohort studies in China. The goal was to build a prediction algorithm based on machine learning techniques to uncover the key determinants of high myopia and to predict, as early and as accurately as possible, the development of high myopia in adulthood. The performance of the algorithm was validated using multi-source datasets from independent ophthalmic centres and population-based research cohorts. The results provide evidence for health policy-making regarding the practical control of school-age myopia and precise individual interventions.
A summary of study procedures is presented in Fig 1. Eight ophthalmic centres were included in the study, including Zhongshan Ophthalmic Centre (ZOC), the Haizhu Optometry Department (HZD), the Huangpu Optometry Department (HPD), the Panyu Optometry Department (PYD), the Dongguan Guangming Ophthalmic Hospital (DGC), the Optometry Centre in Huizhou City (HZC), the Haikou Longhua Optometry Department (LHD), and the Xiuying Optometry Department in Haikou City (HKC). This study also included 2 datasets collected from population-based cohort studies: the Guangzhou Outdoor Activity Longitudinal Trial (GOAL) [5] and the Refractive Error Longitudinal Study (RELS). These 8 ophthalmic centres and 2 cohorts from South China collectively composed a representative medical big data sample for children of Chinese ethnicity. This sample could be generalisable to Chinese children living in Hong Kong, Taiwan, and Singapore, where myopia is similarly a common public health problem in children. The geographical locations and a detailed description of the study population are presented in Fig 2, S1 Table, and S1 Text.
The study adhered to the tenets of the Declaration of Helsinki, and approval for the study protocol was obtained from the Institutional Review Board/Ethics Committee of Sun Yat-sen University (Guangzhou, China). All of the datasets used throughout the study were deidentified prior to transfer to the study investigators.
We extracted data from electronic medical record systems collected between January 1, 2005, and December 30, 2015, at the optometry service of 8 participating ophthalmic centres. To focus on the school-aged population, only individuals aged from 6 to 20 years at the initial examination and with ≥3 visits at ≥1-year intervals were included in the current analysis.
Predictors included age at examination, spherical equivalent (SE), and annual progression rate. Cycloplegic refraction was performed according to a standard protocol in each centre. The right eye was arbitrarily chosen to represent a specific individual.
Using these predictors, we aimed to develop an algorithm to predict SE and presence of high myopia in the subsequent 10 years (with each year as a predictive time point). The presence of high myopia was defined as a SE ≤ −6.0 dioptres.
Electronic medical records from ZOC were used as the training dataset, and 10-fold cross-validation and out-of-bag (OOB) validation methods [6] were applied for internal validation (details are provided in S2 Text). Meanwhile, a methodological comparison of random forest and other conventional algorithms (generalised estimating equation [7] and mixed-effects model [8]; details are provided in S3 Text) was performed using the average performance of the cross-validation. A complete algorithm was trained on the entire ZOC dataset prior to external validation (variable contributions in S1 Fig).
The refraction data records from the other 7 centres were used for external validation. All individuals from the 7 centres with available refraction records at 18 years of age and with at least 2 visits (≥1-year interval) were included. These records were used to explore the accuracy of prediction at a given time before 18 years of age. Two population-based longitudinal cohorts were used for the multi-resource test.
Random forest is an ensemble learning method that operates by constructing a multitude of decision trees at training time and outputting the class that is the mode of the classes (classification) or mean prediction (regression) of the individual trees. Here, we employed the random forest algorithm for the development of the prediction algorithm, which was established in the BrainWave machine learning module [9]. The R randomForest package, which implements Breiman’s classic algorithm, was used to fit the random forest model [10].
Each decision tree in the random forest was built using a bootstrap sample with replacement from the original data. This bootstrap aggregation and random feature selection helped reduce the variance of the algorithm and avoided over-fitting. Consequently, in the random forest algorithm, cross-validation is performed internally, which can be just as effective as using a separate test set to estimate the generalisation error of the training set. Moreover, the random forest algorithm can be used to evaluate the variables in a dataset and to provide a graphical display to assess the importance of each variable.
The 2 random forest parameters, mTry (i.e., the number of input variables randomly chosen at each split) and nTree (i.e., the number of trees to grow for each forest), were set to 2 (square root of 5 features) and 500, respectively. In each tree, each feature received a variable importance (VIMP) score, which can be used to rank and select relatively important features.
Regarding the regression analysis, the most widely used VIMP score of a feature is the average percent increase in the OOB mean square error (MSE) as a result of randomly permuting the OOB feature values [11]. The MSE is the mean of the squared regression residuals, and the VIMP score of a feature indicates its overall predictive ability for the regression. Regarding the classification analysis, the error rate is the proportion of misclassified samples of the total number of samples, and the VIMP score of a feature indicates its overall predictive ability for the classification.
Three evaluation metrics—the coefficient of determination (R2), the root mean square error (RMSE), and the mean absolute error (MAE)—were used to assess the performance of the regression algorithm in predicting a targeted SE [12]. R2 can be expressed as
R2=1−MSE/Var(y)
where MSE is the same as noted above and Var(y) is the variance of the actual value. The RMSE is the square root of the MSE, which penalises large errors but has the same units as the original response variable being predicted; thus, its magnitude is more easily interpreted. The MAE measures the forecast accuracy by averaging the absolute values of the residuals. The MAE is expressed in the same units as the original response variable and provides an average size of the “miss”, regardless of the direction. This variable can be expressed as
MAE=1n∑i=1n|yi−y^i|
where yi is the actual value, and y^i is the predicted value. These 3 evaluation metrics were calculated for the different predicted target times of each algorithm.
For classification performance, the receiver operating characteristic (ROC) curves and area under the curve (AUC) values were calculated as a comprehensive evaluation. All analyses were performed using R statistical software version 3.2.4 [13].
A description of the study population is displayed in Table 1. In total, 687,063 longitudinal electronic medical records of 129,242 individuals from 8 ophthalmic centres and 17,113 follow-up records for 3,215 participants in population-based cohorts were included in the analysis. A total of 517,949 records from ZOC were used as the training set (the follow-up duration ranged from 2 to 11 years, mean ± SD 4.6 ± 1.9 years). The datasets of the remaining 7 centres (169,114 records; the follow-up duration ranged from 2 to 11 years, mean ± SD 5.2 ± 2.1 years) were used for external validation, and the records from the 2 population-based cohorts (17,113 records; the follow-up duration ranged from 3 to 5 years, mean ± SD 4.1 ± 1.2 years) were used for multi-resource validation testing.
For the comparative analysis, the random forest algorithm outperformed the generalised estimating equation and the mixed-effects model in the detection of high myopia (S2 Fig). Therefore, all subsequent analyses were conducted based solely on the random forest algorithm. For classification, AUC values more than 0.9 indicated excellent performance, and values from 0.8 to 0.9 indicated good performance; MAE within ±0.75 dioptres was considered clinically acceptable accuracy (i.e., clinically acceptable prediction) based on the measurement variations of refraction [14].
As presented in Fig 3, our algorithm provided high-precision predictions of high myopia in the cross-validation (the AUC ranged from 0.903 to 0.958 for 3 years, 0.886 to 0.889 for 5 years, and 0.862 to 0.888 for 8 years) and OOB tests (the AUC ranged from 0.934 to 0.986 for 3 years, 0.875 to 0.901 for 5 years, and 0.852 to 0.874 for 8 years). Meanwhile, our algorithm achieved clinically acceptable prediction of the refraction value at each time point (year) after baseline assessment (the MAE ranged from 0.253 to 0.395 for 3 years, 0.394 to 0.496 for 5 years, and 0.503 to 0.799 for 8 years). The regressive performance and calibration curves of the algorithm are presented in Table 2 and S3 Fig, respectively. These calibration results also supported that our algorithm can predict the actual refraction values at time points over 10 years.
The performance of the algorithm in the external validation is presented in Fig 4. Our algorithm achieved stable performance for high myopia detection in the DGC (the AUC ranged from 0.768 to 0.969 for 10 years), the HZD (the AUC ranged from 0.773 to 0.968 for 10 years), the PYD (the AUC ranged from 0.854 to 0.951 for 5 years), the HZC (the AUC ranged from 0.822 to 0.941 for 6 years), the HPD (the AUC ranged from 0.802 to 0.976 for 8 years), the HKC (the AUC ranged from 0.897 to 0.929 for 3 years), and the LHD (the AUC ranged from 0.888 to 0.916 for 2 years). Clinically acceptable prediction of the refraction value was achieved at the majority of the time points examined (the MAE ranged from 0.201 to 0.494 for 3 years, 0.354 to 0.731 for 5 years, and 0.508 to 0.879 for 8 years).
With respect to predicting the presence of high myopia (Table 3), our algorithm provided clinically acceptable prediction over 3 years (the AUC ranged from 0.940 to 0.985), 5 years (the AUC ranged from 0.856 to 0.901), and even 8 years (the AUC ranged from 0.801 to 0.837).
In the multi-resource test (Fig 5), our algorithm presented stable high myopia detection in GOAL (the AUC ranged from 0.784 to 0.869 for 3 years) and RELS (the AUC ranged from 0.752 to 0.845 for 4 years). A clinically acceptable prediction of refraction value was achieved at all time points examined (the MAE ranged from 0.314 to 0.562 for 4 years).
This study, to our knowledge for the first time, demonstrates the utilisation of large-scale electronic medical record data to generate a random forest algorithm for predicting disease prognosis, which, in our analysis, was the risk of developing high myopia in adulthood. Furthermore, this algorithm exhibited high accuracy in a predicting future trait, i.e., the dioptre value at 18 years of age. Our data suggest that this prediction can be performed as early as 8 years prior to an individual turning 18 years old.
Identifying “severe myopia” in younger children is of major clinical importance but poses a significant challenge. The severity of myopia is often estimated as the degree of SE, with an SE of −6.00 dioptres chosen as the cutoff to define high myopia. High myopia carries a much greater risk of developing other ocular complications, including retinal detachment, glaucoma, and pathological myopia [15,16]. Given that myopia is in the development phase during childhood, it is difficult to choose a specific SE cutoff to define “severe myopia” among children.
A few studies have identified children at a greater risk of progressive myopia [17–19]; however, none to our knowledge has attempted to predict actual SE or risk of high myopia in adulthood. The “risk classification” in previous studies was often inferred from the analysis of short-term longitudinal data, or a control group in the instance of intervention randomised trials. As such, the available data are only generalisable among children who meet the inclusion and exclusion criteria of the specific studies [20]. Due to pragmatic feasibility, most myopia control trials can only run for up to 3 years, and, similarly, longitudinal studies on myopia are often shorter than 5 years. Real-world electronic medical record data from established optometry services in tertiary ophthalmic centres are of considerable advantage in terms of the size of the dataset and the length of follow-up.
A prediction is meaningful only when it is accurate and early enough to provide an added clinical benefit. As demonstrated by our results, the accuracy of prediction is reduced when the targeted prediction time increases. However, interestingly, in our analysis, the accuracy, indicated by the AUC, remained high (0.80–0.90) for up to 8 years in both the internal and external validation. Furthermore, the 95% predicted dioptre of refraction was within 0.5 to 0.8 dioptres of the true value at 8 years. Such accurate “long-term” predictions are critically important given that current treatments for myopia control, including low-dose atropine [21] and orthokeratology lenses [22], are effective but often have potential side effects and therefore must be utilised effectively. In addition, accurate early prediction and timely treatment of myopia in its mild stages are important to maximise the treatment benefits.
Methodologically, the random forest algorithm, which is based on random selection and a combination of predictors [9], achieved superior performance in the current analysis compared to conventional methodologies (i.e., a generalised estimating equation and a mixed-effects model). The added value of the random forest algorithm presented a gradual enlargement after 3 years, indicating that myopia development became increasingly non-linear in long-term range. This advantage can be further appreciated in analyses that require the inclusion of more complex potential predictors in the model.
There are some limitations that must be addressed with the development of a prediction model using large-scale real-world clinical data. First, clinical data collected in real-world settings are often subject to bias, with compromised quality. For instance, although a standard clinical protocol was followed, refraction measurement was performed by a number of different optometrists in the present analysis. Despite this, one may argue that this “noise” can be de-emphasised by a stronger “signal” when the sample size is large enough. This effect has been highlighted in previous myopia genetic studies; for example, the genes identified by the Consortium for Refractive Error and Myopia (CREAM) study [23] (in which the refractive error was measured in every participant) were similar to those reported in the 23andMe study [24] (self-reported refractive error). Second, an algorithm developed from a training set may be subject to over-fitting, whereby the correlation or prediction is spurious [25]. This appears not to be the case in the present study, with our algorithm showing balanced contributions of all involved predictors and robust performance when evaluated in independent, external datasets.
In summary, to our knowledge, this study, for the first time, used large-scale data collected from the electronic health records from the largest ophthalmic centres in China to demonstrate the contribution of big data to the better prediction of disease prognosis. In the context of school-age myopia, the most prevalent eye disease in the Chinese population, our study demonstrates that machine learning prediction algorithms further translate the benefit of big data research into clinical practice. The performance of our predictive algorithm is promising, with large sample sizes and diversified data resources. This work proposes a novel direction for the use of medical big data mining to transform clinical practice and guide health policy-making and precise individualised interventions.
|
10.1371/journal.pgen.1004314 | Whole Exome Re-Sequencing Implicates CCDC38 and Cilia Structure and Function in Resistance to Smoking Related Airflow Obstruction | Chronic obstructive pulmonary disease (COPD) is a leading cause of global morbidity and mortality and, whilst smoking remains the single most important risk factor, COPD risk is heritable. Of 26 independent genomic regions showing association with lung function in genome-wide association studies, eleven have been reported to show association with airflow obstruction. Although the main risk factor for COPD is smoking, some individuals are observed to have a high forced expired volume in 1 second (FEV1) despite many years of heavy smoking. We hypothesised that these “resistant smokers” may harbour variants which protect against lung function decline caused by smoking and provide insight into the genetic determinants of lung health. We undertook whole exome re-sequencing of 100 heavy smokers who had healthy lung function given their age, sex, height and smoking history and applied three complementary approaches to explore the genetic architecture of smoking resistance. Firstly, we identified novel functional variants in the “resistant smokers” and looked for enrichment of these novel variants within biological pathways. Secondly, we undertook association testing of all exonic variants individually with two independent control sets. Thirdly, we undertook gene-based association testing of all exonic variants. Our strongest signal of association with smoking resistance for a non-synonymous SNP was for rs10859974 (P = 2.34×10−4) in CCDC38, a gene which has previously been reported to show association with FEV1/FVC, and we demonstrate moderate expression of CCDC38 in bronchial epithelial cells. We identified an enrichment of novel putatively functional variants in genes related to cilia structure and function in resistant smokers. Ciliary function abnormalities are known to be associated with both smoking and reduced mucociliary clearance in patients with COPD. We suggest that genetic influences on the development or function of cilia in the bronchial epithelium may affect growth of cilia or the extent of damage caused by tobacco smoke.
| Very large genome-wide association studies in general population cohorts have successfully identified at least 26 genes or gene regions associated with lung function and a number of these also show association with chronic obstructive pulmonary disease (COPD). However, these findings explain a small proportion of the heritability of lung function. Although the main risk factor for COPD is smoking, some individuals have normal or good lung function despite many years of heavy smoking. We hypothesised that studying these individuals might tell us more about the genetics of lung health. Re-sequencing of exomes, where all of the variation in the protein-coding portion of the genome can be measured, is a recent approach for the study of low frequency and rare variants. We undertook re-sequencing of the exomes of “resistant smokers” and used publicly available exome data for comparisons. Our findings implicate CCDC38, a gene which has previously shown association with lung function in the general population, and genes involved in cilia structure and lung function as having a role in resistance to smoking.
| Chronic obstructive pulmonary disease (COPD) is a leading cause of global morbidity and mortality [1] and whilst smoking remains the single most important risk factor, it is also clear that COPD risk is heritable [2]. The genetics underlying COPD are still not fully understood although genome-wide association studies have identified 26 genomic regions showing robust association with lung function [3]–[6] and, of these, 11 have also now shown association with airflow obstruction [7]–[9]. However, the proportion of the variance accounted for by the 26 common genetic variants representing these regions remains modest (∼7.5% for the ratio of forced expired volume in 1 second (FEV1) to forced vital capacity (FVC)) [5].
Although over a quarter of the population with a significant smoking history go on to develop COPD [10], some individuals are observed to have preserved lung function as measured by a normal or high FEV1 despite many years of heavy smoking. We hypothesised that these “resistant smokers” may harbour rare variants with large effect sizes which protect against lung function decline caused by smoking. Identification of these variants, and the genes that harbour them, could provide further insight into the mechanisms underlying airflow obstruction.
We undertook whole exome re-sequencing of 100 heavy smokers (>20 pack years of smoking) who had healthy lung function when age, sex, height and amount smoked were taken into account. We employed 3 complementary approaches to investigate the genetic architecture of the resistant smoker genotype (Figure 1). Firstly, we screened these 100 “resistant smokers” for novel rare variants (i.e. not previously identified and deposited in a public database) with a putatively functional effect on protein product and tested for enrichment of these novel variants in functionally related genes and pathways. Secondly, using a comparision with two independent control sets with exome re-sequencing data, we looked for signals of association with the resistant smoker phenotype for individual variants (including variants of all minor allele frequencies). Thirdly, we looked for association of the resistant smoker phenotype with the combined effects of multiple rare and common variants within genes.
We found the strongest evidence of association with resistance to smoking for a non-synonymous variant in CCDC38, a gene encoding a coiled-coil domain protein with a role in motor activity, previously identified as showing an association with lung function. We also show evidence of cytoplasmic expression of CCDC38 in bronchial columnar epithelial cells. In addition, we found evidence for an enrichment of novel rare functional variants in resistant smokers in gene pathways related to cilia structure and function. Given that abnormalities of ciliary function are already known to play a role in reduced mucociliary clearance in COPD sufferers and smokers, these data suggest that genetic factors may play a significant role in determining the ciliary response of the airway to inhaled tobacco smoke.
100 individuals from the Gedling [11], [12] and Nottingham Smokers cohorts with good lung function (FEV1/FVC>0.7 and % predicted FEV1>80%) when age, sex, height and smoking history (>20 pack years) were taken into account were selected as “resistant smoker” cases. Characteristics of the 100 resistant smoker case samples are shown in Table S1A and Figure S1. Exome re-sequencing and alignment was undertaken as described in the methods.
Two independent control sets were used; the robustness of findings using the primary control set (n = 166) were further assessed using a secondary control set (n = 230).
We searched for novel variants among the resistant smokers, i.e. genetic variants which were not observed in either control set and which were not documented in public databases. Bioinformatic tools allow for scoring of likely functional impact, including whether a variant is likely to be “deleterious”; here we use the term “putatively functional” since some variants which have a deleterious effect on the function of a given gene may result in a protective phenotype. A total of 24,098 variants which were not already in databases of known variants or within segmental duplications were identified with high confidence using two independent calling algorithms. A total of 6587 coding SNPs were scored using CAROL (including non-synonymous, loss/gain of stop codon, synonymous and splice site/UTR variants) and 1722 were predicted as being putatively functional (CAROL score>0.98) and were within 1533 genes. 16 of these 1533 genes each contained three novel putatively functional variants (Table S2) (no gene contained more than three such variants). GBF1 contained three novel putatively functional variants of which one, chr10:104117872, was identified in two case samples. A further 157 genes each contained 2 novel putatively functional variants and the remaining 1360 genes contained one novel putatively functional variant.
In the resistant smokers, there was no overall enrichment or depletion of novel putatively functional variants among the 26 regions reported to be associated with lung function [5], (16 were observed, the same number would have been predicted based on the sequence length of exons) and no novel putatively functional variants were identified within the CHRNA3/5 region which has been previously associated with smoking [13] and airflow obstruction [9].
Eight of the 1722 novel variants predicted to be putatively functional were identified in >1 case sample. These are listed in Table S3. ATAD3C contained a novel putatively functional variant for which six case samples were heterozygous, SHANK2 contained a novel putatively functional variant for which three cases were heterozygous, and the remaining six genes each contained such a variant for which two cases were heterozygous.
One hundred and ninety two Gene Ontology (GO) terms reached nominal significance for the set of 1533 genes containing novel putatively functional variants in resistant smoker cases. Of these, 22 high level GO terms were significant after Bonferroni correction for multiple testing and are listed in Table 1 [14]. The most significant GO term was the molecular function term “motor activity” which describes molecules involved in catalysis of movement along a polymeric molecule such as a microfilament or microtubule, coupled to the hydrolysis of a nucleoside triphosphate. Other related GO terms also feature amongst the significant signals from this analysis (including “cytoskeleton”, “microtubule motor activity”, “myosin complex”, “axoneme”, “cilium” and “cilium part”) (Table 1 and Table S4).
We tested for association of known and novel exonic variants with the resistant smoker phenotype. After exclusion of variants which were missing in >5% of either cases or controls, 269,822 (of which 215,747 were listed in dbSNP137) variants remained. Of the 269,822 variants, 94,138 were exonic and included in further analyses. Similar distributions of variants across the minor allele frequency spectrum were observed for the cases, primary, and secondary controls (results not shown).
After testing for association with resistant smoker status using primary controls, no SNPs reached genome-wide significance (P<5×10−7, based on Bonferroni correction for 94,138 tests). Substantial under-inflation of the test statistics was seen (lambda = 0.6) (Figure 2A), possibly due to the large number of rare variants (lambda = 0.92 if only variants with MAF>5% [n = 25,646] were considered, Figure 2B). Twenty exonic SNPs showed nominal evidence of association with P<10−3 (Table 2).
The strongest signal from a non-synonymous SNP was within a region previously identified as being associated with lung function [5]. The non-synonymous SNP in CCDC38 (rs10859974, OR = 2.36, P = 2.34×10−4) is 17.43 kb away from, but statistically independent of, rs1036429 (intronic, r2 = 0.064) which has previously shown genome-wide significant association with FEV1/FVC [5]. SNP rs10859974 itself has shown weak evidence of association with FEV1/FVC (P = 0.032) [5]. This SNP is predicted to cause a methionine to valine substitution at protein position 227; the valine allele is predicted to be protective. Investigations into CCDC38 expression in bronchial tissue via immunohistochemistry identified moderate staining of CCDC38 in the cytoplasm of columnar epithelial cells, with weak staining in the sub-epithelial layer (Figure 3). We found no evidence that rs10859974 or any of its proxy SNPs (r2>0.3) were lung eQTLs for CCDC38 itself, although rs11108320 which is intronic in CCDC38 and in strong LD with rs10859974 (r2 = 1) is an eQTL for nearby gene NTN4 (significant at 10% False Discovery Rate (FDR) threshold). Many additional SNPs located near or within CCDC38 and SNRPF were eQTLs for NTN4 (Table S5). Nearby CCDC38 intronic SNPs in weaker LD (r2 = 0.3) with rs10859974 were eQTLs for SNRPF (Table S5).
The strongest signal of association in the single-variant analysis was from a synonymous SNP, rs1287467, in SH3BP5 (OR = 0.47, P = 1.47×10−4) (Table 2). A SNP downstream of SH3BP5 (rs1318937, 1000G CEU MAF = 0.108, 16 kb from rs1287467, r2 = 0.018) has shown evidence of association with alcohol dependence and alcohol and nicotine co-dependence [15].
Synonymous SNP rs2303296 in ITSN2 was the second strongest signal of association (OR = 0.45, P = 2.31×10−4) and had previously shown weak evidence of association with FEV1 (P = 0.02) [5] but was not near to any previously identified genome-wide significant associations with lung function and has not shown evidence of association with COPD [9]. Another SNP in ITSN2, rs6707600 (intronic, 1000 G CEU MAF = 0.017, 89 Kb from rs2303296, r2 = 0.02), has shown some evidence of association with antipsychotic response in schizophrenia patients [16].
The second strongest signal from a non-synonymous SNP was rs4850 in UQCRC2 (OR = 4.87, P = 2.4×10−4). There were no nearby associations reported with any other trait for this gene.
The third strongest signal from a non-synonymous SNP was rs2297950 (OR = 0.51, P = 6.65×10−4) in CHIT1 which encodes chitinase 1 (Chit1). The chitinase pathway has been implicated in asthma and lung function [17] and lung function decline in COPD patients [18]. Chit1 expression in mice has been shown to be correlated with severity of bleomycin-induced pulmonary fibrosis (with overexpression leading to increased severity and Chit1−/− mice exhibiting reduced pulmonary fibrosis) [19].
A non-synonymous SNP in LOXL3, rs17010021, was the only SNP with an association P<10−3 regardless of whether the primary or the secondary controls were used (Table S6). This variant had a minor allele frequency of 0.048 and 0.061 in the primary and secondary control sets respectively, but the minor allele was not observed in any of the resistant smoker cases.
Synonymous SNP rs1051730, in CHRNA3 (15q25.1), has previously shown very strong evidence of association with smoking behaviour (particularly cigarettes per day) [13], [20], [21]. This SNP showed weak evidence of association with the resistant smoker phenotype in our study (P = 0.03 when the secondary control set was used and P = 0.06 when primary control set was used). Association results for SNPs within 500 Kb of rs1051730 are in Table S7.
No nominally significant enrichment of association signals in known pathways was identified in the exome-wide results of the single-variant analysis using MAGENTA [22].
SKAT [23] and AMELIA [24] analyses were undertaken to assess whether multiple variants within a gene collectively showed evidence of association; these tests are agnostic to whether a given variant is previously known. Quantile-Quantile plots for SKAT and AMELIA analyses are shown in Figure 4. Genes with nominally significant association (P<10−3) for SKAT or AMELIA analysis using the primary controls are shown in Table 3 (results of SKAT and AMELIA analyses using the secondary controls are shown in Table S8). No genes showed significant association after Bonferroni correction for multiple testing (P<0.05/18000 = 2.8×10−6) for either analysis (Table 3 and Table S8). Since the genes are likely to be correlated (through LD structure or overlapping reading frames), SKAT provides a resampling function to control Family Wise Error Rate (FWER). No genes were significant after controlling FWER = 0.05. None of the genes in Table 3 and Table S8 were within any of the 26 lung function associated regions [3]–[5], the CHRNA3/5 smoking-associated region [13] or SERPINA1 (mutations in which are known to cause alpha-1-antitrypsin deficiency) [25].
We also checked overlap between the gene-based association testing and single-variant tests. A signal in TMEM252 (which showed P<10−3 in the SKAT analysis regardless of which control set was used) was driven by rs117451470, a non-synonymous SNP, which had P = 2.2×10−3 in the single-variant association analysis (the other SNP in TMEM252, a singleton novel synonymous variant, had P = 0.38 in the single-variant analysis). Signals in UQCRC2 (strongest signal using SKAT), SPATA3D1, PGAP3 and ADCK2 were also driven by variants with P<10−3 in the single-variant analysis. Signals from TMX3, IMPG2 and TCOF1 were not driven by single-variant signals (all SNPs within these genes had P>0.01 in the single variant analysis). IMPG2 was the strongest signal from the AMELIA analysis and all 8 SNPs within IMPG2 had no evidence of association in the single-variant analysis (P> = 0.18).
We tested for enrichment of GO terms within the set of genes showing association with P<0.01 in the SKAT analyses. Ten high level GO terms reached nominal significance (P<0.05) for the set of 150 genes identified using SKAT but none were significant after Bonferroni correction for multiple testing [14].
Understanding why some heavy smokers seem to show resistance to the detrimental effects of cigarette smoke on lung function should provide further insight into the genetics of lung function and COPD. We undertook pathway enrichment, single-variant association testing and gene-based association testing analyses on whole exome re-sequencing data from a set of resistant smokers. Although no individual SNP achieved genome-wide statistical significance (P<5×10−7), our strongest association signal for a non-synonymous SNP was in CCDC38; a gene which has previously shown strong and robust evidence of association with lung function [5]. The intronic SNP previously shown to be associated with lung function (FEV1/FVC) and the non-synonymous SNP showing nominally significant association with the resistant smoker phenotype in this study are located close together (17.4 kb apart) but are not well correlated (the non-synonymous SNP has previously shown nominally significant evidence [P<0.05] of association with FEV1/FVC). A conditional analysis of these two SNPs was consistent with no statistical correlation between these signals. Although the function of CCDC38 is not yet well understood, members of the coiled-coil domain protein family are known to have a role in cell motor activity (e.g. myosin) [26] and cilia assembly [27], [28]. Expression of CCDC38 has been identified in the human bronchi of two subjects, with strong cytoplasmic staining in the epithelium and moderate staining in the airway smooth muscle (Human Protein Atlas [http://www.proteinatlas.org] [29]: ENSG00000165972). We experimentally confirmed these findings using immunohistochemistry on lung sections. We observed moderate cytoplasmic CCDC38 staining in bronchial columnar epithelial cells and some potential airway smooth muscle staining. There is no evidence that SNP rs10859974 is an eQTL for CCDC38 itself, although proxies for rs10859974 are eQTLs for a nearby downstream gene, NTN4, encoding Netrin-4 which may play a role in embryonic lung development [30]. Gene Ontology terms shown to be significantly enriched among the novel putatively functional variants identified only in the resistant smokers also pointed to pathways relating to motor activity and the cytoskeleton, including cilia. Another locus showing association with lung function (1p36.13, [5]) also contains a gene encoding a component of cilia (CROCC which encodes rootletin, another coiled-coil domain protein) and Crocc-null mice have been shown to have impaired cilia with pathogenic consequences to the airways [31]. The enrichment of genes involved in cilia function amongst the results of our analyses supports the importance of cilia function in lung health. Cilia abnormalities are known to be associated with smoking [32], [33], asthma [34], and play a role in COPD [35] where reduced cilia function leads to reduced mucus clearance of the airways. Improving mucociliary clearance is one of the aims of drug therapy for chronic bronchitis in COPD patients (reviewed in [36]).
Impaired cilia function is known to cause a wide range of diseases (collectively known as ciliopathies) many of which include pulmonary symptoms [37]. Primary Ciliary Dyskinesia (PCD) is a rare genetic disorder where respiratory tract cilia function is impaired leading to reduced (or absent) mucus clearance. Mutations in genes which encode components of the cilia have been found to cause several forms of PCD and include the dynein, axonemal heavy chain encoding genes DNAH11 [38], [39] and DNAH5 [40] within which resistant smoker-specific novel putatively functional variants were identified in this study (2 such variants were discovered in DNAH11). Whilst PCD affects resistance to infection and results in bronchiectasis, abnormal lung function can manifest early in life and progressive airflow obstruction has been observed in later life, although aggressive treatment may prevent the latter [41]. Retinitis pigmentosa is a feature of several ciliopathies, including some with pulmonary involvement (for example, Alstrom Syndrome). Low frequency variants in IMPG2 (interphotoreceptor matrix proteoglycan 2) collectively showed strong evidence of association (using AMELIA). Variants in IMPG2 are associated with a form of retinitis pigmentosa [42]. Another retinitis pigmentosa gene, RP1, was amongst the 16 genes containing 3 novel putatively functional variants in the resistant smokers. RP1 encodes part of the photoreceptor axoneme [43], a central component of cilia.
A recent study identified modulators of ciliogenesis using a high throughput assay of in vitro RNA interference of 7,784 genes in human retinal pigmented epithelial cells (htRPE) and identified 36 positive modulators and 13 negative modulators of ciliogenesis [44]. These modulators included many genes which did not encode structural cilia proteins and thus were not obvious candidates for a role in cilia function. None of the genes highlighted by the single-variant or gene-based analyses were confirmed as modulators of ciliogenesis although ITSN2, which contained one of the top signals in our single-variant analysis, was included in the screen and showed suggestive evidence of a positive role in ciliogenesis but this was not confirmed in a second screen. Two of the genes found to harbour a novel putatively functional variant in the resistant smokers were identified as positive modulators of ciliogenesis: GSN (gelsolin) which is a known cilia gene with a role in actin filament organisation and AGTPBP1 (ATP/GTP binding protein 1) which has a role in tubulin modification.
Collectively, our data show an enrichment of novel putatively functional variants in genes related to cilia structure and function in resistant smokers. Association between smoking and shorter cilia has been reported [32]. The largest genome-wide association with lung function to date supports the notion that the majority of associated variants, including those associated with COPD risk, affect lung function development rather than decline in lung function in adults [5]. If confirmed in other studies, it would be interesting to assess whether genetic influences on the function of cilia primarily affect growth or whether these affect more directly the extent of damage caused by tobacco smoke.
Very large GWAS have identified up to hundreds of common variants each with a modest effect on a variety of phenotypes. However, collectively, these still only explain a very modest proportion of the additive polygenic variance. It has been widely hypothesised that rare variation may account for some of this missing variance [45]. Commercially available SNP arrays have tended to include mostly variants with minor allele frequencies upwards of 5% and rare variants have not been reliably imputed from these. Re-sequencing approaches provide the most accurate platform for the study of exome-wide and genome-wide rare variation. However, there is increasing evidence that rare variants may not account for the missing heritability for all traits [46]. Our study did not find evidence for any individual rare variants with large effects in any of the known lung function associated loci or elsewhere in the exome (albeit in a modest overall sample size), although we did identify significant enrichment of novel rare variants in sets of genes with known functions in pathways which are known to have a role in lung health.
For the single variant analyses, we used Fisher's Exact Test. Whilst this is an appropriate test to use for small cell counts (for example, where minor allele counts are low), alternatives have been recently proposed including the Firth test, and although the optimal approach in the size of study we undertook is not clear from the comparisons shown to date, the Fisher's Exact Test can be more conservative than the Firth test and this may have had some impact on the power of the study [47]. Methods for the analysis of rare variant data are continuing to evolve.
Although this is the first exome re-sequencing study of resistance to airways obstruction among heavy smokers, our study does have potential limitations. Sample size was limited both by availability of individuals with such an extreme phenotype as that we were able to study, and also by current sequencing costs. We were able to utilise re-sequencing data available to the scientific community as control data and therefore maximise the discovery potential of our resources by re-sequencing to a sufficient sequencing read depth for confident rare variant calling. By doing so, and selecting an extreme phenotype group from our sampling frame, we adopted a suitable design to test whether there was enrichment of rare variants of large effect in resistant smokers. The same limitations also impact on the availability of suitable replication studies. In particular, it would have been desirable to undertake replication to support the statistically significant findings of the pathway analysis. However, in the absence of a suitable replication resource, the prior evidence for the role of cilia in lung health does lend support to our findings. As it becomes possible to sample and re-sequence from very large biobanks it should become possible to circumvent these issues in years to come, particularly if the cost of sequencing falls.
As limited information was available on smoking status among the controls, we did not restrict controls to heavy smokers and there is therefore potential for genetic associations to be driven via an effect on smoking behaviour. Nevertheless, our design is also consistent with the detection of association due to primary effect on airways and previous genome-wide association studies of lung function not fully adjusted for smoking have detected loci associated with lung function and COPD which were not associated with smoking behaviour [4], [5]. We saw only a weak association with variants at the CHRNA3/CHRNA5 locus (the locus at which variants have shown the strongest effect with smoking behaviour [13], [20], [21]).
Misclassification impacts on power; we would have underestimated the power to detect SNP and gene-based associations if the prevalence of resistance to airways obstruction among heavy smokers was greater than we assumed. In a cross-sectional study of this kind, survivor bias could occur if genetic variants influencing survival were under-represented or over-represented in the resistant smokers, but as the mean age of the resistant smokers was 56.4, any survivor bias, if present, is unlikely to have had a major impact. Finally, although we would expect the allele frequencies of the control sets we used to be representative of a general population control set across the vast majority of the genome, biases could potentially be introduced for any genetic variants related to the ascertainment strategy of the control sets. For the main findings we report in this paper, we also present allele frequencies from a public database (1000 Genomes Project); any such bias does not explain our main findings.
In the first deep whole exome re-sequencing study of the resistant smoker phenotype, we have identified an association signal in a region that has already shown robust association with lung function (CCDC38) and demonstrate significant enrichment of novel putatively functional variants in genes related to cilia structure. These findings provide insights into the mechanisms underlying preserved lung function in heavy smokers and may reveal mechanisms shared with COPD aetiology.
The Gedling study was approved by the Nottingham City Hospital and Nottingham University Ethics committees (MREC/99/4/01) and written informed consent for genetic study was obtained from participants. The Nottingham Smokers study was approved by Nottingham University Medical School Ethical Committee (GM129901/) and written informed consent for genetic study was obtained from participants. The Edinburgh MR-psychosis sample set was compliant with the UK10K Ethical Governance Framework (http://www.uk10k.org/ethics.html) and no restrictions were placed on the use of the genetic data by the scientific community. For TwinsUK, ethics committee approval was obtained from Guy's and St Thomas' Hospital research ethics committee. Tissue for immunohistochemistry was from Nottingham Health Science Biobank (Nottingham, UK) with the required ethical approval (08/H0407/1).
For lung eQTL datasets: At Laval, lung specimens were collected from patients undergoing lung cancer surgery and stored at the “Institut universitaire de cardiologie et de pneumologie de Québec” (IUCPQ) site of the Respiratory Health Network Tissue Bank of the “Fonds de recherche du Québec – Santé” (www.tissuebank.ca). Written informed consent was obtained from all subjects and the study was approved by the IUCPQ ethics committee. At Groningen, lung specimens were provided by the local tissue bank of the Department of Pathology and the study protocol was consistent with the Research Code of the University Medical Center Groningen and Dutch national ethical and professional guidelines (“Code of conduct; Dutch federation of biomedical scientific societies”; http://www.federa.org). At Vancouver, the lung specimens were provided by the James Hogg Research Center Biobank at St Paul's Hospital and subjects provided written informed consent. The study was approved by the ethics committees at the UBC-Providence Health Care Research Institute Ethics Board.
100 individuals with prolonged exposure to tobacco smoke and unusually good lung function (resistant smokers) were selected from the Gedling and Nottingham Smokers studies, described below.
The Gedling cohort is a general population sample recruited in Nottingham in 1991 (18 to 70 years of age, n = 2,633) [11] and individuals were then followed-up in 2000 (n = 1346) when blood samples were taken for DNA extraction, and FEV1 and FVC were measured using a calibrated dry bellows spirometer (Vitalograph, Buckingham, UK), recording the best of three satisfactory attempts [12].
The Nottingham Smokers cohort is an ongoing collection in Nottingham using the following criteria; European ancestry, >40 years of age and smoking history of >10 pack years (currently n = 538). Lung function measurements (FEV1 and FVC) were recorded at enrolment using a MicroLab ML3500 spirometer (Micro Medical Ltd, UK) recording the best of three satisfactory attempts.
Our inclusion criteria was; aged over 40 with more than 20 pack years of smoking and no known history of asthma. A total of 184 samples were eligible for this project after further exclusion of individuals with either FEV1, FVC or FEV1/FVC less than the Lower Limit Normal (LLN) (based on age, sex and height). We calculated residuals after adjusting % predicted FEV1 for pack years of smoking and selected the 100 samples with the highest residuals for exome re-sequencing (Figure S1).
Primary controls were from the Edinburgh MR-psychosis set (n = 166) of the UK10K project (http://www.uk10k.org/) and consisted of subjects with schizophrenia, autism or other psychoses, and with mental retardation. No additional phenotype information was available for the primary controls. The TwinsUK secondary control samples (n = 230) were all unrelated females selected from the high and low ends of the pain sensitivity distribution of 2500 volunteers from TwinsUK [48], [49]. Characteristics of the secondary controls are given in Table S1B (note that phenotype information was only available for a subset of the samples). These secondary controls were not included in the main analyses due to the difference in exome coverage. Further phenotype information was not available for either control sample set.
For the 100 resistant smoker case samples, DNA was extracted from whole blood and the Agilent SureSelect All Exon 50 Mb kit was used for enrichment. The 100 resistant smoker samples were individually indexed and grouped into 20 pools of 5 samples. Each pool was sequenced in one lane (20 sequencing lanes in total) using an Illumina HiSeq2000. Sequences were generated as 100 bp paired-end reads. Exome-wide coverage of 97 out of 100 samples was >20 (Figure S2). Three samples had mean sequence depth coverage <20, of these, one appeared to have had poor enrichment (high number of off-target reads), one had a low overall sequence yield and high number of duplicate reads and one had a high number of duplicate reads (but good sequence yield). To preserve power, and because there was no evidence that the sequence data quality for these samples was lower than for the other samples, these 3 samples were not excluded from further analyses.
A total of 166 exomes from the Edinburgh MR-psychosis study: a subset of the neurodevelopmental disease group of the UK10K project (http://www.uk10k.org/), were used as primary controls. These were enriched using the Agilent SureSelect All Exon 50 Mb kit and sequenced using an Illumina HiSeq2000 to a mean coverage depth of ∼70x (75 bp paired-end reads). The sequencing of the secondary controls from the TwinsUK pain study has been described elsewhere [49]. In brief, raw sequence data was available for 230 exomes which had been enriched using the NimbleGen EZ v2 (44 Mb) array and sequenced on an Illumina HiSeq2000 to a mean depth of coverage of 71x (90 bp paired-end reads).
The sequence alignment of the primary control exomes has been described elsewhere (http://www.uk10k.org/). The 100 resistant smoker case exomes and 230 TwinsUK controls were aligned using BWA v0.6.1 [50] with -q15 for read-trimming. Samtools v0.1.18 [51] was used to convert sort, remove duplicates and index the alignment .bam files. GATK v1.4-30 [52] was used to undertake local realignment around indels and to recalibrate quality scores for all 3 datasets.
In order to identify novel variant calls in the 100 resistant smoker exomes, GATK and SAMtools mpileup were run on a per sample basis for all 100 exomes. Only bases with a base quality score >20 were included. The variants called were then compared with dbSNP137, 1000 Genomes Project (1000G) and NHLBI Exome Sequencing Project calls and all known variants were excluded in order to identify novel rare variants which were unique to the 100 resistant smoker exomes. The novel GATK variant calls were then excluded if they had a Phred scaled probability (QUAL) score <30, quality by unfiltered depth (non-REF) (QD) <5, largest contiguous homopolymer run of the variant allele in either direction >5, strand bias >−0.1 or Phred-scaled P value using Fisher's Exact Test to detect strand bias >60. The novel SAMtools mpileup variants were excluded if they had a QUAL<30, mapping quality <25 or genotype quality <25. Variants called at sites with a depth of coverage less than 4 or greater than 2000 were also excluded. The intersect of variants which were identified and passed filtering using both GATK and SAMtools mpileup was taken forward for further analysis.
CAROL (http://www.sanger.ac.uk/resources/software/carol/) was used to predict the consequence of all coding variants. This method combines the results of the functional scoring tools SIFT and Polyphen2. SNPs were predicted as being putatively functional if they had CAROL score>0.98. Amino acid changes were predicted using ENSEMBL.
For each comparison (resistant smoker cases vs. primary controls and resistant smoker cases vs. secondary controls), variant calling was undertaken across cases and controls together using the GATK v1.5-20 Unified Genotyper. Only bases with a base quality score >20 were included. Coverage was down-sampled to 30 (reads are drawn at random where coverage is greater than 30). This was done to improve comparability between cases and controls and to speed up computation. A minimum QUAL of 30 was used as the threshold for calls. The GATK VQSR approach was used to filter variants across all samples. Variants with QUAL<30 and VQSLOD score equivalent to truth< = 99.9 were excluded (VQSLOD score<2.2989). Only single nucleotide polymorphism variants (SNPs) were called. There was >99% genotype concordance with genotype array data (Illumina 660k) for ∼5000 exonic SNPs with MAF>5% in the 100 resistant smokers.
Single-variant association testing was undertaken using the Fisher's exact test for a comparison of resistant smoker cases and primary controls. A secondary comparison of the resistant smoker cases and the secondary controls was also undertaken although results were interpreted with caution due to disparity of the exome coverage at the pre-sequencing enrichment stage between the cases and secondary controls. Two approaches to analyse the effect of multiple variants within genes were used: SKAT (v0.92) [23] and AMELIA [24]. Variants were assigned to RefSeq genes using Annovar [53]: a total of 16439 genes were identified as containing variants in the resistant smoker cases vs. primary controls analysis. Analysis with SKAT was undertaken using default weighting to account for the assumption that rare variants are likely to have bigger effect sizes.
An alternative method, AMELIA [24], was run using a subset of the variants with MAF<5%. A total of 18182 genes were identified as containing variants (with MAF<5%), of which 7516 contained more than 4 variants and so could be reliably tested by AMELIA.
For both SKAT and AMELIA, only variants which were annotated as exonic, 5′UTR or 3′UTR were included.
Power estimates for the identification of novel putatively functional variants in cases only, single-variant association tests and SKAT analysis were undertaken.
For a given variant unique to, and with a minor allele frequency of 0.005 in, resistant smokers, the probability of identifying at least one copy of the minor allele in 100 such individuals is 0.63 (0.86 for a minor allele frequency of 0.01).
Estimates of power for the single-variant association tests were undertaken for a sample size of 100 cases (resistant smokers) and 166 controls assuming a prevalence of the resistant smoker phenotype of 2% in the controls. Power calculations for detecting single variants were undertaken using Quanto and are shown in Figure S3. As an example, power to detect a variant with an allele frequency of 0.01 and an OR of 10 would be 10% for an alpha level of 5×10−8, and 81% for an alpha level of 0.001.
SKAT power calculations were run using the R package SKAT. The simulated dataset that the R package provides based on the coalescent populations genetic model was used to assess LD and MAF. The “Log” option was used to specify that the logOR distribution varies with allele frequency (logOR increases as minor allele frequency decreases), the effect size of each variant is equal to c|log10(MAF)|, where c is estimated assuming that the maximum OR corresponds to a MAF of 10−4. It was assumed that no logOR for causal variants was negative (results were broadly consistent if 20% of the causal variants were assumed to have negative logOR, results not shown). One thousand simulations were run for a region length of 17.7 kb (median of all gene lengths analysed in the real data), maximum OR of all variants analysed ranged from 5 to 10, significance (alpha) thresholds of 2.8×10−6 (Bonferroni correction for testing of 18,000 genes) and 0.01 (nominal significance threshold used to define genes as input to DAVID Gene Ontology analysis) were used and the percentage of causal variants with MAF<1% (only variants with MAF<1% were considered causal) given were 25% and 50%.
Power to detect a region of length 17.7 kb with a maximum OR of 10, assuming that 50% of variants with MAF<1% are causal is 53% for a Bonferroni-corrected significance threshold of 2.8×10−6, and 89% for a nominal significance threshold of 0.01 (Figure S4).
We tested for enrichment of Gene Ontology terms and enrichment of signals in known biological pathways within the results of the single-variant, gene-based and case-only analyses.
A total of 150 genes had P<0.01 in the SKAT analyses (of these, 28 also had P<0.01 in the AMELIA analysis but many genes were not analysed using both SKAT and AMELIA and so only SKAT results, which included all SNPs with no MAF cut-off, were included in this analysis). A total of 1533 genes contained novel putatively functional variants in the resistant smoker cases. We tested for enrichment of Gene Ontology categories within each of these gene lists using DAVID [14] with an EASE (modified Fisher's Exact) P<0.05.
We tested for pathway enrichment within the single-variant association results using MAGENTA v2 [22]. Briefly, MAGENTA tests for deviation from a random distribution of strengths of association signals (P values) for each pathway and includes all available exome-wide single-variant association results (n = 94,138). Six databases of biological pathways were tested: including Ingenuity Pathway (June 2008, number of pathways n = 92), KEGG (2010, n = 186), PANTHER Molecular Function (January 2010, n = 276), PANTHER Biological Processes (January 2010, n = 254), PANTHER Pathways (January 2010, n = 141) and Gene Ontology (April 2010, n = 9542). Significance thresholds were Bonferroni corrected for each database.
Fixed lung tissue was sectioned and mounted. Slides were treated in Histo-Clear and then re-hydrated using 100% ethanol and 95% ethanol washes. Antigen retrieval was carried out by steaming the tissue samples for 30 minutes in sodium citrate buffer (2.1 g Citric Acid [Fisons - C-6200-53]+13 ml 2M NaOH [Fisher - S-4880/53] in 87 ml H2O). Tissue was then treated with peroxidise blocking solution (Dako - S2023), followed by treatments with a 1 in 50 dilution of rabbit anti-CCDC38 antibody (Sigma HPA039305; 0.2 mg/ml) or a 1 in 50 dilution of the Rabbit IgG Isotype control (Invitrogen 10500C, diluted to 0.2 mg/ml). Secondary antibody staining and DAB treatment was carried out using the EnVision Detection Systems Peroxidase/DAB, Rabbit/Mouse kit (Dako – K5007). Tissue was then counterstained with Mayers Hematoxylin solution (Sigma – 51275) before being dehydrated using 95% ethanol and 100% ethanol washes. Slides were mounted using Vectamount (Vector Laboratories - H-5000).
The description of the lung eQTL dataset and subject demographics have been published previously [54]–[56]. Briefly, non-tumor lung tissues were collected from patients who underwent lung resection surgery at three participating sites: Laval University (Quebec City, Canada), University of Groningen (Groningen, The Netherlands), and University of British Columbia (Vancouver, Canada). Whole-genome gene expression and genotyping data were obtained from these specimens. Gene expression profiling was performed using an Affymetrix custom array (GPL10379) testing 51,627 non-control probe sets and normalized using RMA [57]. Genotyping was performed using the Illumina Human1M-Duo BeadChip array. Genotype imputation was undertaken using the 1000G reference panel. Following standard microarray and genotyping quality controls, 1111 patients were available including 409 from Laval, 363 from Groningen, and 339 from UBC. Lung eQTLs were identified to associate with mRNA expression in either cis (within 1 Mb of transcript start site) or in trans (all other eQTLs) and meeting the 10% false discovery rate (FDR) genome-wide significant threshold.
|
10.1371/journal.ppat.1004346 | The GAP Activity of Type III Effector YopE Triggers Killing of Yersinia in Macrophages | The mammalian immune system has the ability to discriminate between pathogens and innocuous microbes by detecting conserved molecular patterns. In addition to conserved microbial patterns, the mammalian immune system may recognize distinct pathogen-induced processes through a mechanism which is poorly understood. Previous studies have shown that a type III secretion system (T3SS) in Yersinia pseudotuberculosis leads to decreased survival of this bacterium in primary murine macrophages by unknown mechanisms. Here, we use colony forming unit assays and fluorescence microscopy to investigate how the T3SS triggers killing of Yersinia in macrophages. We present evidence that Yersinia outer protein E (YopE) delivered by the T3SS triggers intracellular killing response against Yersinia. YopE mimics eukaryotic GTPase activating proteins (GAPs) and inactivates Rho GTPases in host cells. Unlike wild-type YopE, catalytically dead YopER144A is impaired in restricting Yersinia intracellular survival, highlighting that the GAP activity of YopE is detected as a danger signal. Additionally, a second translocated effector, YopT, counteracts the YopE triggered killing effect by decreasing the translocation level of YopE and possibly by competing for the same pool of Rho GTPase targets. Moreover, inactivation of Rho GTPases by Clostridium difficile Toxin B mimics the effect of YopE and promotes increased killing of Yersinia in macrophages. Using a Rac inhibitor NSC23766 and a Rho inhibitor TAT-C3, we show that macrophages restrict Yersinia intracellular survival in response to Rac1 inhibition, but not Rho inhibition. In summary, our findings reveal that primary macrophages sense manipulation of Rho GTPases by Yersinia YopE and actively counteract pathogenic infection by restricting intracellular bacterial survival. Our results uncover a new mode of innate immune recognition in response to pathogenic infection.
| The type III secretion system (T3SS) is a macromolecular protein export pathway found in gram-negative bacteria. It delivers bacterial toxins into eukaryotic cells to promote pathogenic infection. T3SSs and the bacterial toxins delivered are critical arsenals for many bacterial pathogens of clinical significance, such as Yersinia, Salmonella and Shigella. On the other hand, the mammalian immune system may recognize the T3SS as a danger signal to signify pathogenic infection, and to stimulate appropriate defense against pathogens. Here, we show that the innate immune system recognizes the activity of YopE delivered by the Yersinia T3SS. Modulation of host Rho GTPases by YopE elicits a defensive response, which results in killing of bacteria in host cells. Inhibition of host Rho GTPases by Clostridium difficile Toxin B, another bacterial toxin, mimics the YopE-triggered killing effect. Our study demonstrates that host cells sense manipulation of Rho GTPases by bacterial toxins as a surveillance mechanism, revealing new insights into innate immune recognition of pathogenic infections.
| Innate immunity provides an early and critical protection against pathogenic infection. In the dominant paradigm of innate immunity, host cells detect pathogens by recognition of “microorganism-associated molecular patterns” (MAMPs) via pattern recognition receptors (PRRs) [1]. However, MAMPs, such as flagellin or lipopolysaccharide (LPS), are conserved microbial structures found in both pathogenic and nonpathogenic bacteria. How then do host cells distinguish pathogens from innocuous microbes? Alternate theories propose that, in addition to MAMPs, host cells also respond to distinct pathogen-induced signals, termed “patterns of pathogenesis” [2]–[4]. Several recent studies have demonstrated that host cells sense the activities of bacterial effectors, such as inhibition of host protein synthesis, activation of host Rho GTPases or pore forming activity, resulting in an active response against the pathogenic attack [5]–[10]. The protective immune response that is triggered by the detection of microbial effectors is defined as an “effector-triggered immune response” (ETIR).
In the genus of Yersinia, three species are pathogenic for humans: Yersinia pestis, Yersinia pseudotuberculosis and Yersinia enterocolitica. Y. pestis is the causative agent of plague and is typically transmitted by fleabites or aerosols [11], [12]. Y. pseudotuberculosis and Y. enterocolitica are associated with self-limiting gastroenteritis acquired from contaminated food or water [11]. The virulence of pathogenic Yersinia requires a plasmid (pYV in Y. pseudotuberculosis), which encodes a T3SS and a suite of effectors named Yersinia outer proteins (Yops) [13]. Upon Yersinia infection, Yop effectors are translocated into host cells by the T3SS to modulate host signaling pathways [13]. Four Yop effectors act to target Rho GTPases by distinct mechanisms: YopE mimics the eukaryotic GTPase activating protein (GAP) and promotes GTP hydrolysis to inhibit Rho GTPase activation; YopH, a protein tyrosine phosphatase, impacts Rho GTPase activation by interrupting activating signals for guanine exchange factors (GEFs); YopT, a cysteine protease, proteolytically removes the C-terminal isoprenoid moiety of Rho GTPases, therefore releasing their membrane anchors; YpkA can bind to Rho GTPases with a guanine dissociation inhibitor (GDI) domain [13]. By disturbing Rho GTPase activity, YopE, YopH, YopT and YpkA exert a negative effect on cytoskeleton dynamics, thus contributing to the anti-phagocytic activity of the Yersinia T3SS. In addition, YopJ inhibits NF-κB and MAPK signaling pathways, while YopK regulates effector delivery as well as host responses [10]. Translocators YopB and YopD are required for the formation of the T3SS channel and delivery of effector Yops.
The prototypical bacterial effector YopE is a 219 amino acid protein containing a Rho GAP domain (residues 96 to 219) [14]. YopEGAP shares homology with SptPGAP from Salmonella Typhimurium and ExoSGAP from Pseudomonas aeruginosa. YopE introduces an “arginine finger” into the GTPases catalytic site, which results in efficient GTP hydrolysis and deactivation of GTPases. Exchanging Arg144 in the “arginine finger” with an alanine residue abolishes YopE GAP activity [14]. In mammalian cells, YopE localizes to plasma membrane and unidentified perinuclear compartments, which requires a hydrophobic leucine-rich motif within its membrane localization domain (MLD, residues 53 to 79) [15]–[18]. Stability of YopE in host cells is influenced by allelic variations of residues 62 and 75, as found in different Yersinia strains. The presence of lysine residues at position 62 or 75 can mediate YopE ubiquitination and degradation by the host cell proteasome pathway [19]. Both subcellular membrane localization and stability of YopE are important for its GAP activity [16], [19]. YopE is equally effective on Rac1, RhoA and Cdc42 in vitro [14], whereas it is preferably active on Rac1 and RhoA, but not Cdc42, in vivo [20]. Unlike YopE, YopT seems to be primarily effective on RhoA, but not Rac1 or Cdc42 in vivo [21]. However, overexpressed YopT also acts on Rac1 in Yersinia-infected epithelial cells [22]. Interestingly, under the latter condition, YopT competes with YopE for the same pool of membrane-associated Rac1, promotes translocation of cleaved Rac1 into the nucleus, and therefore interferes with the ability of YopE to inactivate Rac1 [22].
Yersinia is generally considered as an extracellular pathogen, as the bacteria grow primarily in an extracellular form in vivo; however, these bacteria can survive and grow inside phagocytic cells, which may be important for the early stages of colonization [23]. It is suggested that macrophages might serve as permissive sites for bacterial replication or even as transport vehicles from the initial site of infection to deeper lymph tissues [24]. Interestingly, T3SS function decreases survival of Y. pseudotuberculosis in murine macrophages. Under experimental conditions in which T3SS expression is pre-induced, macrophages restrict intracellular survival of wild-type Y. pseudotuberculosis, but not a yopB− mutant (deficient in Yops translocation) or a pYV− mutant (missing the entire T3SS) [25]. Thus, some T3SS-dependent factor encoded in the wild-type strain triggers a bactericidal response in macrophages, the mechanism of which remains unclear. It has been shown that upon internalization of Y. pseudotuberculosis, the T3SS stimulates Ca2+-dependent phagolysosome fusion in macrophages, mediated by the Ca2+ sensor SytVII, leading to increased killing of intracellular bacteria [26]. Also, it has been reported that the Y. pseudotuberculosis T3SS stimulates Ca2+- and caspase-1-dependent lysosome exocytosis, releasing antimicrobial factors [27]. Yet, further studies are needed to determine the molecular basis of innate immune recognition of the Yersinia T3SS, and the role of this process in determining the fate of the bacteria in macrophages.
Here we hypothesize that the activities of the Yersinia T3SS effectors are sensed by host cells as patterns of pathogenesis, which stimulate an intracellular killing response against Yersinia as a type of ETIR. We show that macrophages recognize pathogenic Y. pseudotuberculosis through T3SS functions and elicit an intracellular killing response to counteract infection. We provide evidence that YopE GAP activity is a critical factor sensed by macrophages, with YopH playing a minor role. Overexpression of YopT counteracts the YopE-triggered killing effect possibly by competing for the Rho GTPase target and by reducing YopE translocation. Also, this YopE-triggered intracellular killing response can be mimicked by other bacterial derived toxins like Clostridium difficile Toxin B, indicating that host cells sense manipulation of Rho GTPases as a conserved surveillance pathway to detect pathogens. Thus, our data provide another example of a protective host response induced by pathogenic bacteria through recognition of bacterial effector activities on Rho GTPases, revealing a novel mode of innate immune recognition towards pathogenic infection.
Previous studies have shown that T3SS decreases survival of Y. pseudotuberculosis in murine macrophages [25], [26]. To determine if specific Yop effectors might contribute to decreased survival of Y. pseudotuberculosis in macrophages, the wild-type strain IP2666 and several yop deletion mutants were studied. Initially, the survival of IP2666 (wild-type), IP17 (yopEH−), IP27 (yopEHJ−) and IP37 (yopEHJMKYpkA−) (Table 1) in murine bone marrow-derived macrophages (BMDMs) was compared. Naïve BMDMs were infected with the indicated strains, followed by gentamicin treatment to eliminate extracellular bacteria. At 1 h and 23 h post infection, infected BMDMs were lysed and spread on LB plates to enumerate viable bacteria. CFU at 1 h post infection was considered as the initial intracellular bacterial count. The ratio of CFU between 23 h and 1 h post infection was calculated for each strain. At 1 h post infection, IP2666 showed lower CFU as compared to IP17, IP27 and IP37 (Figure S1A). This is expected because IP2666 expresses Yops with anti-phagocytic functions (YopE and YopH); however the other strains are yopEH− mutants. At 23 h post infection, IP17 displayed significantly higher level of CFU as compared to IP2666, but similar level as compared to IP27 and IP37 (Figure S1B). Consistently, for the ratio of CFU at 23 h/1 h, the level of IP17 was significantly higher than IP2666, but similar to IP27 and IP37 (Figure 1A). To rule out a threshold effect due to the differences in the initial bacterial uptake, BMDMs were infected with IP2666 or IP17 at different MOIs (Figure S2). Even with higher CFU at 1 h post infection, IP2666 (MOI = 10) still showed decreased survival in comparison to IP17 (MOI = 5 or 2.5) at 23 h post infection (Figure S2AB). Therefore, IP2666 shows reduced intracellular survival, in contrast to IPI7, which shows an intracellular growth phenotype, indicating that deletion of yopE and yopH promotes Yersinia survival inside macrophages.
To further elucidate the effects of YopE and YopH on intracellular survival of Yersinia, IP2666 (wild-type), IP6 (yopE−), IP15 (yopH−) and IP17 (yopEH−) (Table 1) were compared by CFU assay (Figure 1B). IP6 showed an intracellular growth phenotype similar to IP17, while IP15 had an intermediate phenotype (Figure 1B). The results were further confirmed by fluorescence microscopy. IP2666, IP6 and IP17 encoding GFP were used to infect BMDMs for different lengths of time. One hour before fixation and examination of the samples by fluorescence microscopy, IPTG was added to induce de novo expression of GFP from viable intracellular bacteria (Figure 2). At 23 h post infection, IP6 and IP17 showed greater survival compared to IP2666 (Figure 2). These results demonstrate that YopE is required for reduced survival of Yersinia in macrophages, and YopH cooperates with YopE in this process.
To determine whether SytVII-mediated phagolysosome fusion contributes to YopE-dependent intracellular killing, SytVII−/− BMDMs were compared to wild-type BMDMs for their ability to restrict intracellular survival of IP2666, IP17 or IP40 (yopB mutant, Table 1). The SytVII−/− genotype was verified by PCR using mouse-tail genomic DNA, in comparison to wild-type mice (Figure S3A). No significant difference was observed by CFU assay for IP2666 survival inside wild-type or SytVII−/− BMDMs (Figure S3B), suggesting that SytVII-mediated phagolysosome fusion does not contribute to the YopE-dependent killing of Yersinia in macrophages.
Under our experimental conditions, Yersinia infection does not cause significant cell death of macrophages (below 2% LDH release after 23 h from IP2666 or IP6 infected macrophages, data not shown). Accordingly, reduced intracellular survival of IP2666 is not due to enhanced Yersinia-induced macrophage cell death. We also investigated the possibility that IP2666 infection induces gentamicin uptake and leads to enhanced bacterial killing by gentamicin. If this is true, with increasing amount of gentamicin, intracellular IP2666 would be more sensitive than IP17, due to more gentamicin uptake. To analyze this possibility, the survival of IP2666 and IP17 in macrophages was compared with increasing amount of gentamicin. IP2666 and IP17 responded similarly to increasing amounts of gentamicin (Figure S4), indicating that reduced intracellular survival of IP2666 is not due to increased gentamicin internalization.
To investigate if the GAP activity of YopE is crucial for macrophages to restrict survival of intracellular Yersinia, experiments were carried out to compare survival of bacteria producing YopE or YopER144A. In YopER144A, a single substitution of arginine to alanine was introduced at amino acid 144 to yield a catalytically dead protein [14]. A plasmid vector encoding yopE or yopER144A was introduced into IP6 (Table 1). The production level of YopE or YopER144A from the vector in trans was similar to the native level in the wild-type strain as shown by SDS-PAGE and immunoblotting (Figure 3A, compare lanes 1, 3 and 4). The survival of IP6+pYopE and IP6+pYopER144A in macrophages was then compared. IP2666 or IP6 containing the empty vector (Table 1) were analyzed in parallel as controls. IP6+pYopE displayed a reduced intracellular survival phenotype, similar to IP2666+empty vector (Figure 3B). In contrast, IP6+pYopER144A showed increased intracellular survival, comparable to IP6+empty vector (Figure 3B). Unexpectedly, the empty vector (pMMB67HE) had a negative effect on Yersinia survival inside macrophages (Figure S5), possibly due to the metabolic burden introduced by the plasmid [28], [29]. Nevertheless, these results demonstrate that YopE GAP activity is indispensable for causing reduced survival of Yersinia in macrophages, and we hypothesize that YopE GAP activity is somehow recognized by macrophages, triggering increased killing of intracellular Yersinia.
Given the activity of YopT towards Rho GTPases and its crosstalk with YopE, the potential influence of YopT on survival of Yersinia inside macrophages was studied. IP2666 is a yopT mutant due to a naturally-occurring deletion in pYV in this strain [30]. Plasmids that overexpress YopT or catalytically-inactive YopTC139S were introduced into IP2666; control strains containing the empty vector or a plasmid producing native levels of YopT under its native promoter were also constructed (Table 1). Analysis of proteins secreted by the bacteria under low calcium conditions using SDS-PAGE and immunoblotting showed that YopT and YopTC139S were overproduced at equal levels, while the native level of YopT was undetectable (Figure 4A, compare lanes 2, 3 and 4). Interestingly, when these strains were used to infect macrophages, overexpression of YopT in IP2666 significantly increased Yersinia intracellular survival, giving the opposite effect of YopE (Figure 4B and Figure S6A). Yersinia survival in macrophages was moderately increased when YopTC139S was overexpressed in IP2666 (Figure 4B and Figure S6A), indicating that YopT catalytic activity is important for counteracting the YopE-triggered killing effect. Expression of YopT at native level in IP2666 also slightly improved Yersinia intracellular survival (Figure 4B and Figure S6A).
Using detergent extraction assay and immunoblotting, lysates of infected macrophage were analyzed to detect the amounts of YopE that were translocated by the different strains. Overexpression of YopT or YopTC139S in IP2666 diminished YopE translocation to 8% or 25% of wild-type level respectively (Figure S7AB). Native level of YopT in IP2666 slightly reduced YopE translocation (75% of wild-type level) (Figure S7AB).
Active and inactive YopT proteins were overexpressed in IP6 or IP37 to further examine the mechanism by which this effector counteracts killing of Yersinia in macrophages. Overexpression of YopT or YopTC139S in IP6 equally enhanced bacterial survival (Figure S6B), while overexpression of active or inactive YopT proteins in IP37 had no effect on Yersinia survival inside macrophages (Figure S6C).
Taken together, these results suggest that YopT has the ability to counteract YopE-triggered intracellular killing effect, which is partially dependent on YopT protease activity. YopT catalytic activity may counteract the YopE effect by competing with YopE for a Rho GTPase target or by reducing YopE translocation. Thus, inactivation of a Rho GTPase by a specific mechanism, i.e. GAP mechanism, appears to be sensed by macrophages, resulting in increased killing of intracellular Yersinia.
Localization to membranes and stability of YopE are critical for functional GAP activity against Rho GTPases in host cells [15], [16], [19]. Since our results suggest that YopE GAP activity is sensed by macrophages, we hypothesize that membrane localization and stability of YopE will impact its ability to stimulate an intracellular killing response. To examine this possibility, plasmids encoding YopE variants that were defective for membrane localization (YopE3N) [17] or less stable (YopER62K) [19] were introduced into IP6. The resulting strains (Table 1) were used to infect macrophages and detergent extraction assays were used to compare the amounts of YopE, YopE3N and YopER62K that were translocated. The yopB mutant IP40, which is defective for Yop translocation, was used to infect macrophages as a negative control. The amount of YopE3N in the soluble fraction was comparable to wild-type YopE, indicating equal translocation of these proteins (Figure 5A, compare lanes 3 and 1). Some YopE proteins with reduced Rho GAP activity are translocated at higher levels as compared to the wild-type protein into epithelial cells infected with Y. pseudotuberculosis [31]. We did not observe hypertranslocation of YopE proteins with reduced Rho GAP activity in our experiments, possibly because YopE has a reduced ability to negatively regulate its own translocation into macrophages as compared to epithelial cells. The amount of YopER62K in the soluble fraction was lower compared to wild-type YopE, probably due to decreased stability as a result of increased ubiquitination (Figure 5A, compare lanes 5 and 1). The appearance of a slower migrating band for YopER62K was consistent with ubiquitination (Figure 5A, lane 5). IP6+YopE3N and IP6+YopER62K displayed improved survival in macrophages in comparison to IP6+YopE at 24 h post infection, as determined using immunofluorescence microscopy to detect intracellular Yersinia (Figure 5B). The results, quantified by the percentage of macrophages containing fluorescent intracellular Yersinia (Figure 5C), or CFU assay (Figure S8A), showed that IP6+YopE3N and IP6+YopER62K had increased bacterial survival compared to IP6+YopE at 24 h post infection. Since membrane localization and stability are important for YopE to efficiently inactivate Rho GTPases, these results provide additional evidence that macrophages sense the inactivation of one or more Rho GTPases, which results in killing of intracellular Yesinia.
YopE variants with altered Rho GTPase specificities [31] were compared to wild-type YopE for their capability to trigger intracellular killing response. YopEL109A has lower GAP activity towards RhoA (70% of wild-type level) and Rac2 (70% of wild-type level); YopE-SptP fusion protein, which contains the secretion and translocation domains of YopE and the GAP domain of SptP, has no GAP activity towards RhoA and decreased activity towards Rac1 (83% of wild-type level) and Rac2 (34% of wild-type level) [31]. The amounts of translocated YopEL109A and YopE-SptP were comparable to wild-type YopE in Yersinia-infected macrophages, and YopE-SptP displayed reduced mobility as expected due to its higher molecular weight (Figure 5A, compare lanes 1, 2 and 4). At 24 h post infection, IP6+YopEL109A and IP6+YopE-SptP showed improved survival inside macrophages compared to IP6+YopE, as demonstrated by immunofluorescence microscopy (Figure 6AB) and CFU assays (Figure S8B). These results indicate that the specificity of YopE GAP activity may impact its ability to trigger the intracellular killing. However, the results obtained with the YopEL109A and YopE-SptP variants did not reveal if inactivation of a specific Rho GTPase by YopE is important for intracellular killing. Since the activity of YopEL109A and YopE-SptP towards Rho GTPases other than RhoA, Rac1 and Rac2 is not clearly known, it is difficult to declare YopE interruption of which specific Rho GTPase is essential for macrophage recognition and intracellular killing.
To explore if this intracellular killing response applies to bacterial toxins targeting Rho GTPases, Clostridium difficile Toxin B was added to Yersinia infected macrophages. Toxin B has been well characterized to inactivate a wide range of Rho GTPases through glycosylation, including Rac1, RhoA/B/C, RhoG, TC10, and Cdc42 [32]–[34]. With Toxin B treatment, the survival of IP6, IP17 and IP40 was dramatically decreased as revealed by CFU assays (Figure 7A and B) and fluorescence microscopy in conjunction with mCherry induction (Figure 7C). Toxin B did not affect Yersinia growth in tissue culture media in the absence of macrophages; Toxin B did not cause significant cytotoxicity in macrophages in these experiments (data not shown). These results suggest that down-regulation of Rho GTPases by Toxin B is perceived by macrophages, inducing an intracellular killing response, mimicking the effect of YopE. Thus, the bactericidal effect triggered by Rho GTPase-inactivating toxins may be a general and conserved response to these bacterial toxins. In addition, the fact that Toxin B decreases IP40 survival inside macrophages implies that T3SS translocon is not essential for macrophage recognition of Rho GTPases-inactivating toxins to cause a bactericidal response.
To identify the Rho GTPase target of YopE critical for causing intracellular killing, specific Rho GTPase inhibitors were studied for their capability to mimic the YopE effect. Treatment with Rac1 inhibitor NSC23766 negatively impacted IP6 and IP40 survival inside macrophages, as demonstrated by CFU assays (Figure 8AB) and fluorescence microscopy with mCherry induction (Figure 8C). The Rac1 inhibitor triggered a reduced bactericidal effect in comparison to Toxin B in the CFU assay (compare Figure 7AB and 8AB). In contrast to the Rac1 inhibitor, the RhoA inhibitor TAT-C3 did not significantly affect Yersinia survival inside macrophages (Figure S9BC). Dramatic morphological changes were observed in TAT-C3 treated macrophages as early as 4 h upon treatment, confirming the efficiency of TAT-C3 towards RhoA (Figure S9A). These results signify that macrophages restrict Yersinia intracellular survival in response to Rac1 inhibition, but not to Rho inhibition.
Several recent studies have shown that the activities of certain bacterial effectors can stimulate transcriptional changes in host cells, resulting in ETIRs [5]–[9]. For example, activation of Rac1 and Cdc42 by SopE from Salmonella enterica serovar Typhimurium is sensed through NOD1 receptor, eliciting NF-κB activation in the host cells as a protective response [9].
To study if YopE stimulates an altered host response that can occur at the transcriptional level, the production of nitric oxide (NO) from macrophages infected with IP2666, IP6, IP17 or IP40 was compared. Specifically, the concentration of nitrite (NO2−), an indicator of NO, was measured by Griess assay. At 23 h post infection, comparing to IP6-, IP17- or IP40-infected macrophages, IP2666-infected macrophages produced significantly higher levels of NO (Figure 9). LPS- and IFN γ-treated macrophages were used as a positive control, while uninfected macrophages were used as a negative control (Figure 9).
To investigate whether YopE dependent-intracellular killing signals through NOD1 receptor, Nod1−/− BMDMs were compared to wild-type BMDMs for their ability to restrict intracellular survival of IP2666 or IP6 by CFU assay. No significant difference was observed for IP2666 survival inside wild-type or Nod1−/− BMDMs (Figure S10).
These results suggest that macrophages respond to wild-type Yersinia differently from yopE− mutant strains and produce higher levels of NO; however, YopE-triggered intracellular killing is not mediated by NOD1 receptor.
The aims of this study were to determine T3SS-dependent factors that restrict Yersinia survival inside macrophages and characterize the mechanism of this “patterns of pathogenesis” triggered host response. Our findings reveal that primary naïve macrophages sense manipulation of Rho GTPases by Yersinia Yop effectors. Three known effector Yops directly inhibit host Rho GTPases: YopE, YpkA and YopT; a fourth effector, YopH, inhibits signals that activate these small GTPases. YopE is an important virulence factor for resistance of Yersinia to the innate immunity, as a Y. pseudotuberculosis yopE null mutant was defective for systemic spread following oral infection in the animal model [35]. However, on the other hand, here we show that YopE GAP activity towards Rho GTPases is recognized by macrophages, stimulating increased killing of intracellular Y. pseudotuberculosis (Figure 10). YopH cooperates with YopE to cause this killing effect, most likely by inhibiting a phosphotyrosine dependent signaling pathway that activates Rho GTPases in response to Yersinia infection (Figure 10). YpkA has very mild effect on Yersinia intracellular survival (data not shown), perhaps due to its low expression level in comparison to other Yops. Interestingly, we have observed that overexpression of YopT counteracts the YopE-triggered intracellular killing effect, which involves the protease activity of YopT. We speculate that an important biological function of YopT is to counteract sensing of YopE by the innate immune system, possibly by preventing YopE access to activated Rho GTPase targets or removing YopE-inactivated Rho GTPases from phagosome membranes (Figure 10). Zhang et al. studied a Y. pseudotuberculosis strain (32777), different from that used here (IP2666), and showed that a mutant of 32777 encoding catalytically inactive YopJ, YopT, YopE and YopH still triggered intracellular bacterial killing, to the same level as wild-type 32777 [25]. We speculate that 32777 has additional Rho GTPase-inactivating effector(s) causing bacterial killing, which remain to be identified.
We have obtained evidence that Toxin B decreases Yersinia survival in macrophages by inactivating several Rho GTPases (Figure 10). To date, at least 20 Rho GTPase proteins belonging to 8 subfamilies have been described in mammals [36], [37]. Interestingly, we have shown that the target preference of YopE impacts its ability to trigger bacterial killing. Thus, one intriguing question to ask is does the innate immune system monitor effector manipulation of a specific Rho GTPase? Our results showed that Rac inhibition, but not Rho inhibition, stimulates the macrophage killing response against intracellular Yersinia (Figure 10). However, the Rac inhibitor only partially promotes killing of Yersinia in macrophages in comparison to Toxin B or YopE, suggesting that disturbance of additional Rho GTPases contributes to the intracellular killing response. Other Rho GTPase candidates may include, but are not limited to, Rac2 and RhoG, which have been shown to be YopE targets and expressed in macrophages [31], [37]–[39]. Further investigation is required to reveal if additional Rho GTPases serve as surveillance points in response to pathogenic effector manipulation [37].
Rho GTPases act as molecular switches that regulate numerous cellular functions, like cytoskeletal dynamics, gene transcription, vesicular trafficking, cell growth and apoptosis [37], [40]. In order to ensure proper signaling responses, the activities of Rho GTPases are tightly regulated by multiple mechanisms, including the canonical regulators (GAPs, GEFs and GDIs) and direct post-translational modifications (like phosphorylation and ubiquitination) [40], [41]. The abnormal inactivation of Rho GTPases by YopE might interfere with multiple Rho GTPase-mediated signaling pathways and lead to many different consequences to trigger intracellular bacteria killing in macrophages. One possibility is that YopE may cause accumulation of inactivated GDP-bound Rho GTPases on phagosome (Figure 10), which could be modified by ubiquitination to stimulate signaling pathways. Activation of Rac1 by cytotoxic necrotizing factor 1 (CNF1) from Escherichia coli induces Rac1 poly- and mono-ubiquitination, the biological function of the latter remains unclear [42]. In line with this, GDP-bound RhoA is targeted by the ubiquitin E3 ligase Cullin-3 for poly-ubiquitination and degradation [43]. Thus, it is tempting to speculate that YopE-inactivated GDP-bound Rho GTPases could be mono-ubiquitinated and serve as signaling components; or they could be poly-ubiquitinated to mediate xenophagic degradation of bacteria-containing vesicles [44]. Alternatively, by modulating vesicular trafficking, YopE activity may interrupt formation of Y. pseudotuberculosis-containing autophagosomes, which have been shown to be impaired in acidification and support survival of the bacteria in macrophages [45]. On the other hand, given that the role of autophagy in Yersinia survival in macrophages is controversial [45], [46], it is possible that YopE activity promotes autophagy to eliminate intracellular bacteria. Deuretzbaher et al. showed that β1-integrin-mediated Y. enterocolitica internalization by macrophages was coupled to autophagy activation, which seemed to be deleterious for bacterial intracellular survival. Another possibility is that the disruption of the actin cytoskeleton by YopE is sensed by the innate immune system. It has been suggested that NOD receptors or inflammasome components associated with the actin cytoskeleton may act as surveillance mechanisms, becoming activated upon perturbations by pathogens [2]. Interesting, a recent study by Shao and colleagues showed that Rho-inactivating toxins such as Clostridium difficile Toxin B and Clostridium botulinum C3 trigger Pyrin inflammasome activation in BMDMs [47]. They further demonstrated that Burkholderia cenocepacia induced inactivation of Rho GTPase stimulates Pyrin inflammasome activation as an immune defense, which limits bacterial intra-macrophage growth and regulates lung inflammation in infected mice [47]. Whether YopE triggers Yersinia intracellular killing through the inflammasome pathway remains to be investigated.
The overall host cell innate immune response to a T3SS-containing bacterial pathogen is unique and multifactorial. MAMPs, the T3SS translocon channel, and the activities of bacterial effectors are likely recognized as combined pathogenic signals by the host cell. A two-signal model, requiring a MAMP and a pattern of pathogenesis, was proposed as an innate immune strategy to evaluate the virulence potential of a pathogen and adjust immune response appropriately to avoid self-damaging inflammation [48]. For example, a type IV secretion system allows Legionella pneumophila to deliver bacterial effectors into the host cell cytosol to inhibit host protein synthesis [6]. In this case, the effector-mediated interference of host protein synthesis, in concert with TLR signaling, results in prolonged activation of NF-κB as an ETIR [6]. Our data suggest that the YopBD translocon is not essential for Toxin B- or Rac inhibitor-triggered bacterial killing in macrophages. Further studies are needed to determine if TLRs or β1-integrins are possible receptors in MAMP-PRR pathways that facilitate the YopE-triggered killing effect. Alternatively, some studies in the literature support the idea that patterns of pathogenesis are sufficient to induce defense responses independently of classical MAMPs [5], [49], [50]. Boyer et al. demonstrated that Escherichia coli CNF1 elicited a vigorous ETIR in flies via activation of Rac2 and the IMD kinase pathway, which was observed even in the absence of PRR ligation [5]. Thus, it is possible that unbalanced disruption of Rho GTPases by YopE is adequate to stimulate a protective immune response, resulting in restriction of Yersinia survival in macrophages.
Roy and colleagues observed that during internalization of Salmonella enterica serovar Typhimurium or Y. pseudotuberculosis, the T3SSs of these pathogens stimulated SytVII-dependent phagolysosome fusion and bacterial killing in macrophages [26]. We have shown that YopE-triggered intracellular bacterial killing does not require SytVII, suggesting that there are at least two independent pathways by which killing of Yersinia internalized by macrophages can be stimulated. The YopE-dependent pathway senses inactivation of Rho GTPases, while the SytVII dependent pathway appears to recognize translocon insertion in the plasma membrane.
Various bacterial T3SS and T4SS effectors modulate Rho GTPase functions and interfere with corresponding host signaling pathways to benefit pathogenic infection [51]. Given that Rho GTPases play multiple roles in many signaling pathways critical for cellular functions, it is not surprising to envision surveillance mechanisms monitoring the status of Rho GTPases. Our work highlights that inactivation of Rac1, and possibly other GTPases, by YopE from Y. pseudotuberculosis is detected by macrophages as a danger signal, stimulating an ETIR that restricts intracellular bacterial survival. Detection of pathogens via Rho GTPase surveillance adds another layer of complexity to the mechanisms of innate immune recognition, improving our understanding of how the innate immune system responds to pathogenic infection.
Use of mice for the preparation of BMDMs was carried out in accordance with a protocol that adhered to the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (NIH) and was reviewed and approved (approval #206152) by the Institutional Animal Care and Use Committee at Stony Brook University, which operates under Assurance #A3011-01, approved by the NIH Office of Laboratory Animal Welfare.
The Y. pseudotuberculosis strains used in this study are shown in Table 1. These bacteria were grown on LB agar plates or in LB broth at 28°C supplemented with 100 µg/ml ampicillin, 25 µg/ml kanamycin or 30 µg/ml chloramphenicol as needed. The plasmids pMMB67HE [52], pYopE [14], pYopER144A [14], pPEYopE [30], pYopT [53], pPTYopT [30], pYopTC139S [18], p67GFP3.1 [23] and p207mCherry [54] have been previously described.
A new series of plasmids expressing yopE mutants were created as described below. Plasmids encoding yopEL109A, yopER62K and yopE3N were generated as follows. DNA fragments encoding yopEL109A, yopER62K or yopE3N were obtained by PCR using primers yopE-3 (5′-CGGATCCCATATGAAAATATCATCATTTATTTC-3′) and yopE-EcoRI (5′-CGCGGAATTCCCATATCACATCAATGACAGTAATTT-3′). Recombinant plasmid DNA (pBAD33/YopEL109A, a gift from Joan Mecsas, Tufts University), or Y. pseudotuberculosis virulence plasmid DNA (from 32777 yopER62K or 32777 yopE3N, Zhang et al. submitted) was used as template for the PCR to obtain yopEL109A, yopER62K and yopE3N, respectively. The resulting DNA fragments were inserted into pETBlue2 vector using blunt end ligation. To create a plasmid encoding the yopE-sptP fusion, a DNA fragment containing the first 100 codons of yopE (yopE1–100) was amplified from IP2666 virulence plasmid DNA with primers yopE-infusion-5 (5′-TAATAAATAGTCATATGAAAATATCATCATTTATTTCTACATCACTG-3′) and yopE-infusion-3 (5′-AGGTTGCTTACTTTCCGTAGGACTTGGCATTTGTGC-3′). A DNA fragment containing codons 166–293 of sptP (sptP166–293) was amplified with primers sptP-infusion-5 (5′-ATGCCAAGTCCTACGGAAAGTAAGCAACCTTTACTCAGTATCG-3′) and sptP-infusion-3 (5′-CAGCCAAGCTGAATTTTAGCCGGCTTCTATTTTCTCAAGTTC-3′) using chromosomal DNA from Salmonella enterica Typhimurium strain 14028 as template. A DNA fragment encoding the yopE1–100sptP166–263 fusion was made by overlapping PCR using the yopE1–100 and sptP166–293 fragments as templates and primers yopE-infusion-5 and sptP-infusion-3. The product was inserted into pETBlue2 by blunt end ligation. The sequences of the inserts in the plasmids described above were confirmed by DNA sequencing. DNA fragments encoding yopEL109A, yopER62K, yopE3N or yopE1–100sptP166–263 were obtained from the pETBlue2 vectors by digestion with NdeI and EcoRI, and ligated between the NdeI and EcoRI sites in pPEYopE, thereby replacing the wild-type yopE gene, and placing the mutant alleles under control of the native yopE promoter. The resulting plasmids pYopEL109A, pYopER62K, pYopE3N and pYopE-SptP were introduced into E. coli S17-1 λpir by electroporation and subsequently transferred into IP6 (Table 1) by conjugation as described previously [55].
Bone marrow-derived macrophages (BMDMs) were isolated and cultured from femurs of C57BL/6 wild-type mice (Jackson Laboratory) or SytVII−/− C57BL/6 mice (a generous gift from Dr. Norma Andrews, University of Maryland), or Nod1−/− C57BL/6 mice (a generous gift from Dr. Andreas Baumler, University of California-Davis) as previously described [56]. 24 h before infection, macrophages were seeded into 24-well tissue culture plate at a density of 1.5×105 cells/well in Dulbecco's modified Eagle medium supplemented with 10% fetal bovine serum (Hyclone), 15% L-cell conditioned medium, 1 mM sodium pyruvate and 2 mM glutamate.
Y. pseudotuberculosis strains were grown at 28°C in LB broth with aeration overnight. The next day, overnight cultures were diluted 1∶40 into fresh LB broth containing 2.5 mM CaCl2 and sub-cultured at 37°C for 2 h to induce yop gene expression. Bacteria were washed once and resuspended in HBSS to obtain optical density at OD 600 nm. Next, bacteria were diluted into cell culture medium to infect macrophages at an MOI of 10, unless specified. After centrifugation for 5 min at 700 rpm to facilitate bacterial contact with macrophages, another 15 min incubation was performed at 37°C, giving the total infection time of 20 min. The end of 20 min incubation is considered as 0 h post infection. To eliminate extracellular bacteria, unless specified, the cells were then incubated in medium containing 8 µg/ml gentamicin for 1 h, and then maintained in fresh medium containing 4.5 µg/ml gentamicin until the end of incubation. When indicated, 40 ng/ml Toxin B (Calbiochem), 100 µM NSC23766 (Calbiochem), or 10 µg/ml TAT-C3 was added at 0 h post infection and maintained throughout the experiment. TAT-C3 was purified and kindly provided by Dr. Gloria Viboud, Stony Brook University [57].
At the time points indicated in the figures, the infected BMDMs were washed twice with HBSS, lysed and scraped with 500 µl 0.1% Triton X-100 in HBSS to release intracellular bacteria. After collecting the lysates, 500 µl HBSS was used to rinse the wells and collect any residual bacteria. The lysates and the wash were combined, serially diluted and spread on LB plates, and then incubated at 28°C for 2 days to enumerate output CFU.
The primary antibodies used were a cocktail of two monoclonal mouse anti-YopE antibodies designated 202 and 149 (unpublished data), the monoclonal mouse anti-YopH antibody designated 3D10 (a gift from Dr. Richard Siegel, NIH) diluted 1∶1000, a polyclonal rabbit anti-YopT antibody diluted 1∶500 [30], and a polyclonal rabbit anti-β-actin antibody (Cell signaling) diluted in 1∶1000. The secondary antibodies used were a goat anti-mouse antibody conjugated to IRD800 (Rockland) diluted 1∶5000 and a donkey anti-rabbit antibody conjugated to IRD800 (Rockland) diluted 1∶5000.
The protein samples were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis, transferred to nitrocellulose membranes, and subjected to immunoblotting with specific primary and secondary antibodies. The membranes were then scanned and analyzed with the Odyssey system (Li-Cor Biosciences).
BMDMs were prepared and infected as described above, except that they were seeded into wells with glass coverslips, which had been washed with acetone and heated at 180°C for 4 h to remove LPS. At indicated time points, infected BMDMs were washed three times with PBS and fixed with 2.5% PFA for 10 min. When needed, 0.5 mM Isopropyl-β-D-thiogalactopyranoside (IPTG) was added at 1 h before fixation to induce de novo GFP expression, or at 2 h before fixation to induce de novo mCherry expression. When indicated, immunofluorescence staining was performed as described previously [23]. Briefly, the fixed cells were permeabilized with 0.1% Triton X-100 in PBS for 1 min, followed by blocking with 3% bovine serum albumin in PBS for 10 min. The cells were then incubated with a polyclonal rabbit anti-Yersinia antibody SB349 diluted 1∶1000 for 30 min. After washing with PBS, the cells were incubated with FITC conjugated anti-rabbit antibody (Jackson Laboratories) diluted 1∶250 for 40 min. After washing, the coverslips were inverted onto 6 µl Prolong Gold anti-fade reagent (Invitrogen) on a microscope slide. The slides were examined by fluorescence microscopy using a Zeiss Axioplan2 microscope with a 32× objective. Three randomly selected fields of each slide were examined. In each field, about 50 BMDMs were examined from merged images of phase contrast and fluorescence, which were captured with a Spot camera (Diagnotic Instruments, Inc) and processed with Adobe Photoshop. Percentage of cells containing bacteria was quantified using three independent experiments.
Detergent extraction assays were performed as previously described in [58]. BMDMs were infected as described above, except that they were seeded in 6 well plates at a density of 8×105 cells/well and infected at an MOI of 30 for 2 h. The infected cells were washed twice with ice-cold HBSS and lysed with 50 µl 1% Triton X-100 in HBSS containing EDTA-free protease inhibitor cocktail (ROCHE). After 10 min on ice, the cells were scraped from the plate to collect the lysates. The soluble and insoluble fractions of the lysates were separated by centrifugation (14000 rpm, 10 min, 4°C) and subsequently analyzed using immunoblotting as described above.
Chromosome DNA was isolated from C57BL/6 wild-type or SytVII−/− mouse tails and used as templates for PCR amplification. Briefly, tail tips were digested in 500 µl lysis buffer (0.1M NaCl, 0.05M Tris-HCL pH 7.7, 1% SDS and 2.5 mM EDTA) with 40 µg/ml freshly added proteinase K (Sigma), and incubated at 55°C overnight. The resulting supernatant were collected and mixed with 500 µl isopropanol to precipitate chromosomal DNA. After centrifugation (14000 rpm, 10 min, RT), the pellets were washed twice with 70% ethanol, air-dried for 5 min, and dissolved in 100 µl TE buffer. Genotyping PCR were performed with the following primers: P1 (5′-CATCCTCCACTGGCCATGAATG-3′), P2 (5′-GCTTCACCTTGGTCTCCAG-3′), P3 (5′-CTTGGGTGGAGAGGCTATTC-3′) and P4 (5′-AGGTGAGATGACAGGAGATC-3′). PCR products were analyzed by agarose gel electrophoresis.
NO levels generated by infected macrophages were determined by measuring the accumulation of nitrite (NO2−) using the Griess assay as described previously [59]. Control macrophages were treated with E.coli LPS (100 µg/µl, Sigma) and IFN γ (0.1 units/µl, ROCHE) throughout the experiment. At 23 h post infection, conditioned medium were collected and centrifuged (14000 rpm, 10 min, RT). 100 µl of the supernatant was mixed with 100 µl Griess reagent (0.5% sulfanilamide and 0.05% N-(1-naphthyl)ethylenediamide in 2.5% acetic acid) and incubated for 10 min at room temperature. The samples were then measured at OD550 nm. The concentration of NO2− was calculated by using a standard curve prepared with sodium nitrite.
The GenBank accession number for the YopE protein studied in this work is CAA68609.1.
|
10.1371/journal.pgen.1003684 | A Model-Based Analysis of GC-Biased Gene Conversion in the Human and Chimpanzee Genomes | GC-biased gene conversion (gBGC) is a recombination-associated process that favors the fixation of G/C alleles over A/T alleles. In mammals, gBGC is hypothesized to contribute to variation in GC content, rapidly evolving sequences, and the fixation of deleterious mutations, but its prevalence and general functional consequences remain poorly understood. gBGC is difficult to incorporate into models of molecular evolution and so far has primarily been studied using summary statistics from genomic comparisons. Here, we introduce a new probabilistic model that captures the joint effects of natural selection and gBGC on nucleotide substitution patterns, while allowing for correlations along the genome in these effects. We implemented our model in a computer program, called phastBias, that can accurately detect gBGC tracts about 1 kilobase or longer in simulated sequence alignments. When applied to real primate genome sequences, phastBias predicts gBGC tracts that cover roughly 0.3% of the human and chimpanzee genomes and account for 1.2% of human-chimpanzee nucleotide differences. These tracts fall in clusters, particularly in subtelomeric regions; they are enriched for recombination hotspots and fast-evolving sequences; and they display an ongoing fixation preference for G and C alleles. They are also significantly enriched for disease-associated polymorphisms, suggesting that they contribute to the fixation of deleterious alleles. The gBGC tracts provide a unique window into historical recombination processes along the human and chimpanzee lineages. They supply additional evidence of long-term conservation of megabase-scale recombination rates accompanied by rapid turnover of hotspots. Together, these findings shed new light on the evolutionary, functional, and disease implications of gBGC. The phastBias program and our predicted tracts are freely available.
| Interpreting patterns of DNA sequence variation in the genomes of closely related species is critically important for understanding the causes and functional effects of nucleotide substitutions. Classical models describe patterns of substitution in terms of the fundamental forces of mutation, recombination, neutral drift, and natural selection. However, an entirely separate force, called GC-biased gene conversion (gBGC), also appears to have an important influence on substitution patterns in many species. gBGC is a recombination-associated evolutionary process that favors the fixation of strong (G/C) over weak (A/T) alleles. In mammals, gBGC is thought to promote variation in GC content, rapidly evolving sequences, and the fixation of deleterious mutations. However, its genome-wide influence remains poorly understood, in part because, it is difficult to incorporate gBGC into statistical models of evolution. In this paper, we describe a new evolutionary model that jointly describes the effects of selection and gBGC and apply it to the human and chimpanzee genomes. Our genome-wide predictions of gBGC tracts indicate that gBGC has been an important force in recent human evolution. Our publicly available computer program, called phastBias, and our genome-wide predictions will enable other researchers to consider gBGC in their analyses.
| Gene conversion is the nonreciprocal exchange of genetic information from a ‘donor’ to an ‘acceptor’ sequence, primarily resulting from the repair of mismatched bases in heteroduplex recombination intermediates during meiosis [1]. In many cases, the process of resolving mismatches between G/C (guanine or cytosine; denoted ‘strong’ or ‘S’) and A/T (adenine and thymine; ‘weak’ or ‘W’) alleles appears to be biased in favor of S alleles [1]–[3]. Such GC-biased gene conversion (gBGC) elevates the fixation probabilities for S alleles relative to W alleles at positions of W/S polymorphism, and, if it acts in a recurrent manner over a sufficiently long time, can result in a significant excess of W→S over S→W substitutions and a consequent increase in equilibrium GC content. It has been known since the 1980s both that gene conversion occurs in various eukaryotes [4] and that mismatch repair can be significantly biased [5]. As complete genome sequences have become widely available, evidence has accumulated that gBGC may have played an important role in genomic evolution across many branches of the tree of life. In particular, it has been argued that gBGC has significantly influenced the genomic distribution of GC content, the fixation of deleterious mutations, and rapidly evolving sequences in many species [6]–[13].
Aside from limited experimental evidence of a GC-bias in meiosis, mostly from yeast [14], much of what is known about gBGC comes from two indirect sources of information: global patterns of variation within or between species suggesting a fixation bias favoring S alleles [11], [12], [15]–[17] and the existence of numerous loci exhibiting dense clusters of substitutions with a pronounced W→S bias [7]–[9], [13]. Both types of evidence correlate strongly with recombination rates, consistent with the hypothesis that they are caused by gBGC, although other recombination-associated factors might also contribute [16]. However, these observations provide limited information about the general prevalence, strength, and functional consequences of gBGC in humans and other mammals. Genome-wide patterns of variation are influenced by diverse forces that act in a highly heterogeneous manner across the genome, and it is difficult to measure the specific contribution of gBGC to these patterns. Clusters of biased substitutions perhaps provide more direct evidence of a local influence from gBGC. However, such clusters have so far been identified by considering either genomic windows of fixed size or pre-identified genomic segments (such as protein-coding exons or fast-evolving noncoding regions), which has limited the regions that can be detected. In addition, many studies have considered only fairly small numbers of clusters showing extreme substitution rates and W→S biases.
For modelers of molecular evolution, gBGC is an anomaly—a process separate and distinct from the fundamental processes of mutation, recombination, drift, and selection that underlie most models, yet one with the potential to profoundly influence patterns of variation within and between species. Like selection, gBGC acts in the window between the emergence of genetic polymorphism due to mutation and its elimination due to the fixation or loss of derived alleles. Unlike selection, however, gBGC is neutral with respect to fitness. The influence of gBGC at individual nucleotides can be modeled approximately by treating it as a selection-like force that depends only on whether a new mutation is W→S, S→W, or neither [13], [16], [18]. However, this approach ignores the close association of gBGC with the notoriously difficult-to-model process of recombination, which leads to a complex correlation structure along the genome (i.e., gBGC “tracts” separated by regions of no gBGC). Owing to these difficulties, with a few exceptions [9], [13], [19], gBGC has generally been ignored in phylogenetic or population genetic models, and considered at most in post hoc analyses (e.g., by examining identified genomic regions for an excess of W→S substitutions). These approaches are clearly limited in efficiency and effectiveness, and there is a need for improved models of gBGC that can be applied on a genome-wide scale. There is also a need for high quality annotations of gBGC-affected regions that can be used by investigators in other comparative and population genomic analyses.
Another reason to develop improved models of gBGC is that gBGC-induced nucleotide substitutions provide a unique window into historical recombination processes, by serving as a proxy for average recombination rates along a lineage of interest. By contrast, the other main sources of information about recombination—sperm typing [20], genotypes for known pedigrees [21], and patterns of linkage disequilibrium in present-day populations [22]—provide information about recombination that goes back no farther than the coalescence time between individuals. Pronounced differences between the human and chimpanzee recombination maps suggest that recombination rates in hominoids have changed rapidly [23]–[25]. gBGC may provide useful information about the recombination processes during the critical period between the divergence of humans and chimpanzees (4–6 million years ago [Mya]) and the coalescence time for human individuals (roughly 1 Mya, on average). Notably, archaic hominin genome sequences are of limited use for this purpose, because they are still few in number and result in only a modest increase in coalescence times.
In this article, we address these issues by introducing a novel model-based approach for the identification of gBGC tracts. Our approach makes use of statistical phylogenetic models that jointly consider gBGC and natural selection [13]. In addition, it approximates the recombination-associated correlation structure of gBGC along the genome using a hidden Markov model. We have implemented this approach in a computer program called phastBias, which is available as part of the open-source PHylogenetic Analysis with Space/Time models (PHAST) software package (http://compgen.bscb.cornell.edu/phast) [26]. Using simulations, we show that phastBias can identify tracts of various lengths from unannotated multiple alignments with good power. We then analyze genome-wide predictions of gBGC tracts in the human and chimpanzee genomes, comparing them with recombination rates, patterns of polymorphism, functional elements, fast-evolving sequences, and other genomic features. This analysis sheds light on the prevalence and fitness consequences of gBGC, and on recombination processes during the time since the human/chimpanzee divergence. Our predictions of gBGC tracts are freely available as browser tracks (http://genome-mirror.bscb.cornell.edu). We anticipate that these tracks will be useful for avoiding false positives in scans for positive selection, understanding the evolution of specific loci, and investigating the broader evolutionary forces shaping the human genome.
We model gBGC tracts using a phylogenetic hidden Markov model (phylo-HMM) with four states, representing all combinations of gBGC or no gBGC in a specified “target” genome (e.g., human or chimpanzee), and of evolutionary conservation or no evolutionary conservation across the phylogeny (Figure 1; Methods). The phylo-HMM framework [27] allows the distinct rates and patterns of nucleotide substitution for each state to be described using a full statistical phylogenetic model, and it captures the pronounced correlations along the genomes in these patterns using a first-order Markov model. Our phylo-HMM can be thought of as a straightforward generalization of the two-state model used by the phastCons program for prediction of evolutionarily conserved elements [28] that additionally predicts gBGC tracts in the target genome. We directly consider evolutionary conservation together with gBGC because the dramatic reduction in substitution rates in functional elements would otherwise be a confounding factor in the identification of gBGC tracts. The model allows conserved elements and gBGC tracts to overlap or occur separately. The joint effects of gBGC and selection are modeled by treating gBGC as a selection-like force that specifically favors the fixation of G and C alleles, as in other recent work. In particular, the influence of selection is described using a population-scaled selection coefficient, , and the influence of gBGC is described using an analogous population-scaled GC-disparity parameter, (where is the effective population size) [13] (see also [16], [18]). The parameter measures the strength of gBGC, and values cause W→S substitution rates to increase and S→W substitution rates to decrease. A key feature of our approach is that it permits identification of gBGC tracts of any length based on characteristic substitution patterns, independent of predefined windows or genomic annotations.
Because the signal for gBGC in the data is typically quite weak, we make several assumptions to reduce the complexity of the model. Briefly, we model negative selection as uniformly decreasing evolutionary rates on all lineages, we ignore positive selection, and we assume that the disparity parameter is the same for all gBGC tracts. In addition, we pre-estimate the parameters describing the neutral phylogeny and evolutionary conserved elements using restricted models, we fix the tract-length parameter based on our prior expectation for tract lengths, and we treat the parameter as a “tuning” parameter to be set by trial and error (see summary of model parameters in Table 1). Our simulation study indicates that fairly high accuracy in tract prediction is possible despite these simplifying assumptions and approximations (see below and Methods for details). We have implemented our model in a program called phastBias in the PHAST package [26]. PhastBias makes use of existing features in PHAST for alignment processing, phylogenetic modeling, efficient HMM-based inference, and browser track generation.
While the absence of high-quality annotations of gBGC tracts makes it difficult to assess prediction accuracy, we are able to gain some insight into the performance of phastBias using simulated data. To make our simulated data as realistic as possible, we started with real genome-wide alignments, and simulated new human sequences only, using our phylogenetic model to define neutral and conserved sequences, and interspersed gBGC tracts of fixed lengths (see Methods). This strategy ensures that features such as variation in mutation rates, changes in equilibrium GC content, conserved elements, indels, alignment errors, and missing data are all retained in the nonhuman sequences. We used phastBias to predict human-specific tracts based on these partially simulated alignments and compared our predictions with the “true” tracts assumed during simulation. We found that the nucleotide-level false positive rate was always very low in these experiments (/bp, Figure S1), so we measured the specificity of our predictions using the nucleotide-level positive predictive value (PPV), defined as the fraction of all bases predicted to be in gBGC tracts that truly belong in gBGC tracts. As a measure of power, we used the nucleotide-level true positive rate (TPR), the fraction of bases in true gBGC tracts that were correctly predicted as being in tracts.
First, we explored the performance of phastBias on simulated gBGC tracts of various lengths, generated with several different values of the GC-disparity parameter (denoted ). Under our model, increasing produces tracts with more substitutions and greater GC bias in their substitution patterns. As expected, both our power to detect gBGC and the specificity of our predictions increases with the lengths of the true tracts and with (Figure 2). We found that power and specificity were both quite good for tracts of 1,000–1,500 bases or longer, provided gBGC is reasonably strong (). Current estimates of the lengths and GC-disparity of real gBGC tracts [8], [29] suggest that phastBias should have good power for many tracts (see Discussion).
Next, we examined how our choice of the tuning parameters for expected tract-length () and gBGC strength () influence prediction performance. We found that the performance of the method was not highly sensitive to the value of , so we decided to fix the expected tract length at 1 kilobase (kb) (by setting ) based on empirical evidence indicating that mammalian gene conversion tracts are approximately this size [1], [29]. By contrast, the choice of had a much stronger influence on the observed prediction performance. Power was highest for small values of , regardless of the value used to simulate the tracts () (Figure S2). However, this increase in power comes at only a modest cost in PPV, which remains fairly high (>90%) except when the elements are both short and under weak gBGC (e.g., mean length bases, ). These results suggest that phastBias is inherently somewhat conservative with its predictions, and that setting to a relatively low value helps to improve sensitivity for tracts having a range of true gBGC intensities, at minimal cost in specificity.
We applied phastBias to genome-wide alignments of the human, chimpanzee, orangutan, and rhesus macaque genomes, and used it to predict tracts in the human and chimpanzee genomes likely to have experienced gBGC since the divergence of these two species 4–6 Mya (see Methods). In separate runs, we selected either the human or the chimpanzee genome as the “target,” and we set the tuning parameter to values of 2, 3, 4, 5, and 10 (in increasing strength of gBGC). As expected from our simulation study, the number, lengths, and genomic coverage of the predicted tracts depend fairly strongly on the choice of . In particular, coverage decreases from more than 1% to 0.07% as is increased from 2 to 10 (Table 2). Because the tracts predicted with high are largely found within those predicted with lower (Table S1), and because a value of appears to result in good power while controlling false positives (see above), we will focus on the tracts predicted with for the remainder of the article. The absolute sensitivity of these predictions of course depends on unknown properties of true gBGC tracts, but our simulation experiments indicate that power is fairly good, at least for the subset of tracts 1 kb or longer with a reasonably pronounced GC-disparity (Figure 2).
With , the predictions for the human genome include 9,439 gBGC tracts covering 0.33% of the genome (Table 2). These predicted tracts average 1,018 bp in length (median 788 bp), consistent with our choice of , but they display a fairly broad length distribution (Figure 3), indicating that our choice of tuning parameters is not overly restrictive. Most predicted tracts contain exclusively or predominantly W→S substitutions (Figure S3). The statistics for the chimpanzee genome are similar, but in this case there are somewhat fewer tracts (8,677), their lengths are reduced (mean = 842 bp, median = 663 bp), and genomic coverage is about 25% lower (at 0.25%). The reduced coverage of the chimpanzee genome holds even if we consider only tracts that completely fall within regions of high-quality, syntenic alignment between the two genome assemblies. These differences between the human and chimpanzee predictions could reflect differences between species in the degree to which recombination events are concentrated in recombination hotspots [25] (see Discussion).
The human and chimpanzee predictions are broadly distributed across the two genomes, but show a clear tendency to cluster near the ends of chromosomes (Figure 4; Text S1, Figures S4 and S5), consistent with previous findings [12], [15], [30]. In human, the median distance from the nearest telomere is only about one third that observed for a set of GC-content-matched control regions (9.6 megabases (Mb) vs. an average of 30.4 Mb over 1000 replicates, ). Similarly, the median distance between tracts is less than one third that for the controls, even after merging tracts less than 1 kb apart to account for possible biases from the HMM-based prediction method (24.3 kb vs. an average of 86.0 kb, ). The chimpanzee predictions are similarly distributed. In human, there is an obvious cluster of predicted tracts near the centromere of chromosome 2, reflecting the telomeres of two ancestral chromosomes that fused at this site along the human lineage after the human/chimpanzee divergence [15], [31]. However, the tract density in this region is somewhat lower in human than in the orthologous telomeric regions in chimpanzee (Figure S6), consistent with a reduction in the human recombination rate following the fusion event [12], [15] (see Discussion).
Together, the human and chimpanzee tracts account for about 1.2% of all human/chimpanzee nucleotide differences apparent in our genome-wide alignments (435,729 differences). About half (214,195) of the nucleotide differences within the tracts can be confidently explained by W→S substitutions on either the human or chimpanzee lineage, of which slightly more than half (115,699) fall on the human lineage. Thus, even with our limitations in power, our predictions suggest a non-negligible influence of gBGC on overall levels of human/chimpanzee nucleotide divergence.
The predicted human gBGC tracts are substantially enriched for recombination hotspots from the HapMap project [32]: 1,228 (13%) overlap a hotspot, compared with an average of 796 for the GC-matched control regions (). In addition, the average recombination rate [33] within these tracts is more than twice the rate in the control regions (3.85 centimorgans per megabase (cM/Mb) vs. 1.61 cM/Mb, ; Table 3). A parallel analysis of the chimpanzee gBGC tracts based on the genome-wide recombination rate map from the PanMap Project [25] showed, similarly, that recombination rates in predicted gBGC tracts were more than twice as high as in control regions (Table 3). Pedigree-based human recombination maps [21] produced similar results (data not shown).
At fine scales, the human and chimpanzee tracts show a modest, but significant, degree of overlap (Figure 4): 605 (6.4%) of the human tracts directly overlap a chimpanzee tract, compared with an average of 86 for the control regions (). Shared recombination hotspots account for only a small minority (<1%) of the overlapping tracts. However, the correlation in tract locations between species is much stronger at broader scales. For example, if the fractions of nucleotides in gBGC tracts (“gBGC density”) are compared in orthologous genomic blocks of various sizes, the human/chimpanzee Pearson's correlation increases from for 10 kb blocks to for 100 kb blocks, and to for 1 Mb blocks (Figure S7). These observations mirror those for human and chimpanzee recombination rates, which correlate well at scales of 1 Mb or larger but much more poorly at finer scales [23]–[25].
To gain further insight into the conservation of the gBGC tracts, we mapped the human gBGC tracts to orthologous locations in the chimpanzee genome, and the chimpanzee tracts to orthologous locations in the human genome. We then compared the recombination rates in these “ortho-tracts” with those in control regions, as with the tracts directly predicted for each species. Unlike recombination hotspots [25], the predicted gBGC tracts do show significantly elevated recombination rates at orthologous positions in the other species (Table 3). However, these recombination rates are not nearly as elevated as those for the directly predicted tracts. An analysis of the correlation between gBGC tract densities and recombination rates within and between species yielded similar results. Human gBGC tract densities are significantly correlated with human recombination rates, and this correlation increases with block size. A similar pattern is present in chimpanzee. When these correlations are examined across species (e.g., human gBGC densities vs. chimpanzee recombination rates), they are weaker but still significant (Figure S8). Differences in recombination rates between species are modestly predictive of differences in gBGC densities ( at 1 Mb; Figure S9). In general, we find much stronger correlations of gBGC- and recombination-associated features within species than between species, but these features nevertheless exhibit residual correlations between species, probably because they reflect average recombination rates over millions of years (see Discussion).
In both human and chimpanzee, the predicted tracts show a weak positive correlation with GC-content on a megabase scale. This correlation is somewhat stronger for human (Pearson's correlation for 1 Mb blocks: ) than for chimpanzee (), mirroring observations of a stronger correlation of recombination rate with GC-content in human than in chimpanzee [25].
To shed light on the functional implications of gBGC, we examined the degree of overlap of the predicted human gBGC tracts with various sets of genomic annotations (listed in Methods). In comparison with the control regions, we found that the human gBGC tracts were significantly depleted for overlap with known protein-coding exons, core promoters (1 kb upstream of annotated transcription start sites), miscellaneous RNAs, LINEs and SINEs, while they were significantly enriched for overlap with introns, lincRNAs, and a collection of ChIP-seq-supported transcription factor binding sites (Figure S10). However, all of these enrichments and depletions were modest in magnitude, with fold-changes of about 0.8–1.3. Overall, the gBGC tracts appear to be fairly representative of sequences of the same GC content. It is possible that the depletion for gBGC tracts in protein-coding exons and promoters could result in part from strong purifying selection counteracting GC-biased fixation.
To distinguish between fixation- and mutation-related biases, we compared the derived allele frequencies at polymorphic W→S and S→W sites in the predicted tracts and control regions. To control for the possibility of an ascertainment bias from polymorphic sites at which the derived allele is present in the human reference genome, we performed this analysis twice: once with the original gBGC tracts, and once with predictions based on alignments in which polymorphic sites in the human genome had been masked with ‘N’s.
Based on pilot data from the 1000 Genomes Project [33] (YRI population), the predicted gBGC tracts displayed significantly elevated derived allele frequencies at sites of inferred W→S mutations compared with sites of inferred S→W mutations (W→S DAF skew of ; Figure 5A). This skew in DAFs was significantly greater than that observed genome-wide () or in recombination hotspots (; Figure 5B), and it was larger than observed in any of the 1000 control region replicates (). The tracts are also far more biased than any of the regions considered by Katzman et al. [17], which were identified using sliding windows of fixed size and likely contained a mixture of gBGC tracts and non-tracts. Results were similar for the CEU (W→S DAF skew of ) and CHB-JPT populations (). These results held for the tracts based on the polymorphism-masked alignments, although the magnitude of the skew was somewhat reduced in this case ( for YRI; Figure S11). Together, these results strongly indicate an on-going preference for the fixation of G and C alleles in the predicted gBGC tracts.
There is much less polymorphism data available for chimpanzees than for humans, but data for 10 individual chimpanzees from the PanMap project [25] indicates a similar ongoing fixation bias within the predicted chimpanzee tracts (Figure S12). As in human, the W→S DAF skew in the predicted chimpanzee tracts is significantly stronger than that observed in recombination hotspots. We also compared the W→S DAF skews of the tracts predicted for each genome and the “ortho-tracts” mapped from the other genome. As with recombination rates, we found that, in both species, the predicted tracts have significantly greater W→S DAF skews than the ortho-tracts (Figure 5B and Figure S12B). These findings are consistent with gBGC currently acting on a subset of our predicted tracts in association with transient, species-specific recombination hotspots.
Theoretical modeling has shown that gBGC, in principle, can overcome negative selection and result in the fixation of weakly deleterious alleles [3], [8], [10]. However, there is currently little direct empirical evidence of a contribution of gBGC to fixed or segregating deleterious alleles [11]. Our genome-wide tract predictions enabled us to investigate the link between gBGC and deleterious alleles by testing for enrichments for disease-associated genomic regions in gBGC tracts.
We examined the relationship between the gBGC tracts and four sets of putatively disease-associated genomic regions: 10,711 polymorphic sites from dbSNP annotated as “pathogenic” or “probable pathogenic” [34]; 43,952 polymorphic sites from the Human Gene Mutation Database (HGMD) [35] (see also [11]); 11,444 genomic regions from the Genetic Association Database (GAD) [36]; and 6,435,165 polymorphic sites with evidence of functional importance (classes 1–5) in RegulomeDB [37]. For the dbSNP pathogenic and HGMD comparisons, we considered sets of control regions that overlapped the same number of exonic SNPs as the gBGC tracts. This control is designed to avoid misleading findings of significance that simply reflect the GC content, exon coverage, and/or rates of polymorphism in the gBGC tracts, since these disease-associated region sets are mostly found in coding regions. Similarly, we used control regions matched to SNPs considered by RegulomeDB, since it only includes non-coding SNPs (Methods).
We found that the gBGC tracts overlapped significantly more putatively disease-related SNPs from the dbSNP, HGMD, and RegulomeDB collections, and significantly more of the GAD regions, than did the matched control regions (Table 4; for each). In the cases of the two collections of disease-associated SNPs (dbSNP and HGMD), the enrichment within the predicted gBGC tracts was particularly striking (fold-enrichments of 2.4 and 1.9, respectively), while in the other cases it was more modest but still significant. These results suggest that gBGC may contribute in important ways to elevated allele frequencies, and perhaps, to the eventual fixation of deleterious mutations.
Many fast-evolving regions of the human genome display an excess of W→S substitutions, leading to the suggestion that gBGC may play a role in their evolution [6], [7], [9], [13], [38], [39]. Supporting this hypothesis, our predicted gBGC tracts overlap 13 of the 202 (6.4%) HARs identified by Pollard et al. [38], more than observed for any of the 1000 GC-control region replicates (). Notably, the HARs overlapped by gBGC tracts included HAR1, HAR2, and HAR3, the three fastest evolving sequences in this set. We also examined an expanded set of 721 HARs [40] and found that gBGC tracts overlapped 75 of them (10%; ; see example in Figure 6). Next, we compared the gBGC tracts with 10 protein-coding genes identified as showing signatures of positive selection on the human branch based on a likelihood ratio test [41]. One of these genes is overlapped by a gBGC tract, significantly more than expected based on exon-aware controls (). The overlapped gene, ADCYAP1, was also highlighted by another group [9] as showing strong evidence of an influence from gBGC. We repeated our analysis with 157 genes identified in another recent study as showing signatures of human-specific positive selection [42], and found that the gBGC tracts overlapped 11 (7%) of these genes, somewhat more than average for the exon-aware control replicates (7.4, ). Considering our limitations in power (see Discussion), these results indicate the gBGC has contributed to a substantial fraction of fast-evolving sequences in the human genome.
Our predicted tracts for human and chimpanzee are available as a UCSC Genome Browser track at http://genome-mirror.bscb.cornell.edu (Figure 6). This track displays both our discrete predictions of gBGC tracts and a continuous-valued plot indicating the posterior probability that each position is influenced by gBGC. Using this track it is possible to browse the predicted tracts in their full genomic context, perform queries intersecting them with other browser tracks, and download them for further analysis. We expect this track to be particularly useful for other investigators who wish to exclude gBGC-influenced regions of the genome from other molecular evolutionary analyses, such as the identification of genes under positive selection. The tracts themselves will also be directly useful for studying the evolution of recombination rates and their relationship to substitution rates and patterns.
This paper describes an analysis of predicted gBGC tracts in the human and chimpanzee genomes, based on a new computational method called phastBias. PhastBias makes use of a hidden Markov model and statistical phylogenetic models that consider the influence of both natural selection and gBGC on substitution rates and patterns. Unlike previous methods for identifying signatures of gBGC, it does not depend on a sliding window or predefined annotations of protein-coding genes or conserved noncoding elements [9], [13], [15], [19], but instead can flexibly identify tracts of various sizes directly from genome-scale multiple alignments. The method appears to have good power for tracts of about 1 kilobase or longer, provided gBGC has acted with a reasonably high average intensity along the lineage of interest. Our predictions in the human and chimpanzee genomes cover about 0.3% of each genome and explain 1.2% of human/chimpanzee single nucleotide differences. Consistent with the hypothesis that they are caused by gBGC, the predicted tracts are correlated with recombination rates, tend to fall in subtelomeric regions, and exhibit an ongoing fixation bias for G and C alleles. In addition, they are enriched for disease-associated human polymorphisms, and they tend to overlap previously identified fast-evolving coding and non-coding regions, suggesting that gBGC has contributed significantly to both deleterious mutations and rapid sequence evolution. Overall, our analyses indicate that gBGC has been an important force in the evolution of human and chimpanzees since their divergence 4–6 million years ago.
Many attributes of the predicted gBGC tracts are consistent with the hypothesis that recombination is the driving force behind the observed patterns of biased substitution. Nevertheless, the tract locations are only partially correlated with recombination rates in human and chimpanzee. Moreover, while the tracts are enriched for recombination hotspots in both species, there are thousands of hotspots that do not overlap a gBGC tract, and the majority of tracts do not overlap a hotspot. These differences can be explained by several factors. First, the hotspots we have analyzed reflect recombination patterns in modern human populations, while the gBGC tracts reflect average patterns since the divergence of humans and chimpanzees. Many current hotspots presumably have not had sufficient time to produce a detectable signature of biased substitution, while many extinct hotspots contributed to gBGC for long periods of time in the past. Second, models of gBGC suggest that it can occur in conjunction with both crossover and noncrossover recombination events, but current recombination maps reflect crossover events only [3]. An imperfect correlation of these types of events, together with statistical noise in current estimates of crossover rates, likely accounts for some of the absence of correlation between recombination rates and gBGC tracts. Third, biased substitution rates are influenced by many factors other than recombination, such as mutation rates, natural selection, and GC content [43]. For example, strong purifying selection at or near a hotspot could eliminate the signature of gBGC. Finally, limitations in power for both recombination events and gBGC tracts undoubtedly reduce the apparent correlation between these features.
The locations of the human and chimpanzee tracts are strongly correlated on megabase scales, but, like recombination rates, they differ significantly on fine scales, and few human and chimpanzee tracts directly overlap one another (Figure 4; Figure S7). Nevertheless, even at fine scales, the human and chimpanzee gBGC tracts agree better than recombination hotspots, which are essentially uncorrelated between the two species [25]. This observation probably stems from the fact that gBGC tracts reflect time-averaged recombination rates, and historical recombination rates were presumably better correlated than those in present-day humans and chimpanzees. In general, the predicted gBGC tracts provide a valuable window into historical recombination processes, but this window is “blurred” by time-averaging over millions of years. Nevertheless, together with other sources of information about historical recombination processes—such as new methods based on patterns of incomplete lineage sorting (K. Munch, T. Mailund, J.Y. Dutheil, and M.H. Schierup, submitted)—predictions of gBGC tracts may help to provide a more detailed picture of the evolution of recombination rates in hominoids.
The different time scales associated with crossover-based recombination maps and our predicted gBGC tracts are particularly well illustrated by the region of the chromosome 2 fusion in human (Figure S6). Consistent with its location near a centromere in the human genome, this region displays no elevation of crossover rates in human populations, while the orthologous regions of the chimpanzee genome show elevated crossover rates typical of telomeres. Accordingly, this region exhibits little W→S DAF skew in human, but a clear skew in chimpanzee. However, the density of predicted gBGC tracts in this region is elevated in both species, only slightly more so in chimpanzee than human, suggesting that this region was telomeric for most of the approximately 6 million years during which human-specific recombination-associated substitutions could have occurred. Thus, our observations indicate that the fusion event is fairly old relative to intraspecies coalescence times but young relative to the human/chimpanzee divergence time. They are qualitatively consistent with Dreszer et al.'s [15] estimate of 0.74 Mya (95% confidence interval: 0–2.81 Mya) for the date of the fusion event and inconsistent with the argument that this event contributed to the initial speciation of humans and chimpanzees [44].
Despite the overall similarity of the human and chimpanzee predictions, the coverage of the predicted tracts is about 25% lower in the chimpanzee genome. A possible cause of this difference is the greater concentration of recombination events in hotspots in humans [25]. This difference could lead to a stronger population-level signal for gBGC in humans, allowing for more predictions and longer predicted tract lengths. It has been proposed that the difference in the concentration of recombination events may derive from differences in the activity of the hotspot-specifying protein PRDM9, which shows substantially greater allelic diversity in chimpanzees than in humans [25]. Consistent with this hypothesis, Auton et al. [25] found a much weaker signal for sequence motifs potentially involved in PRDM9 binding at chimpanzee hotspots than at human hotspots. In an attempt to shed light on the ancestral binding preferences of PRDM9, we applied motif discovery methods to the predicted gBGC tracts in the human and chimpanzee genomes. However, in both species this analysis turned up only a few motifs, none of which resembled the well-defined motifs reported for the human genome [25], [45]. This absence of strong motifs may occur because the ancestral recombination hotspots in both species are more like those in present-day chimpanzees than humans. Alternatively, it may simply reflect the difficulty of motif discovery given rapidly evolving PRDM9 binding preferences and the time-averaged nature of the gBGC tracts.
Given what is currently known about gBGC, it is impossible to obtain direct measurements of the completeness and accuracy of our predicted tracts. Our simulation experiments suggest that both sensitivity and specificity are reasonably good for tracts at least 1 kb in length with , but we often miss shorter or less biased gBGC tracts (Figure 2), and the true distributions of tract lengths and values are unknown (although average estimates of [46] and [8] have been reported for highly recombining regions). It is important to bear in mind that represents an average along an entire branch of the phylogeny. Many regions may have experienced quite strong gBGC but for short evolutionary intervals, resulting in small average values of and poor detection power. Thus, while our genome-wide predictions improve on what is currently available, it seems plausible that they still represent the “tip of the iceberg”—a relatively small subset of all genomic regions significantly influenced by gBGC, perhaps unusual for their length or GC-disparity.
It is worthwhile to consider two other indirect sources of information about our power for gBGC tract prediction. First, Katzman et al. [17] found that about 20% of the 40 kb genomic intervals they examined show significant W→S DAF skew. If we conservatively assume one 1–2 kb tract per gBGC-influenced window, this observation would imply that at least 0.5–1.0% of the human genome has been influenced by gBGC on population genomic time scales, compared with the phastBias estimate (for ) of 0.3%. Second, using a method optimized for the analysis of individual HARs, Kostka et al. [13] estimated that 24% of HARs experienced significant gBGC (19% exclusively and 5% in combination with positive selection), or 3.7 times as many as overlap our phastBias predictions (6.4%). Thus, these two imperfect indicators of power suggest that, with , phastBias underpredicts gBGC tracts by a factor of at least about 2–4. The genomic coverage of our predictions may be closer to the truth (1.1%; Table 2), but these predictions appeared to be of poorer quality on inspection, apparently because the phylo-HMM states with and without gBGC were insufficiently distinct to control false positive rates.
While the likelihood ratio tests of Kostka et al. [13] appeared to have greater power for gBGC in HARs overall, phastBias sometimes achieves improved sensitivity by considering the entire genome (including flanking sequences) rather than just a designated collection of elements. Indeed, of the thirteen HARs that overlap one of our gBGC tracts, three were not identified by Kostka et al., apparently for this reason. These instances of improved sensitivity are especially noteworthy given that phastBias must address the more difficult problem of unconstrained genome-wide prediction, with the attendant potential for large numbers of false positives predictions.
In principle, gBGC can overcome purifying selection and help to drive deleterious alleles to high frequencies [3], [8], [10], but it has been difficult to find direct empirical evidence for a reduction in fitness (genetic load) caused by gBGC. Our predicted gBGC tracts are significantly enriched for disease-associated polymorphisms in current human populations, suggesting that gBGC has helped to drive at least some of these alleles to appreciable frequencies, and, indeed, may still be active in maintaining them. We attempted to establish an orthogonal link between gBGC and deleterious alleles by looking for evidence of purifying selection in chimpanzees and other species at the locations of W→S substitutions within the predicted human tracts (Text S1). The idea behind this analysis was that, if a substantial number of these mutations were driven to fixation by gBGC despite negative selection against them, one would expect an excess of evolutionary conservation, a deficiency of polymorphisms, and/or a skew toward low-frequency derived alleles at orthologous locations in other species, relative to an appropriate control. However, this analysis yielded inconclusive results: the human tracts are significantly enriched for overlap with evolutionarily conserved elements at locations of W→S substitutions (Figure S13), but evolutionary conservation scores and chimpanzee polymorphisms do not display the expected patterns (Figures S14, S15, and S16). It seems likely that the signal for excess conservation in the gBGC tracts is simply too weak to detect by these methods, owing to the sparseness of functional sites within the tracts and the difficulty of establishing appropriate control regions. Nevertheless, it may be possible in future work to develop refined comparative genomic methods for measuring the genetic load associated with gBGC.
Our phylogenetic hidden Markov model has four states: one that assumes both evolutionary conservation and gBGC (), a second with gBGC but no conservation (), a third with conservation but no gBGC (), and a fourth with neither conservation nor gBGC () (Figure 1). To avoid over-parameterization, we make the following simplifying assumptions. First, we model gBGC only on the lineage leading to a pre-defined “target” genome (human or chimpanzee), because gBGC is expected to be a transient phenomenon, typically affecting a single lineage in any genomic position of interest. gBGC tracts are allowed to occur on other lineages, but these tracts are expected to have a negligible influence on inferences in the target genome and are not directly modeled. Second, negative selection, in contrast to gBGC, is assumed to apply uniformly across all branches of the phylogeny. Third, positive selection is ignored. We omit positive selection and lineage-specific negative selection from the model because they are expected to be fairly rare, to leave a relatively weak signal in the data at human-chimpanzee evolutionary distances [47], and to primarily operate at a somewhat different genomic scale from gBGC (e.g., at the level of individual binding sites or clusters of amino acids, rather than genomic tracts of hundreds or thousands of bases). We expect our modeling framework to be robust to occasional sequences under positive or lineage-specific selection, because the primary signal for tract prediction is a W→S substitution bias, and selection generally will not produce such a bias consistently across many bases. Finally, we assume that the strength of gBGC and the strength of negative selection in the target genome are constant across the genome. A similar homogeneity assumption is employed in phastCons and appears to have a minimal impact on power and accuracy for element identification [28].
With these assumptions, the phylogenetic models for the four states are defined as follows (with further mathematical details given in Text S1).
The state-transition probabilities are defined by four parameters, denoted , , , and (Figure 1, Table 1). The parameters and are inherited from phastCons [28] and describe the conditional probabilities of transitioning from a conserved state to a neutral state, and from a neutral state to a conserved state, respectively. The parameters and are analogous, defining the conditional probabilities of transitioning out of, and into, a gBGC tract, respectively. The sixteen possible state transition probabilities are obtained by multiplying the appropriate pairs of conditional probabilities and enforcing the standard normalization constraints (Figure 1). This “cross-product” construction corresponds to a prior assumption of independence for the two types of transitions (conservation no conservation and gBGC no gBGC).
Given a multiple sequence alignment, standard algorithms for statistical phylogenetics and hidden Markov models can be used to calculate the likelihood of the data under this model, to predict the most likely state path (Viterbi), or to calculate the marginal posterior probability of each state at each alignment column (reviewed in [27]).
In principle, the nine free parameters in our model (Table 1) could all be estimated directly from the data by maximum likelihood, using an expectation maximization or numerical optimization algorithm. In practice, however, parameter estimation is difficult because there are no validated gBGC tracts to use for supervised training of the model, and the signal in the data is not sufficiently strong to support a fully unsupervised estimation procedure. Instead, we partition the parameters into three groups: those for the neutral substitution process, those for the model of conserved elements, and those specific to the gBGC tracts. The first two groups of parameters are pre-estimated from the data without consideration of gBGC, by what can be considered an empirical Bayes approach. The parameters in the third group are then estimated by a combination of methods.
Specifically, the free parameters for the neutral substitution process (, , and ) are estimated per alignment block (see below) using phyloFit [26], after conditioning on the tree topology and branch-length proportions (as described above). This strategy assumes that conserved elements and gBGC tracts are sparse and have at most a minor effect on average substitution rates for large genomic blocks. The three additional parameters that describe conserved elements (, , and ) are inherited directly from phastCons and therefore were simply set to the values used for the Conservation tracks in the UCSC Genome Browser. The remaining parameters include the GC-disparity and the gBGC transition probabilities and . As discussed in the Results section, we found that —which can be interpreted as an inverse prior expected length for gBGC tracts—has only a weak influence on our predictions (within a reasonable range) and decided to simply fix it at 1/1000, corresponding to a prior expectation of 1 kb tracts. We treated as a “tuning” parameter and considered various possible values in a plausible range. The final parameter, , was estimated from the data (separately for each alignment block) by expectation maximization, conditional on fixed values of all other parameters.
To predict gBGC tracts based on our model, we computed marginal posterior probabilities for the four model states at each genomic position using the forward/backward algorithm. We then computed the marginal posterior probability of gBGC by summing the probabilities for states and , and we predicted tracts by applying a threshold of 0.5 to this probability (i.e., the predicted tracts are maximal segments in which every position has a posterior probability of at least 50% of gBGC). We settled on this strategy after discovering that the more conventional Viterbi algorithm performed poorly in this setting, evidently due to uncertainty about the endpoints of tracts. This uncertainty causes the probability mass for a putative gBGC tract to be distributed across many possible HMM state paths, and as a result, the Viterbi algorithm often fails to predict a tract even when the posterior probability of gBGC is close to one. A potential drawback of our thresholding strategy is that fluctuating posterior probabilities could lead to highly fragmented tract predictions. However, we found that the posterior probability function was quite smooth in practice (probably owing to small values of the state transition probabilities) and fragmentation was not a problem. For example, at , only about 2% of the predicted human tracts fall within 50 base pairs of another tract. Nonetheless, when analyzing the genomic distribution of gBGC tracts relative to one another and to telomeres, we merged adjacent tracts (within 1 kb) in order to reduce any bias introduced by over fragmentation (Text S1).
Our analyses of both simulated and real data were based on genome-wide alignments obtained from the UCSC Genome Browser (http://genome.ucsc.edu) [49]. We began with the 44-way vertebrate alignments produced with multiz [50] (hg18 assembly) and extracted the rows corresponding to the human, chimpanzee, orangutan, and rhesus macaque genomes, discarding alignment columns containing only gaps in these sequences. We also discarded columns in which the human genome contained a gap. Human-referenced alignments were used for both the human and chimpanzee gBGC tract predictions, as chimpanzee-based multiple alignments are not available. For convenience in processing, the resulting four-way alignments were partitioned into blocks of approximately 10 megabases (Mb) in length. The boundaries between blocks were required to occur in regions uninformative about gBGC (due to greater than 1 kb with lack of alignment with the other species). We experimented with several alternative block sizes, ranging from 1–30 Mb, and found that the predictions were fairly robust to the choice of block size (Table S2).
We simulated human sequences with gBGC tracts for each 10 Mb block in the real genome-wide alignments as follows. First, we identified positions at which any sequence contained a CpG dinucleotide, because substitution rates are likely to be substantially elevated at such sites. Next, we used phastCons to identify conserved elements in the four species. We then fitted a phylogenetic model to the alignment columns in each of four categories (neutral/non-CpG, conserved/non-CpG, neutral/CpG, conserved/CpG) by estimating , , and for the most data-rich category (neutral/non-CpG), then estimating a separate for the CpG category (using phyloFit) and applying a branch-length scale-factor of 0.31 to the conserved categories. Next, we defined an alternative “gBGC” instance of each of the four estimated models by modifying the substitution rate matrix for the human branch according to our model of gBGC [13] and a given choice of (here denoted ). In this way, we obtained eight phylogenetic models, representing all combinations of conservation/no conservation, CpG/no CpG, and gBGC/no gBGC.
We generated synthetic human sequences by assigning one of these eight models to each alignment column, as follows. The conservation and CpG status of each column was maintained as originally annotated, so that the synthetic alignments would resemble the original ones as much as possible. The gBGC status was set to “no gBGC” for most columns, but set to “gBGC” for tracts of fixed size at randomly selected locations, at an average gBGC coverage of 0.1%. We then simulated a new human base for each alignment column conditional on the assigned phylogenetic model and the observed chimpanzee, orangutan, and rhesus macaque bases. This was accomplished using the ‘postprob.msa’ function in RPHAST, which computes the marginal distribution over bases at any node in the phylogeny conditional on a given phylogenetic model and collection of observed bases, using the sum-product algorithm. This function computes the desired distribution for the human base if the human sequence is masked and treated as missing data in the input. A particular base was selected by sampling from this marginal distribution.
We performed this simulation procedure for combinations of and fixed tract lengths of 200, 400, 800, 1600, 3200, and 6400. For each set of simulated alignments, we predicted gBGC tracts as described in the previous section, assuming several different values for the tuning parameter . For each data set and value of , we calculated the true positive rate (number of correctly predicted gBGC bases/total number of gBGC bases), false positive rate (number of incorrectly predicted gBGC bases/total number of non-gBGC bases), and positive predictive value (number of correctly predicted gBGC bases/number of predicted gBGC bases).
We compared the predicted gBGC tracts with exon and intron definitions from Gencode version 3c and Ensembl genes [51], and with annotations of lincRNAs, miRNAs, miscRNAs, small non-coding RNAs, NMD transcripts, and pseudogenes from Gencode version 14 [52]. We also compared them with LINE and SINE elements from the rmskRM327 table in the UCSC Table Browser [53], and with a set of high-confidence predictions of transcription factor binding sites based on ChIP-seq data from ENCODE [54]. In addition, we compared the tracts with genome-wide recombination rate estimates from the 1000 Genomes Project [33], recombination hotspots from the October 2006 release of HapMap [32], and chimpanzee recombination rate estimates from the PanMap project [25].
Disease-associated SNPs were obtained from several sources. SNPs annotated with “pathogenic” or “probable pathogenic” clinical significance were downloaded on October, 25, 2011 from dbSNP [34]. The HGMD dSNPs were obtained from the Supplementary Material of reference [11]. Regions of the human genome with positive genetic associations with disease were taken from the Genetic Association Database [36] on February 2, 2012. The level of evidence for the function of non-coding SNPs was downloaded from the RegulomeDB [37] web site on December 12, 2012. All data not in reference to the GRCh36/hg18 assembly were mapped to hg18 using the ‘liftOver’ tool from the UCSC Genome Browser.
To evaluate the statistical significance of various properties of interest, we compared the predicted gBGC tracts with sets of control regions matched to them in number, length distribution, and chromosome assignment. We also ensured that the control regions were matched to the gBGC tracts by GC content (by stratifying predictions and controls into 100 bins), which is known to correlate strongly with several relevant genomic features. We obtained a null distribution for each statistic of interest (such as the number of tracts overlapping exons, or the number human tracts overlapping orthologous chimpanzee tracts), by computing a value of the statistic for each of 1000 randomly sampled replicates of the control regions. One-sided empirical p-values were computed as the fraction of sampled control sets for which the statistic was at least as extreme as observed in the predicted tracts. As noted in the text, we occasionally considered alternative sets of control regions designed to accommodate known biases in genomic regions of interest. For example, when evaluating the significance of overlap with disease-associated SNPs from HGMD and dbSNP, we used control regions matched to the predicted tracts in terms of their degree of exon overlap, since these sets consist mostly of coding SNPs. Similarly, for RegulomeDB, which is focused on non-coding SNPs, we used control regions that matched the overlap of the gBGC tracts with the set of SNPs considered by RegulomeDB.
Our analysis of human derived allele frequencies was based on genotype data and ancestral allele predictions from the low-coverage pilot data set from the 1000 Genomes Project released in July 2010 [33]. These comprise SNP calls for the 22 autosomes in three HapMap population panels: YRI (59 individuals), CEU (60 individuals), and CHB-JPT (60 individuals). The chimpanzee derived allele frequency analysis was based on genotype data for 10 individuals downloaded from the PanMap project [25]. SNP locations were mapped to the human genome, and the 1000 Genomes predicted human chimpanzee ancestral allele was used to identify the derived allele. Sites with a low quality genotype call (GQ quality score less than 5), more than two alleles, or no predicted ancestral allele were not considered. We computed the W→S DAF skew of all human and chimp gBGC tract SNPs as normalized values from a Mann-Whitney test on the derived allele frequencies of W→S and S→W SNPs, as previously described [17]. A W→S DAF skew of 0.5 indicates no bias, and values greater than 0.5 indicate that W→S mutations are favored.
|
10.1371/journal.pntd.0004322 | Intradermal Immunization of Leishmania donovani Centrin Knock-Out Parasites in Combination with Salivary Protein LJM19 from Sand Fly Vector Induces a Durable Protective Immune Response in Hamsters | Visceral leishmaniasis (VL) is a neglected tropical disease and is fatal if untreated. There is no vaccine available against leishmaniasis. The majority of patients with cutaneous leishmaniasis (CL) or VL develop a long-term protective immunity after cure from infection, which indicates that development of an effective vaccine against leishmaniasis is possible. Such protection may also be achieved by immunization with live attenuated parasites that do not cause disease. We have previously reported a protective response in mice, hamsters and dogs with Leishmania donovani centrin gene knock-out parasites (LdCen-/-), a live attenuated parasite with a cell division specific centrin1 gene deletion. In this study we have explored the effects of salivary protein LJM19 as an adjuvant and intradermal (ID) route of immunization on the efficacy of LdCen-/- parasites as a vaccine against virulent L. donovani.
To explore the potential of a combination of LdCen-/- parasites and salivary protein LJM19 as vaccine antigens, LdCen-/- ID immunization followed by ID challenge with virulent L. donovani were performed in hamsters in a 9-month follow up study. We determined parasite burden (serial dilution), antibody production (ELISA) and cytokine expression (qPCR) in these animals. Compared to controls, animals immunized with LdCen-/- + LJM19 induced a strong antibody response, a reduction in spleen and liver parasite burden and a higher expression of pro-inflammatory cytokines after immunization and one month post-challenge. Additionally, a low parasite load in lymph nodes, spleen and liver, and a non-inflamed spleen was observed in immunized animals 9 months after the challenge infection.
Our results demonstrate that an ID vaccination using LdCen-/-parasites in combination with sand fly salivary protein LJM19 has the capability to confer long lasting protection against visceral leishmaniasis that is comparable to intravenous or intracardial immunization.
| Leishmaniasis is a disease with a wide spectrum of clinical manifestations caused by different species of protozoa belonging to the Leishmania genus that are transmitted by sand fly vectors. Visceral infections of Leishmania cause significant mortality and morbidity and development of a vaccine to prevent leishmaniasis has become a high priority. We have previously reported that intravenous immunization with a live attenuated parasite vaccine comprised of Leishmania donovani parasites lacking the centrin gene conferred protection in mice, hamsters and dogs. In the current report, we describe the immunological response and associated protection to the ID immunization with attenuated parasites in combination with a sand fly salivary protein (LJM19). We observe that protection against experimental ID challenge with L. donovani resulting from ID immunization with live attenuated parasites in combination with LJM19 is comparable to intracardial immunization and offers improved protective immunity compared to immunization with salivary protein alone and non-immunized hamsters. This study supports the potential use of the genetically attenuated vaccine and a recombinant sand fly salivary protein for control of visceral leishmaniasis.
| Leishmaniasis is a disease with a wide spectrum of clinical manifestations caused by different species of protozoa belonging to the Leishmania genus that are transmitted by sand fly vectors [1]. The disease causes high morbidity and significant mortality throughout the world, where 350 million people in 98 countries are at risk of contracting the infection. Moreover, approximately 1.0 to 1.5 million cases of cutaneous leishmaniasis (CL), and 200,000 to 500,000 cases of visceral leishmaniasis (VL), are registered annually [2].
VL is fatal if not treated [2]. The treatment of leishmaniasis is still based on the use of the parenteral administration of pentavalent antimonial compounds. However, side effects associated with the treatment and increased parasite resistance have made control and elimination of VL a serious challenge [3,4]. Therefore, the development of new strategies to prevent leishmaniasis has become a high priority [5].
The development of a vaccine for VL has been the focus of several research groups. Among the various types of vaccines, genetically modified live-attenuated vaccines provide the immunized host with diverse and complex antigens and induce a potent protective immunity in murine models [5,6]. Importantly live attenuated parasites cause no pathology in experimental infections [7–14], while inducing protection reflected by a significant reduction of parasite burden in animals challenged with virulent wild type strains [10,12,14–18].
We have previously reported on the LdCen-/- parasites as a live attenuated candidate vaccine in several animal models [12,14,18]. Infection with LdCen-/- was non-pathogenic i.e., safe and highly immunogenic in mice, hamsters and dogs [12,13,18]. In addition, immunization with LdCen-/- induced protection against homologous challenge with wild type L. donovani and conferred cross-protection against infection with a heterologous challenge with L. braziliensis, L. mexicana and L. infantum [14,18]. However, previous studies with LdCen-/- parasites as immunogens were performed without any adjuvants. Since the adjuvants can activate a range of innate immune pathways it is difficult to predict on an empirical basis which adjuvant will work most effectively with live attenuated parasites. Since the adaptive response is the primary determinant of protective immunity generated by vaccination, immunomodulatory reagents that could supplement LdCen-/- induced immunity without causing rapid elimination of the vaccine antigen due to innate immune reactions could make LdCen-/- more effective as an anti-Leishmania vaccine.
Saliva from sand flies contains potent pharmacologic components that facilitate blood meal acquisition and modulates the host inflammatory and immune responses [19,20]. Arthropod vector saliva also plays an important role in pathogen transmission from the sand fly to the vertebrate host [21]. Recent reports have shown the importance of some salivary proteins from sand fly vectors such as LJM19, LJM11 or LJM17 as potential targets for vaccine development against Leishmania infection [11,19,22–30]. A specific immune response against salivary proteins has been reported in various animal models. For example, hamsters immunized with plasmid DNA coding for LJM19, a Lu. longipalpis salivary protein, protected them from disease after challenge with wild type Leishmania infantum chagasi parasites plus saliva through the induction of a LJM19-specific immune response [26]. By comparison, salivary protein LJM11 provided partial protection that was not long lasting against virulent challenge [26]. Importantly, immunization with LJM19 induced higher ratios of IFN-γ/IL-10 and IFN- γ/TGF-β in the spleen, conditions consistent with a Th1 polarization [26,28]. These results suggested that salivary gland proteins such as LJM19 could be a potent supplement to the protective immunity with live attenuated Leishmania parasites.
In our previous studies, we have tested different routes of immunization, including intravenous (tail vein), intracardial and subcutaneous [12,13]. However, intradermal immunization can offer improved protective immunity and simplify the logistics of delivery as was previously demonstrated [26,28,29,31,32]. Therefore, it is of value to evaluate live attenuated parasite vaccines for their efficacy following intradermal immunization.
Since our previous studies have shown that exposure to live attenuated parasites injected by intravenous or intracardial routes and without an adjuvant induced a strong protective immunity, we asked whether an intradermal immunization with LdCen-/- parasites in combination with LJM19 could further enhance vaccine induced protection. In the present study, we report for the first time the immunogenicity and protection outcome in hamsters intradermally primed with salivary protein LJM19 and boosted with genetically modified live attenuated L. donovani parasites (LdCen-/-) in combination with recombinant LJM19. Immunized hamsters demonstrated a strong immune response comparable to that of intracardial immunization with LdCen-/- and resulted in long term protection against infection with virulent L. donovani parasites.
Two-month-old female Syrian golden hamsters (Mesocricetus auratus) were obtained from the Harlan Laboratories and kept in the Food and Drug Administration (FDA) animal facility. The experimental procedures used in this study were reviewed and approved by the Animal Care and Use Committees of the FDA and the National Institute of Allergy and Infectious Diseases (NIAID).
The animal protocol for this study has been approved by the Institutional Animal Care and Use Committee at the Center for Biologics Evaluation and Research, US FDA (ASP 1995#26). Further, the animal protocol is in full accordance with ‘The guide for the care and use of animals’ as described in the US Public Health Service policy on Humane Care and Use of Laboratory Animals 2015 (http://grants.nih.gov/grants/olaw/references/phspolicylabanimals.pdf).
Lu. longipalpis sand flies, Jacobina strain were reared at the Laboratory of Malaria and Vector Research, NIAID. Salivary glands were dissected from 5- to 7-day-old females and stored in PBS at -70°C. Before use, salivary glands were sonicated and centrifuged at 12,000×g for 2 min. The supernatant was collected and used immediately.
LdCen-/- and L. donovani (Ld1S) promastigotes were grown at 26°C in medium 199 supplemented with 20% FCS. Three- to 4-month-old hamsters were immunized intradermally in the ear at two-week intervals between immunizations, using a 29-gauge needle (BD Ultra-Fine) in a volume of 20μl, using the following protocols. Group 1: prime with 2μg of LJM19 protein and boost with 107 stationary phase LdCen-/-promastigotes plus 2μg of LJM19. Group 2: 107 stationary phase LdCen-/- promastigotes via intracardial injection. Group 3: 2μg of LJM19 protein (two times). Group 4: BSA (control) (S1 Fig). Each experimental group consisted of 6 hamsters. Five weeks after the last immunization, the animals were challenged ID with 105 stationary phase Ld1S promastigotes in combination with 0.5 pairs SGH.
At 5 weeks post immunization, and 1 and 9 months post challenge, the parasite load was measured in the ear, lymph node (from challenged ear), spleen and liver by the limiting dilution assay as previously described [33].
Whole blood from immunized hamsters was collected at the indicated time points before sacrifice in 1.1mL microcentrifuge tubes with serum gel-clotting activator (Sartesdt, GE). The serum was separated by centrifugation and used in IgG assays. Total IgG, IgG1 and IgG2 responses to L. donovani soluble antigens were measured by ELISA as described [34]. The following clones were used in the study. IgG cocktail (Catalog# 554010); IgG1 (clone-G94-56); IgG2 (clone-G192-3) (BD Biosciences).The cut-off value of reactivity for SLA antigens was calculated as the mean plus 2 SD of the OD values observed in naive controls. Sera from 9 naïve hamsters were used for determining cut-off values.
Splenocytes were collected from hamsters, macerated, lysed in Trizol for RNA extraction. Total RNA was extracted from the ear, lymph node (superficial parotid lymph node), spleen and liver of infected hamsters using TRIzol reagent (Invitrogen). First-strand cDNA synthesis was performed with ≈1–2μg of RNA using a Transcriptor High Fidelity cDNA Synthesis Kit (Roche). Amplification conditions consisted of an initial pre-incubation at 95°C for 10 min, followed by amplification of the target DNA for 40 cycles of 95°C for 15 s and 60°C for 1 min with the LightCycler 480 (Bio-Rad). The efficiency of each reaction was determined. The expression levels of genes of interest were normalized to β-Actin levels. The results are expressed in fold change of 2-ΔCt over control.
Oligonucleotide primers used for real time PCR were: β Actin, reverse, ACA GAG AGA AGA TGA CGC AGA TAA TG, forward, GCC TGA ATG GCC ACG TAC A; IFN-γ, reverse, TGT TGC TCT GCC TCA CTC AGG, forward, AAG ACG AGG TCC CCT CCA TTC; TNF-α, reverse, TGA GCC ATC GTG CCA ATG, forward AGC CCG TCT GCT GGT ATC AC; IL-10, reverse, GGT TGC CAA ACC TTA TCA GAA ATG, forward, TTC ACC TGT TCC ACA GCC TTG; IL-4, reverse, ACA GAA AAA GGG ACA CCA TGC A, forward, GAA GCC CTG CAG ATG AGG TCT; IL-12/IL-23p40, reverse, AAT GCG AGG CAG CAA ATT ACT C forward, CTG CTC TTG ACG TTG AAC TTC AAG, iNOS, reverse, ACC ACA CAG CCT CCG AGT CC, forward, CTG CCA GAT GTG GGT CTT CC. The primers and probes were synthesized at the Center for Biologics and Evaluation and Research, FDA Core facility.
Statistical analysis was performed using GraphPad Prism 5.0 software (GraphPad Software Inc., USA). Non-parametric Kruskal-Wallis test followed by Dunns test were used to compare data from four groups (G1, G2, G3 and G4). Differences were considered significant when a p value ≤ 0.05 was obtained.
Since the efficacy of genetically modified Leishmania parasites in general, and LdCen-/- parasites in particular have not been tested through an intradermal route, we sought to determine the persistence of LdCen-/- parasites at the site of injection and its dissemination to other organs where relevant immune reactions might occur. Thus, we measured the parasite load in the ear, lymph node, spleen and liver 5 weeks post immunization (5wpi) in hamsters primed with LJM19 and boosted with LJM19 plus LdCen-/- (G1). We observed an average of 300 and 30 viable LdCen-/- parasites in the ear and draining lymph node, respectively, measured by limiting dilution (Fig 1A and 1B). However, we could not recover any LdCen-/- parasite from the spleen and liver from either G1 or from animals immunized with LdCen-/- intracardially.
The analysis of sera from hamsters primed with LJM19 and boosted with LdCen-/- plus LJM19 indicated the occurrence of strong Leishmania-specific antibody responses. The SLA-specific IgGTotal and IgG1 production was significantly higher for immunized animals in G1 when compared to animals immunized IC with LdCen-/- (IC, G2), at 5wpi (p<0.05) (Fig 2A and 2B). However, increased levels of IgG2 were detected in G1 and G2 groups compared to G3 and control group G4 (p<0.01 and p<0.001, respectively; Fig 2C) that show base line reactivity. Importantly, the IgG2/IgG1 ratio was significantly higher in G1 and G2 groups compared to G4, the control group (p<0.05) and to G3 (p<0.05) that received LJM19 alone (Fig 2D). Of note, the IgG2/IgG1 ratio was higher in G2 compared to G1 (p<0.05) (Fig 2D).
In order to evaluate the protective immunity induced after ID immunization, we analyzed the mRNA expression of both Th1 and Th2 cytokines (IFN-γ, iNOS, IL-12/IL-23p40, IL-4 and IL-10) in the ear, the site of injection at 5wpi. The IFN-γ, iNOS and IL-12 expression levels were significantly higher in immunized hamsters in G1 compared with G2, G3 and G4 groups (p<0.05; Fig 3A,3B and 3C, respectively). Expression of IL-4 and IL-10 cytokines was also higher in G1 animals after immunization, when compared to animals in groups G2, G3 and G4 (p<0.05; Fig 3D and 3E, respectively). However, the level of IFN-γ was higher than either IL-10 or IL-4 in G1 animals. Additionally, G1 animals showed up to a ~16-fold, ~12-fold and ~13-fold increase in IFN-γ, iNOS and IL-12 respectively, compared to BSA-immunized control group G4 (Fig 3F). Similarly, mRNA levels of IL-4 and IL-10 were up-regulated ~9- and 8-folds, respectively, in G1 compared to G4 animals.
The number of viable L. donovani parasites was determined by the limiting dilution assay in the draining lymph node and spleen of immunized hamsters a month post challenge (mpc) (Fig 4). The number of live parasites in the lymph node (Fig 4A) and in the spleen (Fig 4B) were significantly lower (p<0.05) in the G1 and G2 groups either immunized with LJM19 then boosted with LdCen-/- plus LJM19 or immunized with LdCen-/- alone respectively, when compared with the groups of hamsters that received LJM19 alone or BSA alone.
We measured antibody levels in the sera of immunized animals one mpc. No difference was observed in the levels of IgGTotal and IgG1 between the groups (Fig 5A and 5B). The level of IgG2 was elevated in G1 immunized animals when compared to G3 and G4 hamsters (p<0.05 and p<0.01, respectively; Fig 5C). In addition, G1 group presented a significantly higher IgG2/IgG1 ratio in comparison to G3 and G4 control groups (p<0.05 and p<0.01, respectively) (Fig 5D), indicative of a Th1-type immune response post challenge.
The mRNA expression level of Th1 and Th2 cytokines was estimated by qRT-PCR one mpc. In the lymph node (LN), G1 and G2 groups presented a high expression of IFN-γ, when compared to G3 and G4 groups (p<0.01) (Fig 6A). A moderate increase in the expression levels of iNOS mRNA transcripts in LN was observed only in G1, and G3 immunized hamsters compared to G4 group (p<0.05) (Fig 6B). Interestingly, cytokine IL-12 mRNA levels in LN were significantly higher in G1, G2 and G3 groups compared to the G4 control group (p<0.001) (Fig 6C). Concomitantly, the mRNA levels of the Th2 cytokines IL-4 and IL-10, primarily regulatory cytokine, were higher in LN of G4 control animals in comparison to G1, G2 and G3 immunized groups (p<0.05) (Fig 6D and 6E). Additionally, IFN-γ, iNOS and IL-12 from G1 were significantly up-regulated by ~8 folds, ~14 folds and ~5 folds, respectively, compared to G4 group (Fig 6F).
In the spleen, IFN-γ was up-regulated one mpc in groups G1 (p<0.001), G2 (p<0.01) and G3 (p<0.01), compared to control group (G4) (Fig 6G) In addition, G1 presented a higher expression of iNOS when compared to G2, G3 and G4 after challenge (p<0.001, p<0.005 and p<0.01, respectively) (Fig 6H). Interestingly, IL-12 mRNA levels were not significantly different in spleens from animals of the 4 groups though G1 animals exhibited a trend for increased IL-12 expression (Fig 6I). The IL-4 expression was decreased in G1 and G2 groups, but not significantly (Fig 6J) compared to G3 and G4. However, there was a significant reduction in IL-10 expression (p<0.05) in G1 when compared to G2, G3 and G4 groups (Fig 6K). The IFN-γ, iNOS and IL-12 mRNAs from G1 were significantly up-regulated by ~19 folds, ~32 folds and ~10 folds, respectively, compared to G4 group (Fig 6L).
In the liver, IFN-γ expression was upregulated one mpc in G1 and G2 groups in comparison with G3 and G4 animals (p<0.01) (Fig 6M). However, iNOS expression was significantly higher in G2 compared to G1, G3 and G4 (Fig 6N). Liver cells from group G1 immunized hamsters induced a significantly high expression of IL-12 (Fig 6O). IL-4 expression was significantly higher in G4 animals compared to G1, G2 and G3 hamsters (p<0.01), and lower in G2 group compared to G1 group, but was not significantly different compared to G3 group (Fig 6P). On the other hand, hamsters from control group G4 showed a significant up-regulation (p<0.001) in IL-10 expression compared to G1, G2 and G3 (Fig 6Q). The IFN-γ, iNOS and IL-12 mRNAs from G1 were moderately up-regulated by ~1.3 folds, ~0.39 folds and ~1.2 folds, respectively, compared to control group G4 (Fig 6R).
The animals in G1 and G2 groups showed robust protection 9 mpc as evident by a significant decrease in the lymph node parasite load (Fig 7A), spleen (Fig 7B) and liver (Fig 7C) in comparison to G4 animals (p<0.0001). Immunization with salivary gland protein LJM19 alone (G3) also provided significant protection (p<0.01), albeit weaker than that observed in animals in groups G1 and G2, compared to G4 BSA immunized animals.
Further, we wanted to test whether LdCen-/- immunization causes re-establishment of homeostatic conditions by comparing the spleen sizes. Both G1 and G2 presented a non-inflamed spleen (median 3.4 and 3.5cm, respectively), as compared to a highly inflamed spleen in the control group G4 (median 7.9cm) (Fig 7D). G3 animals showed an intermediate spleen size (median 5.2 cm).
Previous work from our laboratories has shown that the LdCen-/- live attenuated vaccine is immunogenic in mice, hamsters and dogs [12,13,18]. Similarly, LJM19 protein from saliva of the vector Lu. longipalpis protected hamsters against challenge with L. infantum and L. braziliensis [26,28]. In the present work, we examined the value of combining the two immunization strategies for their potential to elicit protective immune responses in a hamster model of Leishmania donovani infection. ID needle inoculation of the ear has been extensively employed as the route of infection that most closely replicates the physiological ID and intra-epidermal deposition of parasites by the bite of an infected sand fly [33,35–37]. Additionally, the ID route presents the most practical route for vaccine delivery [26,28,29,31,32]. We hypothesized that a prime/boost strategy with LJM19 followed by LdCen-/- parasites plus LJM19, all delivered intradermally, would induce long-lasting protective immunity against L. donovani particularly since LdCen-/- parasites can undergo limited replication in the immunized host and provide an array of antigens very similar to those produced by a virulent parasite. As such, in this prime/boost protocol, priming with LJM19 would generate a specific adaptive immune response to the sand fly salivary gland protein as was observed in previous studies [26,28] that could result in a potent supplement to the specific adaptive immune response to antigens of the LdCen-/- parasites.
We had previously observed LdCen-/- parasites in the spleen up to 5 wpi after intracardial (IC) injection [12]. In the current study, we observed parasites in the immunized ear and the draining lymph node but not in the spleen, at 5wpi after ID injection, suggesting that either the parasites take a longer time to disseminate to the viscera and reach the spleen or alternatively they do not visceralize. Of interest, recent studies with dermotropic parasite strains (L. donovani isolated from a cutaneous lesion and L. major) that fail to persistently visceralize nevertheless produced protective immunity in a low-dose infection followed by challenge with L. donovani and L. infantum in mouse models [38,39]. This suggests that visceralization may not be a necessary pre-condition for protective immunity against VL to develop in the immunized mice. Consistent with this hypothesis, our results indicate that since the draining lymph nodes represent the immunological niche where relevant reactions between APCs that acquired the antigens and naïve T cells could occur, recovery of attenuated parasites from the lymph nodes 5 weeks post immunization suggested that parasite persistence, i.e. antigen availability, was adequate for protective immunity to be established.
Immunization of hamsters with attenuated parasites associated with LJM19 protein elicited a biased Th1-type immune response at 5wpi at the site of injection. As expected, immunization by LJM19 alone provided protection against L. donovani parasites, however, it was considerably weaker compared to the one observed following a prime/boost ID immunization with LJM19 followed by LJM19 and LdCen-/- or an IC immunization with LdCen-/-. Of note, boosting with LdCen-/- along with LJM19 protein through the ID route resulted in a higher pro-inflammatory response compared to immunization with LdCen-/-alone through the IC route. Indeed, immunization with LdCen-/- through the intravenous (IV) route in mice and IC route in hamsters [12] and subcutaneous route in dogs [13] has been shown to promote a pro-inflammatory response, with the presence of IL-12p40, IFN-γ, iNOS and TNF-α. Additionally, our finding that increased levels of IL-4, IFN-α, iNOS and IL-12/IL-23p40 in immunized hamsters would suggest a mixed immune response (Th1 biased) triggered by LdCen-/- vaccination, as we observed in dogs immunized previously [18,38,40]. Our results suggest that the ID mode of immunization, at least in combination with LJM19, is equally efficacious compared to the IC mode. Additionally, the high ratio of IgG2/IgG1 observed in groups G1 and G2 is considered an additional immune biomarker of protection [5,41–43].
During natural transmission, an infected sand fly deposits saliva and parasites into the skin of the host while feeding. To mimic the natural mode of infection with Leishmania, we injected L. donovani wild type parasites into the ear of hamsters along with sand fly salivary gland extract. After one month of infection, hamsters immunized with LdCen-/- either alone (IC) or with LJM19 protein (ID) demonstrated a reduced parasite burden in the lymph node and spleen. In our study, similar to the response observed post-immunization, challenged hamsters presented a significant increase of IgG2 production, and a high ratio of IgG2/IgG1, as well as an enhanced production of IFN-γ and iNOS. Higher levels of IgG2 might also contribute to pathogen clearance in vaccinated animals [44,45].
Previously, it was reported that iNOS and concomitant high levels of NO were produced by macrophages in protected mice vaccinated with attenuated parasites after challenge with L. donovani [12,14]. In the present study hamsters immunized with either LdCen-/- alone or in association with recombinant LJM19 displayed increased IFN-γ and iNOS expression in lymph nodes and spleen one month after challenge with wild type parasites. Of importance, five weeks post-immunization we observed a higher production of IL-12 only in animals immunized with LdCen-/- in association with recombinant LJM19. It can be speculated that LJM19 might be pre-conditioning the innate immune arm and thus allowing the antigen presenting cells such as DCs to produce IL-12 that is necessary for initiating a strong adaptive Th1 cell immunity. Certainly, the ability of LJM19 to produce a Th1 response in hamsters has been previously demonstrated [26]. As such, it may be argued that LJM19 might be enhancing the immunogenicity of LdCen-/- as a vaccine. An increased IFN-γ/IL-10 ratio has been observed when DNA vectors expressing KMP11 along with LJM19 were used as immunogens compared to either KMP11 or LJM19 alone at 5 months post challenge [46]. This increased IFN-γ/IL-10 ratio did not result in reduced splenic parasite burden between KMP11+LJM19 and either antigen or LJM19 alone groups at 5 months post challenge. Further in their study the authors also observed increased IFN-γ/IL-10 ratio after 5 months of infection in non-immunized animals which does not explain the role of increased IFN-γ levels in protection. The observed differences between da Silva [46] and our study in splenic parasite burden could be due to the ability of the live attenuated parasites to induce sustained immunological reactions (Fig 6G and 6K) because of a longer availability of a multitude of antigens compared to recombinant antigens that tend to have limited availability and diversity that is reflected in parasite control up to 9 months post challenge. Further, in the LdCen-/- immunized animals (G1 and G2) down-regulation of IL-10 and a concomitant increase in IL-12 in lymph node, spleen and liver may explain the greater parasite killing observed at the challenge site. In murine and human VL, production of Th1 cytokines is desirable for resolution of infection [47–50]. In addition, IL-12 results in the generation of Th1 cells that produce both IFN-γ and IL-12, thus favoring the development of a protective cellular immune response against Leishmania [51–53]. An important consideration is that the sustained Th1-type immune response and long-term protection generated here against L. donovani after ID injection of LdCen-/- parasites in combination with LJM19 is comparable to that observed following IC or intravenous immunization with LdCen-/- parasites alone [12,14]. A similar induction of Th1 immunity was also observed in dogs immunized subcutaneously with LdCen-/- parasites and challenged with L. infantum [18]. At 9 months post challenge, the parasite load in the lymph nodes and in the spleen was significantly reduced in all the immunized groups compared to the control group. The control of parasitemia in the spleen translated into lack of splenomegaly in immunized and challenged animals compared to control challenged animals. Importantly, the significant reduction of parasitemia after immunization with LJM19 alone in our current study argues that LJM19 contributes to the observed protection in G1 hamsters. This is corroborated by Gomes at al. [26] who observed a decrease in parasite load in the spleen and liver in hamsters immunized with LJM19 after 2 and 5 months post I.D. inoculation of L. infantum chagasi with sand fly salivary gland homogenate. Additionally, in an independent study, Tavares et.al [28] showed that hamsters immunized with LJM19 induced protection against infection with L. braziliensis. Taken together our data indicate the induction of a long-lasting protective immune response in the spleen, liver and lymph nodes in hamsters immunized intradermally with LJM19 and LdCen-/- after challenge with virulent parasites and reveal that a stronger immune response is elicited when Leishmania donovani live attenuated parasites are combined with a salivary gland protein. In summary, we have demonstrated the capability of a combined vaccine composed of live attenuated LdCen-/- parasite and a defined salivary gland protein from Lu. Longipalpis (LJM19) delivered intradermally to confer strong long-lasting protection against L. donovani infection in a hamster model.
|
10.1371/journal.pbio.2005359 | Comparative genomics and the nature of placozoan species | Placozoans are a phylum of nonbilaterian marine animals currently represented by a single described species, Trichoplax adhaerens, Schulze 1883. Placozoans arguably show the simplest animal morphology, which is identical among isolates collected worldwide, despite an apparently sizeable genetic diversity within the phylum. Here, we use a comparative genomics approach for a deeper appreciation of the structure and causes of the deeply diverging lineages in the Placozoa. We generated a high-quality draft genome of the genetic lineage H13 isolated from Hong Kong and compared it to the distantly related T. adhaerens. We uncovered substantial structural differences between the two genomes that point to a deep genomic separation and provide support that adaptation by gene duplication is likely a crucial mechanism in placozoan speciation. We further provide genetic evidence for reproductively isolated species and suggest a genus-level difference of H13 to T. adhaerens, justifying the designation of H13 as a new species, Hoilungia hongkongensis nov. gen., nov. spec., now the second described placozoan species and the first in a new genus. Our multilevel comparative genomics approach is, therefore, likely to prove valuable for species distinctions in other cryptic microscopic animal groups that lack diagnostic morphological characters, such as some nematodes, copepods, rotifers, or mites.
| Placozoans are a phylum of tiny (approximately 1 mm) marine animals that are found worldwide in temperate and tropical waters. They are characterized by morphological simplicity, with only a handful of cell types, no neurons, no tissue organization, and even no axial polarity. Since the original description of Trichoplax adhaerens 135 years ago, no additional accepted species has been established, leaving the Placozoa as the only animal phylum with only a single formally described species. While classical morphological species identification has failed to reveal further species, single-gene DNA sequence analyses have identified a broad and deep genetic diversity within the Placozoa. To address the significance of this deep genetic diversity in this morphologically uniform phylum, and to better understand its consequences for speciation processes, general biology, and species delimitation in the Placozoa, we sequenced the genome of the placozoan isolate “H13,” a lineage distantly genetically related to T. adhaerens. Our multilevel genomic comparisons with the T. adhaerens genome show considerable differences in the general structure of the genome and the makeup and history of various gene families of biological relevance to habitat adaptation. Based on comparative genomics, we here describe the second placozoan species and show that it belongs to a new genus.
| Placozoans Grell, 1971, are small, benthic marine animals found worldwide in various habitats [1–6]. To date, only a single species has been described, Trichoplax adhaerens Schulze 1883. Animals are flat and have a typically disc-like morphology but have the capacity to change shape [7–9]. The lack of symmetry axes, neurons, and defined muscle cells, and the presence of only six morphologically distinguishable somatic cell types ([9,10]; Fig 1, S1 Fig), makes the Placozoa morphologically the most simply organized animals. The prominent placozoan modes of reproduction are asexual, i.e., binary fission and budding [8,9,11–13] that produce genetically identical clones. Sexual reproduction has rarely been observed under laboratory condition [14–19], but both oocytes and sperm cells have been reported [14,17,19], and fertilization, likely coupled with genetic exchange, was confirmed based on structural similarities of the placozoan eggshell with the fertilization membrane of other animal groups [16].
No sexually reproducing individual has ever been reported from the wild. However, the occurrence and success of sexual reproduction in the field have been demonstrated by DNA sequence analyses, since nuclear-encoded marker genes have revealed the occurrence of allele sharing and mixing of heterozygous alleles in a natural placozoan population isolated from a Caribbean habitat [20]. These molecular signatures for genetic exchange prove that sexual reproduction does occur and that the life cycle is completed in the natural environment. However, all efforts to follow the placozoan embryonic development in the laboratory have failed to date. All embryos died at an early stage during development, never reaching beyond the 128-cell stage [19]. The fragmentation of the nucleus in the zygote [21] was previously suggested as the reason for the termination of development, although this has been questioned [19]. This ambiguity and scarcity of information has, therefore, left us with a large knowledge gap regarding the life history of the Placozoa and has resulted in speculations of the existence of a missing life stage (compare [22]).
The genome of the diploid T. adhaerens was sequenced previously [22], revealing that this morphologically very simple animal harbors a rich repertoire of gene families [22]. These families are known from bilaterian animals and are typically associated with a considerable cell type diversity, a complex body plan, developmental processes, and behavioral responses to external stimuli [10,23–31]. Additionally, single-gene molecular phylogenetics have identified a sizeable cryptic diversity within placozoans collected worldwide; but while their gross morphology is highly plastic, morphologically, all isolates fit the description of T. adhaerens [32] (Fig 1). The high intraspecific shape variability, coupled with an ultraconserved internal structure (Fig 1, S1 Fig), does not allow the establishment of reliable diagnostic morphological characters in the Placozoa, hindering attempts to characterize their diversity.
While these single-marker studies provided clear indications that additional species may be uncovered in the Placozoa, two fundamental questions remain: how different are placozoans at the nuclear genome level, and what can we learn from comparative genomics about the evolution and diversity of placozoans? To address these questions, we generated a high-quality draft genome of a placozoan lineage that is genetically distantly related to T. adhaerens [3,5] and performed a multilevel comparison, including genome synteny, gene clustering, gene ontology (GO) term enrichment, allele sharing, and cross-phylum comparative distance analyses. This approach, together with the morphological characterization of the lineage H13, allowed us to assign a taxonomic status to morphologically cryptic taxa and led to the establishment of the second placozoan species in a new placozoan genus.
Based on mitochondrial 16S ribosomal DNA (rDNA) analyses, the genetic lineage H13 is among the most distantly related haplotype to T. adhaerens (lineage H1) [5], whose nuclear genome has been sequenced previously [22]. We hypothesized that the substantial 16S rDNA divergence might also be reflected on the whole-genome scale and, therefore, targeted H13 for nuclear genome sequencing. To assemble the genome of H13—a new species described here, called H. hongkongensis nov. gen., nov. spec. (Fig 1, S1 Fig; see species description in Material and methods; Tables 1 and 2)—we generated 24 Gb of paired-end reads and 320 Mb of Moleculo (Illumina Artificial Long Synthetic) reads. Our final, highly complete 87-megabase assembly contained 669 high-quality and contamination-filtered contigs with an N50 of 407 kb (S1 Table; S2–S4 Figs), 7 megabases smaller than the T. adhaerens contig assembly. The overall calculated genome heterozygosity (based on single-nucleotide polymorphism [SNP] counts, see S2 Table) was 1.6%, which is moderate for a marine animal but about average when compared to arthropods and high in comparison to terrestrial chordates [33]. This value cannot be compared to T. adhaerens because of the low genome coverage of the latter, which does not allow haplotype phasing.
We annotated the genome with a combination of 15.3 Gb of RNA-Seq and ab initio methods to yield 12,010 genes (S1 Table, S1 & S2 Data). A high percentage of raw reads mapped back to the genome (S3 Table), and between 90.8%–95.3% of the 978 genes in the BUSCO v3 Metazoa dataset were identified in the transcriptome and the ab initio gene models, respectively (S4 Table). Together, this suggests an almost complete assembly and annotation, in which 96.5% of the genes in the H. hongkongensis genome were expressed in what are commonly considered adult animals. In our gene set, H. hongkongensis had 490 more genes than the 11,520 genes reported in the original T. adhaerens annotation from 2008 [22]. We reannotated T. adhaerens with AUGUSTUS and found an additional 1,001 proteins and also managed to complete formerly partial proteins (for T. adhaerens Blast2GO protein annotations see S3 Data). This approach added 4.4 Mb of exons to the T. adhaerens annotation, an increase of 28% of exonic base pairs to the original annotation. The new T. adhaerens annotation now has 511 more genes than H. hongkongensis, which accounts for some portion of the size difference between the two genomes.
Moleculo reads also enabled us to assemble very large reference contigs, the longest being over 2 Mb. We compared the organization of genes in H. hongkongensis to the 10 longest scaffolds in the T. adhaerens genome (size range 2.4–13.2 Mb; accounting for 66% of the T. adhaerens assembly). We found 144 contigs >100 kb from H. hongkongensis that aligned to these 10 scaffolds, accounting for 69% of the H. hongkongensis assembly (Fig 2A). Mean gene collinearity (i.e., the same genes in the same direction) in this reduced genome representation was in the range of 69.5% to 78.8% (mean 73.6% ± 5.5%; see S5 Table). The mean number of genes per syntenic block was 33.8 (±25.2) in the reduced set and 33.9 (±24.7) when comparing full genomes (S5 Fig), which indicates that the reduced set is representative for both complete genomes.
Although much of the gene order is conserved between the two species, we counted 2,101 genes (out of the 8,260 genes in the 10 scaffolds) that were inverted or translocated within the same scaffold relative to the order in the T. adhaerens scaffolds. These numbers seem low when compared to the fast-evolving bilaterian genus Drosophila [34,35] or Caenorhabditis [36], but they are in the range of rearrangements found between mouse and human [37]. Comparison to Bilateria, however, might be misleading (see also results on genetic distances below), and genome rearrangement events might be more favored in some bilaterian taxa because of inherent genomic traits such as transposon-induced rearrangement hotspots [38]. Nonetheless, the high percentage of rearrangements between T. adhaerens and H. hongkongensis is clear evidence for a deep genetic separation of both lineages.
To estimate how divergent the two placozoan genomes are at the sequence level, we calculated genetic distances for 6,554 one-to-one orthologs. Between H. hongkongensis and T. adhaerens, genetic distances ranged from 0.9% to 80.1% (mean 28.3% ± 12.9%) for proteins and 7.4% to 80.7% (mean 28.5% ± 9.9%) for coding sequences (CDSs), respectively (Fig 2B). To assess if certain genes are under positive (diversifying) selection, indicative of functional evolution, we calculated the ratio of nonsynonymous to synonymous nucleotide substitutions (dN/dS ratio [39]) for each H. hongkongensis and T. adhaerens one-to-one ortholog pair. Results show that most orthologs (97%) are under strong purifying selection (dN/dS < 0.5). One might hypothesize that strong purifying selection pressure is the reason for the phenotypic stasis we see in modern placozoans. However, more placozoan genomes in the phylum are clearly needed to test this hypothesis. Despite this strong tendency toward purifying selection, a high proportion of orthologs (46%) showed larger protein distance than CDS distance and, therefore, an accumulation of double or triple mutations in already mutated codons, which led to amino acid substitutions (S6 Fig).
Only 3 of the 6,554 one-to-one orthologs had dN/dS ratios slightly >1, indicating positive selection (S7 Data; see S6 Fig for an estimate of mutation saturation in codons). One of these seems placozoan specific, since it could not be annotated because of missing UniProt BLAST hits and InterPro domains, respectively. For the second, GO annotation and InterPro IDs indicate a role in telomere maintenance. The third positively selected gene (CYP11A1) is putatively a cholesterol side-chain cleavage enzyme acting in the mitochondrion.
The roughly 4x coverage of the genome with long Moleculo reads (N50 of 5.4 kb) allowed the assembly of large haplocontigs (i.e., contigs representing both haplotypes of the genome). This phasing information for large parts of the genome facilitated the isolation of 2,870 one-to-one orthologs with both full-length alleles after a highly stringent filtering procedure. Only by using the phasing information we were able to show that many orthologs with high allelic variation in H. hongkongensis were also profoundly different between the species (S7 Fig). This indicates that genetic sequence adaptation already takes place at the population level and is further magnified between species in the same genes.
The Markov cluster (MCL) analysis identified 6,644 true one-to-one orthologs (for an overview of ortholog categories, see Material and methods and [40]) for both placozoan species (55% of all proteins in H. hongkongensis and 53% in T. adhaerens, respectively) (S8 Fig). A fraction of 465 (3.8%) H. hongkongensis and 1,036 (8.3%) T. adhaerens proteins, respectively, did not have reciprocal BLAST hits. The difference in the non-BLAST hits almost perfectly matches the differences in total gene numbers, which is probably an indication that genes without a homolog in H. hongkongensis account at least partially for the slightly higher gene number in T. adhaerens. A high proportion of proteins had BLAST hits to the UniProt database, and only 15.4% (1,859) and 19.0% (2,384) of H. hongkongensis and T. adhaerens proteins, respectively, did not have BLAST hits to metazoans included in UniProt.
Placozoan-specific duplications constitute a significant proportion of both proteomes, with 3,943 (32.8%) co-orthologs in H. hongkongensis and 3,484 (27.8%) in T. adhaerens. The enrichment analyses for the proteins in each non-BLAST-hit bin identified unique GO terms in all three GO categories among the first five most significantly enriched GO terms (S4 & S5 Data). The same applies to one-to-many and many-to-one co-orthologs in both species.
The enrichment analyses further indicate that both placozoan species have multiple co-orthologs associated with G-protein-coupled receptor (GPCR) signaling. A rich repertoire of GPCRs has been identified in T. adhaerens [22], but here, we were able to identify independent GPCR duplications in H. hongkongensis and T. adhaerens, respectively (S6 Data). Furthermore, we identified multiple enriched GO terms related to synaptic activity in all co-ortholog categories (S5 Data) and both placozoan species. This points to a plethora of independent duplication events in gene families related to sensory capacities. Despite lacking neurons (based on traditional morphological classifications), T. adhaerens has previously been shown to stain positive for FMRFamide [10,41] and recently even to change behavior when exposed to physiologically relevant levels of neuropeptides [31].
Based on the identification of vast and independent gene family expansions in both placozoans, we propose that adaptation in the Placozoa, ultimately leading to speciation, is coupled with independent gene duplications as suggested, for example, for bacteria, yeast, plants, and other animals (compare [42–45]). H. hongkongensis was isolated from a stream running through a mangrove with rapid drops in salinity and temperature, especially during heavy rainfall in the summer. We hypothesize that the presence of multiple divergent copies of genes involved in various processes, such as behavior and metabolism (compare [42,43]), in addition to a situation-dependent expressional fine-tuning of these copies was necessary for adaptation to this habitat and would facilitate speciation. We furthermore propose that the presence of multiple copies of genes and their expression does not affect the phenotype but instead provides a genetic toolkit for gradual physiological responses to (changes in) the environment.
All internal Linnaean ranks within the Placozoa are, as yet, undefined [5]. Despite efforts to identify them, reliable diagnostic morphological characters, commonly used for defining animal species, are lacking in the Placozoa [46]. Thus, all present taxonomic definitions in the phylum must solely rely on diagnostic molecular characters. In other taxonomic groups (e.g., bacteria and archaea [47], protists [48,49], and fungi [50]), purely sequence-based approaches and working models for the distinction of taxa have been proposed and are generally well established and widely accepted [51]. In animals, such methods (which may be based on distances, on trees, or on allele sharing; [52]) are currently under development and have been used in rare cases to identify and describe cryptic species [53].
In a first step to converting the identified genomic differences into a taxonomically meaningful system, we studied reproductive isolation by addressing allele sharing within placozoan isolates from different localities. To identify reproductive isolation, a conspecificity matrix (CM) was generated [54]. The CM was based on three nuclear genes encoding ribosomal proteins and clearly identified reproductive isolation between placozoan clades (Fig 3). This approach extends a previous study that has uncovered sexual reproduction only within one placozoan haplotype (H8) [20] and provides clear evidence that the previously established placozoan clades (based on 16S genotyping) are reproductively isolated biological species.
We have shown that biological species exist in the Placozoa. Previous studies have furthermore provided first indications for the existence of deeper differences between placozoan lineages [1,3], with as-yet-unknown correspondence to, for example, the Linnaean ranks of genus, family, order, and class. However, these observed deeper divergences were based on single marker genes only, and no diagnostic morphological traits could be identified to establish a firm, higher-level, systematic framework in the Placozoa. To further estimate the level of taxonomic relatedness between T. adhaerens and the new placozoan species H. hongkongensis (strain H13), and in an attempt to initiate a higher-level taxonomic system for the Placozoa, we performed cross-phylum multimarker sequence divergence analyses. To do so, we compared the variation between the two placozoans to variation within the other three nonbilaterian phyla, Cnidaria, Ctenophora, and Porifera (compare [1]), as well as the bilaterian phylum Chordata. Marker sets included a nuclear protein set of 212 concatenated proteins (dataset 1, a taxon-extended matrix from [56]; S7–S9 Tables; see Fig 4) as well as 5 selected genes with different substitution rates (S9–S14 Figs), all commonly used for DNA barcoding and molecular systematics.
Across individual markers, it appears that the phylogenetic ranks are most robust in the Cnidaria, in which the partitioning of molecular variation matches the established taxonomy, in that Linnaean ranks consistently correspond to the greater distance between groups (Fig 4; S9–S14 Figs). The same is true for the Chordata, which was included in our distance calculations for the 212 nuclear protein set as an example of a bilaterian phylum with a high taxonomic coverage (many genomes are available for this group). However, distances in chordates are, in general, much lower when compared to the overall more similar nonbilaterian phyla. This indicates that (i) genetic distances and corresponding Linnaean rank assignments in Chordata cannot be compared to nonbilaterian lineages and (ii) that comparisons among nonbilaterians are better suited to guide taxonomic ranking of the two placozoan species. We consequently used genetic distances in the Cnidaria as an approximation and comparative guideline for the higher systematic categorization of the new placozoan species.
Genetic distances between H. hongkongensis and T. adhaerens were higher than those for the Cnidaria in five of the six marker sets at the generic level but lower at the family level for all markers (S14 Fig, S10 Table), which, cautiously interpreted, supports genus-level genetic differences between the two placozoans.
A clear split of the Placozoa in the molecular groups “A” and “B” was previously shown by the rearrangement pattern of mitochondrial genomes [61] and compensatory base changes in the internal transcribed spacer 2 (ITS2) [55]. The conspecificity analysis, the high amount of genomic rearrangement, and the large-scale independent gene duplication history, as well as the genetic distances in six independent datasets, strongly support this split (Fig 3). Since clades were identified as the primary taxonomic units—i.e., biological species—these two previously identified higher-level placozoan “groups” consequently represent at least the genus level in the Linnaean hierarchical system. We therefore establish the new genus Hoilungia for the former group “A” (clades III–VII), which is, so far, the single sister genus to Trichoplax (former group “B”; clades I and II).
Future research efforts focusing on genome sequencing of additional placozoan clades/species will likely help to establish a broader and more detailed systematic framework for the Placozoa and provide further insights into the mechanisms and driving forces of speciation in this enigmatic marine phylum.
Recent discussions about the phylogenetic position of placozoans have largely been based on the T. adhaerens genome. A better sampling of placozoan genomic diversity is, however, needed [62] to address their placement in the metazoan tree of life. In this context, it is important to first assess if adding another placozoan genus would break up the long placozoan branch. The inclusion of a single representative of a clade with a very long terminal branch, or fast-evolving taxa that can have random amino acid sequence similarities, may result in erroneous groupings in a phylogeny (so-called “long-branch attraction artefacts”) [63,64]. To address these questions, we generated a highly (taxa) condensed version of the full protein matrix from Cannon and colleagues [56] (termed dataset 2; with less than 11% missing characters and 194 genes). We additionally created a Dayhoff 6-state recoded matrix [65,66] of this second set to reduce amino acid compositional heterogeneity, which is also known to be a source of phylogenetic error [67,68]. Phylogenetic analyses were performed on these two matrices (protein and Dayhoff-6 recoded), using the site-heterogeneous CAT-GTR model in PhyloBayes-MPI [69] and using the site-homogenous GTR model both in Phylobayes-MPI and RAxML (RAxML, protein only) [70], as well as the LG model in RAxML (protein only). The resulting trees (S15–S20 Figs) of the highly dense gene matrix (S21 Fig) suggest a sister group relationship of the Placozoa to a Cnidaria + Bilateria clade with both CAT-GTR (Protein, Dayhoff-6 recoded, S15–S17 Figs) and GTR models (Protein, S18 Fig) in PhyloBayes, or these relationships are unresolved (RAxML, protein, both GTR, S19 Fig, and LG, S20 Fig). This is in agreement with some previous findings [56,64,71–74] and with recent studies using a large gene set and intense quality controls [64] as well as improved modeling of compositional heterogeneity [68]. In addition, the sister group relationship of the Placozoa to the Cnidaria + Bilateria clade is corroborated by independent data—namely, the analysis of metazoan genome gene content [73,75,76].
Phylum: Placozoa, Grell 1971 [77]
Type Family: Trichoplacidae, Bütschli and Hatschek 1905 [78], synonymized with “Trichoplaciden” (in German original), Haeckel 1896 [79].
Diagnosis: We assign all currently known 19 placozoan genetic lineages (16S haplotypes H1-H19; [5]) to the Trichoplacidae. The description of T. adhaerens Schulze 1883 applies to all.
Type Genus: Hoilungia, nov. gen., Eitel, Schierwater, and Wörheide
Hoilungia is the second genus of the family Trichoplacidae.
Etymology: Hoilungia, pseudo-Latinized from “Hoi Lung,” Cantonese, meaning “sea dragon,” which is based on the shape-shifting dragon king in Chinese mythology.
Diagnosis: Gross and fine morphology appear similar among all placozoans studied to date. We therefore use molecular diagnostics to define Linnaean ranks. Among all tested markers, the mitochondrial large ribosomal subunit 16S rDNA appears to be the most variable among placozoans and other nonbilaterian phyla, and the mean pairwise distance is closest to that calculated for the nuclear dataset in most cases (S14 Fig). This marker also best mirrored classical taxonomy in the Porifera and Cnidaria (S11 Fig; in Ctenophora, 16S rDNA is highly derived and hard to identify [80]). According to these data, molecular diagnostics based on differences in the 16S rDNA appear to be suitable for current and future designation of species in the Placozoa, which is in agreement with previous results [3]. Diagnostics are here, therefore, defined by nucleotide substitutions in the 16S rDNA. Full-length 16S rDNA sequences of T. adhaerens and H. hongkongensis (clonal strain “M2RS3-2”), as well as for the undescribed Placozoa sp. H4 and sp. H8, were aligned with MAFFT v7.273 [81] using the GINSI option and otherwise default settings. Ambiguously aligned 5′ and 3′ sequence ends were removed. To this alignment, we added all currently available placozoan 16S haplotype sequences [5] using MAFFT [added option:—add]. The final alignment contained all 19 placozoan haplotypes and had a length of 2,551 nucleotides (including gaps). The region for identification of diagnostic nucleotides was restricted to a part of the 16S alignment that was previously shown to be suitable and sufficient for molecular haplotype discrimination [1,3,5]. We furthermore restricted the identification of diagnostics to stem regions of this rDNA to omit uncertainties in future taxonomic assignment due to ambiguously aligned loop regions. To identify molecular diagnostics for the genus Hoilungia, we screened for molecular synapomorphies (nucleotide exchanges) within the placozoan 16S group “A” (clades III–VII; [5,61]) versus group “B” (clades I and II).
Molecular diagnostics for Hoilungia and Trichoplax are summarized in Table 1.
Type species: H. hongkongensis, nov. spec., Eitel, Schierwater, and Wörheide.
Diagnosis: To identify molecular species diagnostics, we determined unique substitutions (based on the alignment used for genus diagnostics before) for H. hongkongensis (clade V) in comparison to the other Hoilungia clades (III, IV, VI, and VII).
Molecular diagnostics for H. hongkongensis are summarized in Table 2.
Type locality: A single specimen of H. hongkongensis (clonal strain “M2RS3-2”) was isolated in the Ho Chung River close to a small mangrove at Heung Chung village, Hong Kong (22.352728N 114.251733E), on June 6, 2012.
Type specimen: One specimen of H. hongkongensis (clonal strain “M2RS3-2”) has been mounted and deposited at the Bayerische Staatssammlung für Paläontologie und Geologie in München, Germany, under voucher number SNSB-BSPG.GW30216. Clonal individuals have been stored in ethanol as paratypes under voucher number SNSB-BSPG.GW30217 in addition to a DNA extraction under voucher number SNSB-BSPG.GW30218.
Etymology: hongkongensis, from “Hong Kong,” and “-ensis,” Latin, suffix referring to place of origin, as specimens are at present endemic to Hong Kong. The full name “Hoilungia hongkongensis” thus means “Hong Kong sea dragon.”
Two strains were used for this project: The “M2RS3-2” strain was used for the DNA sequencing (the “DNA strain”) and the “M153E-2” strain (the “RNA strain”) for the transcriptome. Both strains descend from a single placozoan individual each, which was isolated from mangroves/mangrove associates at two different sites in Hong Kong (SAR, China). The DNA strain was isolated from a dead mussel shell collected in the Ho Chung River close to a small mangrove at Heung Chung village (22.352728N 114.251733E) on June 6, 2012. The habitat undergoes daily changes in salinity, and on the day of collection, the salinity was 20 psu. The RNA strain was isolated from collection traps (for details on slide sampling, see [82]) connected to mangrove associates (Hibiscus sp.) and high shore mangrove (Excoecaria sp.) trees at Tai Tam Tuk (22.244708N 114.221978E) on March 30, 2012. Both clonal cultures were cultured in 14 cm glass Petri dishes as described [19], with a pure Pyrenomonas helgolandii algae culture (strain ID 28.87, Culture Collection of Algae, Georg-August-Universität Göttingen). The two different strains were used for DNA and RNA sequencing, respectively, to identify polymorphisms in these strains living in the same habitat but at two hydrogeographically distinct sampling sites (northeast versus southeast Hong Kong).
Animals were transferred in 20% BSA in artificial seawater, high-pressure frozen in a Wohlwend HPF Compact 02, and stored in liquid nitrogen. Samples were processed from −90 °C to room temperature for Epon embedding in a Leica AFS unit as follows: they were fixed and contrasted in 0.1% tannic acid in acetone for 24 h and washed 4 times for 15 min in acetone; samples were then incubated in 2% Osmium tetroxide in acetone while the temperature was increased stepwise to −40 °C within the next 23 h; samples were then washed and progressively infiltrated in Epon:acetone mixes (1:2, 2:1) and pure Epon while temperature was further raised from −40 °C to room temperature over 6 h. They were then polymerized in Epon. Seventy-nm ultrathin sections were cut on a Leica Ultracut and picked up on a copper slot grid 2 × 1 mm coated with a polystyrene film. Sections were poststained with uranyl acetate 2% in distilled water for 10 min, rinsed several times with distilled water followed by Reynolds lead citrate in distilled water for 10 min, and rinsed several times with distilled water. Micrographs were taken with a Transmission Electron Microscope Philips CM100 at an acceleration voltage of 80 kV with a TVIPS TemCam-F416 digital camera.
A "lavalamp" kmer/GC plot was generated (S2 Fig) to yield a high-resolution plot of read counts per %GC and 31 bp kmer coverage using the Jellyfish kmer counter and a set of custom Python scripts (kmersorter.py and fastqdumps2histo.py; for details on the procedure, see https://github.com/wrf/lavaLampPlot). In contrast to the conceptually similar approach Blobtools [108], we used raw reads instead of contigs to yield a high-resolution plot of read counts per %GC and 31 mer coverage. The plot identified two read clouds with high counts at a kmer coverage of 80–140x (heterozygous “read cloud”) and 160–260x (homozygous “read cloud”), respectively. Additional “read clouds” at 270–320x and 380–410x coverage mark repetitive sequence stretches. Another “read cloud” was found at a low coverage of 20–50x. Reads within this cloud and their pairs were extracted with kmersorter.py [added options: -s 0.16 -b 50 -w 0.40 -T -k 31] and fastqdumps2histo.py. Bowtie2 v2.2.5 [109] [added options: -q—no-sq] was used to map the 580,092 extracted reads to the 19 previously identified bacterial contigs (see section “Contamination screening”). More than 86% of these reads mapped to the bacterial contigs, confirming the bacterial origin of the reads within the low-coverage “read cloud.” Read counts identified a relatively high abundance of bacterial cells, and the GC content was similar to the host genome.
To estimate the per-base genome coverage, paired-end reads were mapped to the softmasked reference assembly with Bowtie2 v2.2.5 [added options: -q—no-unal—no-sq) and sorted with SAMtools v1.3.1 [110]. The bam file was used to create a bedgraph file in BEDtools v2.25.0 [111] by invoking the genomecov operation [added options: -ibam stdin -bga]. A custom Python script (bedgraph2histo.py) [added options: -m 2000] was used to create a coverage histogram table. Since 81.4% of the genome falls within the second peak (165–332x coverage with a maximum at 248x), most of the genome was merged in the reference assembly (S3 Fig).
To identify collinearity between the two placozoan species, all H. hongkongensis contigs >100 kb were aligned to the longest 10 T. adhaerens scaffolds (accounting for 70.3 Mb or 66.5% of the genome assembly; including 5.7-Mb gaps) with default settings. For generating the alignments, LASTZ v1.02.00 [115] (implemented as a plugin in Geneious) was used. Of the 222 H. hongkongensis contigs >100 kb, a total of 144 (accounting for 60.6 Mb or 69.4% of the genome assembly) aligned to the 10 longest T. adhaerens scaffolds. Aligned H. hongkongensis contigs were extracted from the assembly, sorted, and occasionally reverse complemented to be oriented according to the T. adhaerens scaffolds. Gene annotations (GFF) of contigs as well as protein sequences were extracted for the target scaffolds/contigs sets of both species. A MCScanX run [116] [added option: -a] was performed for each target set, using the extracted T. adhaerens and H. hongkongensis GFFs together with the reciprocal best 5 BLASTP hits [added options: -evalue 1e-10 -max_target_seqs 5 -outfmt 6] between and among proteins of both placozoans. Dual synteny line plots of the resulting collinearity files were visualized in VGSC v1.1 [117] [added options: -tp DualSynteny] and combined to Fig 2A. In addition, bar plots were generated for the 10 T. adhaerens scaffolds and the matching 144 H. hongkongensis contigs in VGSC [added option: -tp Bar]. Bar plots were mapped onto the DualSyntheny plots to show collinearity within each set and macrosynteny between both genomes. The percentage of collinearity between the T. adhaerens scaffolds and H. hongkongensis contigs was calculated in MCScanX, and results for the 10 scaffolds are given in S5 Table. The mean collinearity was calculated as the sum of the individual collinearities for the 10 T. adhaerens scaffolds multiplied by a size-correction faction for each scaffold (i.e., percent coverage of the evaluated 70.4 Mb of the T. adhaerens genome).
Syntenic block sizes and the number of blocks were calculated using the custom Python script microsynteny.py (described in [118]) with skipping no more than 1 gene [added option: -s 1] and otherwise default options.
To identify allele sharing or reproductive isolation, 3 genes encoding for ribosomal proteins were amplified via PCR, using degenerate primers designed based on the T. adhaerens genomic sequence, as well as a previously sequenced EST library of lineage H4 [19]. Primer sequences to amplify gDNA (including intronic sequence) for the ribosomal proteins L9 and L32, as well as ribosomal protein P1, were as follows:
PCRs were run with an initial denaturation of 3 min at 94 °C; followed by 40 cycles of 30 s of denaturation at 94 °C, 30 s of annealing at 60 °C, and 1.5 min of elongation at 72 °C; and finished with a final elongation for 3 min at 72 °C. The BIOTAQ system was used (Bioline, London). A list of samples used for amplification is provided as S6 Table. Sequencing was performed by Macrogen (South Korea). Alleles were identified as double peaks in standard sequencing in the case of heterozygous alleles. The phasing of SNPs was inferred from homozygous sequences as well as the sequence of allelic variants in closely related haplotypes, for which phasing information was available because of the long Sanger reads.
To check for reproductive isolation and to identify conspecific isolates, haplowebs [123] were generated for each marker as well as a CM [54] for combined markers using the online tool HaplowebMaker (https://eeg-ebe.github.io/HaplowebMaker/; Spöri & Flot, in prep.). The resulting conspecificity scores were plotted in R using the heatmap3 package [124], sorted according to a UPGMA tree (JC69 model) of the three concatenated genomic sequences (with indels removed). If present, both alleles of an isolate were merged, and the consensus sequence was used to generate the tree.
dN/dS ratios—as well as fractions of unchanged codons, synonymous, and nonsynonymous sites—were calculated based on a custom Python script (alignmentdnds.py) using regapped CDS alignments and untrimmed protein alignments (S6 Fig). Codons with any ambiguous bases and gapped sites were ignored.
Clustering into homologs and co-orthologs was performed with a custom python script (makehomologs.py) [added options:-s 1 -p 234 -H 200]. The script calls the MCL v12-068 algorithm [127], which uses the output of a local all-versus-all BLASTP search [added options: -evalue 1e-3 -outfmt 6] of all H. hongkongensis and T. adhaerens proteins.
To identify enriched GO terms in non-BLAST hits as well as in four co-ortholog categories (one-to-many, many-to-one, many-to-many, and many-to-zero), an enrichment analysis was performed for the three main GO categories (Biological Process, Cellular Component, and Molecular Function) using topGO [128]. Only enriched GO terms with a p-value <0.05 were kept, based on the classic Fisher test.
Ortholog categories (see also [40]) are defined as (1) one-to-one: Only one ortholog is found in each species; (2) one-to-many: One ortholog in this species, but many co-orthologs in the other species. The gene was duplicated in the other species from the ancestral copy after speciation; (3) many-to-one: More than one co-ortholog in this species but only one in the other species. The gene was duplicated in this species from the ancestral copy after speciation; (4) many-to-many: More than one co-ortholog in this and the other species. At least two gene duplications could be found from an ancestral gene in the common ancestor of both species—one duplication in this species, and a second one in the other species; (5) many-to-zero: Many co-orthologs in this species but none in the other. In this case, the gene was duplicated from an ancestral copy in this species after speciation and likely lost in the other species.
To estimate molecular differences between H. hongkongensis and T. adhaerens and to bring these into a taxonomic context, we measured genetic distance using an extended data matrix of 212 nuclear proteins set up by Cannon and colleagues [55]. This data matrix was chosen as it includes a comparable number of sites for a diverse taxonomic range and is, therefore, also suitable for phylogenetic analyses. In addition, genetic distances were measured for 5 standard barcoding (“selected”) markers—namely, nuclear ribosomal subunits 18S (S9 Fig) and 28S (S10 Fig), mitochondrial large ribosomal subunit 16S (S11 Fig), and the mitochondrial proteins cytochrome c oxidase subunit 1 (CO1) (S12 Fig) and NADH dehydrogenase subunit 1 (ND1) (S13 Fig). An overview of means for all distances of all six marker sets is provided as S14 Fig. The incorporation of datasets from four individual categories (nuclear protein versus nuclear rDNA versus mitochondrial protein versus mitochondrial rDNA) enabled the comparison among markers with different substitution rates.
To assess the effect of adding a second placozoan species on the placement of the Placozoa in the animal tree of life and to estimate branch lengths to the two placozoan species, dataset 1 was further condensed to generate a highly complete protein matrix (dataset 2). This set had only 10.8% missing characters in 58 taxa, including 32 nonbilaterians and 2 outgroups with an almost complete gene set (194 genes, see also gene density matrix in S21 Fig).
It has been demonstrated that the CAT model (specifically CAT-GTR) implemented in PhyloBayes [132] fits phylogenomic amino acid supermatrices containing nonbilaterians best [73,133], and obviously, only best-fitting evolutionary models should be used in probabilistic phylogenetic analyses to reduce systematic errors [133]. However, the computational burden of reaching convergence of analyses using the CAT-GTR model can be prohibitive. It is also well known that phylogenomic datasets frequently suffer from compositional heterogeneity that might negatively influence phylogeny estimation [134–136]. Compositional heterogeneities can be reduced by the so-called Dayhoff recoding [65,137,138], which combines amino acids with similar physicochemical properties into one of six categories. Through this reduction of character space, lineage-specific compositional heterogeneities are lessened—at the cost, however, of losing phylogenetic signal [67]. However, another advantage of Dayhoff recoding is a significant reduction of computation time needed to reach convergence.
The protein as well as the Dayhoff 6-state recoded dataset 2 were analyzed with PhyloBayes-MPI v1.7 [69,132], employing the CAT-GTR model, on the Linux cluster of the Leibniz Rechenzentrum (http://www.lrz.de) in Garching bei München, running 2 chains (each on 112 CPUs) each until reaching convergence, as estimated by using tracecomp and bpcomp programs of the PhyloBayes package (see PhyloBayes manual for details).
Furthermore, to evaluate the effect of using less-fitting site-homogeneous evolutionary models on the phylogenetic relationships of the Placozoa, we conducted a PhyloBayes-MPI analysis as above but with the GTR model (see for example [73], [68]), and also two maximum-likelihood analyses in RAxML: one with the GTR model using RAxML-NG v0.5.1b [139] [added options:—model PROTGTR+G+I—bs-trees 100—data-type AA] and one with the LG model using RAxML v8.2 [70] [added options: -f a -x 670 -m PROTGAMMAILG -p 220 -N 100]. The LG model was used as it was the best-fitting site-homogeneous model in 210 of the 212 gene partitions determined by ProtTest v3.4 [140]. Phylogenetic trees are shown as S15–S20 Figs.
|
10.1371/journal.pntd.0006772 | Ligand binding properties of two Brugia malayi fatty acid and retinol (FAR) binding proteins and their vaccine efficacies against challenge infection in gerbils | Parasitic nematodes produce an unusual class of fatty acid and retinol (FAR)-binding proteins that may scavenge host fatty acids and retinoids. Two FARs from Brugia malayi (Bm-FAR-1 and Bm-FAR-2) were expressed as recombinant proteins, and their ligand binding, structural characteristics, and immunogenicities examined. Circular dichroism showed that rBm-FAR-1 and rBm-FAR-2 are similarly rich in α-helix structure. Unexpectedly, however, their lipid binding activities were found to be readily differentiated. Both FARs bound retinol and cis-parinaric acid similarly, but, while rBm-FAR-1 induced a dramatic increase in fluorescence emission and blue shift in peak emission by the fluorophore-tagged fatty acid (dansyl-undecanoic acid), rBm-FAR-2 did not. Recombinant forms of the related proteins from Onchocerca volvulus, rOv-FAR-1 and rOv-FAR-2, were found to be similarly distinguishable. This is the first FAR-2 protein from parasitic nematodes that is being characterized. The relative protein abundance of Bm-FAR-1 was higher than Bm-FAR-2 in the lysates of different developmental stages of B. malayi. Both FAR proteins were targets of strong IgG1, IgG3 and IgE antibody in infected individuals and individuals who were classified as endemic normal or putatively immune. In a B. malayi infection model in gerbils, immunization with rBm-FAR-1 and rBm-FAR-2 formulated in a water-in-oil-emulsion (®Montanide-720) or alum elicited high titers of antigen-specific IgG, but only gerbils immunized with rBm-FAR-1 formulated with the former produced a statistically significant reduction in adult worms (68%) following challenge with B. malayi infective larvae. These results suggest that FAR proteins may play important roles in the survival of filarial nematodes in the host, and represent potential candidates for vaccine development against lymphatic filariasis and related filarial infections.
| Human lymphatic filariasis (LF) and river blindness (onchocerciasis) are highly debilitating neglected tropical diseases. As with all parasitic nematodes, Wuchereria bancrofti, Brugia malayi and Brugia timori, the etiological agents of LF and Onchocerca volvulus the causative agent of river blindness, possess limited lipid metabolic pathways and hence rely on lipids scavenged from their human hosts. Two unusual lipid-trafficking proteins from Brugia malayi (Bm-FAR-1 and Bm-FAR-2) were expressed as recombinant proteins, and their ligand binding activities along with their structural characteristics were examined. Their immunogenicity in infected children and those who are classified as endemic normal or putatively naturally immune were also evaluated. Their immunogenicity and immunoprotective efficacies were also evaluated in a B. malayi gerbil infection model. The possible role these proteins play in the survival of filarial nematodes in the host, and their prospects of being candidates for vaccine against these highly pathogenic infections are discussed.
| Human lymphatic filariasis (LF) and river blindness (onchocerciasis) are highly debilitating diseases in tropical developing countries with an estimated disease prevalence of 29.38 and 14.65 million cases that cause 1.2 and 0.96 million years lived with disability (YLD), respectively [1]. As with all parasitic nematodes, the etiological agents of LF such as Wuchereria bancrofti, Brugia malayi and Brugia timori, and that of river blindness, Onchocerca volvulus, possess limited lipid metabolic pathways and hence rely on lipids scavenged from their hosts [2]. Several structurally novel families of lipid-binding proteins in nematodes have been reported [3], including the fatty acid- and retinoid-binding protein family (FAR) that have been identified from many species of filarial nematodes including those from the genera Onchocerca, Brugia, Wuchereria, Loa, Acanthocheilonema and Litomosoides [4]. FAR proteins represent a structurally novel class of approximately 20 kDa lipid-binding proteins that are only found in nematodes [5], isoforms of which are known to be differentially expressed during development of parasitic and free-living species [5–7].
Ov-FAR-1 (previously known as Ov20 or Ov-RBD-1), a FAR protein from the filarial nematode O. volvulus was initially identified as a 20 kDa, structurally novel small helix-rich fatty acid and retinol (vitamin A)-binding protein secreted by the adult worm [8]. Soon thereafter, Bm-FAR-1, was described in B. malayi, and the two proteins were found to have similar secondary structure and ligand-binding characteristics [2, 4]. PCR-based strategies have since been used to isolate cDNAs encoding FARs from the filarial nematodes Onchocerca, Brugia, Wuchereria, Loa, Acanthocheilonema and Litomosoides [2, 4].
The ligand-binding properties of the filarial FAR proteins have been suggested to contribute, not only to their survival in the host, but also to pathogenesis in mammalian hosts [5, 8, 9]. These parasites appear to require retinoids and fatty acids for a variety of metabolic and developmental needs, including growth, development, differentiation, embryogenesis, and glycoprotein synthesis [2, 5, 10, 11]. FAR proteins have been shown to be released from the parasites into their hosts [2, 8, 12], suggesting that their FARs may also play an important role in modifying the local inflammatory and immunological environment of the surrounding host tissue by sequestering and/or delivering pharmacologically active lipids [5, 12]. Relevant to this hypothesis is the finding of high concentrations of retinol within onchocercal nodules [13]. Given the role of retinoids in vision, tissue differentiation and collagen synthesis [9], such sequestration of retinol might exacerbate vitamin A deficiency in infected humans, thereby contributing to the clinical manifestation of river blindness. It has been found that patients with onchocerciasis have lower serological level of vitamin A [14, 15].
The probable dependence of the filarial parasites on the FAR proteins for metabolic needs, and their potential roles in development and immune modulation of the host makes them pertinent targets for anthelmintic drugs and vaccine development. We therefore produced two FAR proteins from B. malayi in recombinant forms, biophysically characterised their hydrophobic ligand binding properties, and tested their immunogenicity and immunoprotective efficacy against infection with B. malayi infective larvae in gerbils. We found that despite their amino acid sequence relatedness and similar structural characteristics, and a precedent in another species, rBm-FAR-1 and rBm-FAR-2 could be clearly discriminated by their ligand-binding properties. We provide evidence that this disparity applies widely in filarial parasites. Immunization with recombinant Bm-FAR-1 conferred significant protection against infection in gerbils, but only when formulated in a water-in-oil adjuvant rather than alum. Our results indicate that Bm-FAR-1 is a candidate for a vaccine development against lymphatic filariasis, and related filarial helminthiases, and that Bm-FAR-2 might have to be further optimized before confirmed to being a promising vaccine candidate.
All the animals in this study were handled according to the National Institutes of Health (USA) guidelines and the animal experimentation was performed with prior approval from the Louisiana State University Institutional Animal Care and Use Committee under the protocol number 12–037.
The protocols used in all the population studies were approved by the Institutional Review Board (IRB) of the National Institute of Allergy and Infectious Diseases (Cook Islands studies, clinical protocol number 92-I-0155), the New York Blood Center's IRB (clinical protocol number 321 and 603–09), and by the National Institutes of Health (USA) accredited Institutional Review Board of the Medical Research Council Kumba, Cameroon (Kumba studies, clinical protocol number 001). Informed written consent was obtained from all adult subjects, and for children consent was obtained through both verbal assent and written consent from each subject’s legal guardian.
The encoded amino acid sequences of Bm-FAR-1 (GenBank accession# XP_001899742) and Bm-FAR-2 (XP_001900470) were aligned with FAR proteins from other nematodes including Wuchereria bancrofti (Wb-FAR-1:Q8WT54.2; Wb-FAR-2:EJW79208.1), Onchocerca volvulus (Ov-FAR-1: Q25619.1; Ov-FAR-2: ACT55269.1), Loa loa (L1-FAR-1: AAK84218.1; Ll-FAR-2: XP_003137038.2), Dirofilaria immitis (Di-FAR-1: nDi.2.2.2.t00119*; Di-FAR-2: nDi.2.2.2.t02086*); Brugia pahangi (Bp-FAR-1: Q8WT55.2; Bp-FAR-2: BPAG_0000817001-mRNA-1*), Litomosoides sigmodontis (Ls-FAR-1: Q8WT56.2; Ls-FAR-2: nLs.2.1.2.t01040-RA*), Acanthocheilonema viteae (Av-FAR-1: Q8MZJ8.1; Av-FAR-2: nAv.1.0.1.t08895-RA*), Ascaris suum (As-FAR-1: ERG87764.1; As-FAR-2: ERG85173.1), Ancylostoma caninum (Ac-FAR-1: AAM93667.1; Ac-FAR-2: ANCCAN_08556*), Necator americanus (Na-FAR-1 (PDB: 4UET_A, XP_013293708.1; Na-FAR-2: NECAME_14206*), Toxocara canis (Tc-FAR-1: KHN72925.1; Tc-FAR-2: KHN88420.1), Haemonchus contortus (Hc-FAR-: CDJ83169.1; Hc-FAR-2: HCOI00378700.t1*), Caenorhabditis elegans (Ce-FAR-1: NP_001254978.1; Ce-FAR-7 (PDB: 2W9Y_A), Strongyloides ratti (St-FAR-1: CEF68237.1), Ostertagia ostertagi (Oo-FAR-1: CAD20464.1). Similar sequences marked with [*] were obtained by BLAST searching nematode genomes databases at Wormbase (http://parasite.wormbase.org/index.html). A phylogenetic tree was generated using MEGALIGN from Lasergene 14, DNASTAR), and rendered using FigTree 1.4.3. For structural comparison of Bm-FAR-1 and Bm-FAR-2 with structure-defined orthologues Na-FAR-1 (PDB: 4UET_A) [16] and Ce-FAR-7 (PDB: 2W9Y_A) [17], the sequences were aligned with Clustal Omega and the secondary structural features were predicted based on the coordinates of Na-FAR-1 and Ce-FAR-7 using ESPript [18]. The structural theoretical ‘homology’ models of Bm-FAR-1 and Bm-FAR-2 were generated using Swiss model [19, 20].
For production of recombinant proteins, the DNAs encoding Bm-FAR1 and Bm-FAR-2 (minus their signal peptides) were codon optimized for yeast preference and synthesized by GenScript (Piscataway, NJ, USA), and then inserted in-frame into the PichiaPink expression vector pPinkα-HC (Invitrogen, Carlsbad, USA) using XhoI/KpnI sites. The recombinant proteins with 6His-tag at C-terminus were induced in PichiaPink strain 4 (with protease PEP4 and PRB1 deleted to reduce degradation) with 5% methanol and purified with immobilized metal affinity chromatography (IMAC) as previously described [21, 22]. Due to the high levels of glycosylation of the yeast expressed Bm-FAR-2 recombinant protein, it was re-cloned in E. coli expression vector pET41a (Novagen, USA) with glutathione-S-transferase-tag deleted (NdeI/XhoI), and then transformed into BL21(DE3) cells (Novagen, USA). Recombinant Bm-FAR-2 protein (rBm-FAR-2) was induced with 1 mM IPTG and purified with IMAC as described [23]. The rBm-FAR-1 expressed in yeast, and the rBm-FAR-2 expressed in E. coli were used for subsequent binding activity and vaccination experiments. The recombinant proteins of FAR orthologues in O. volvulus (rOv-FAR-1 and rOv-FAR-2) were produced in E. coli using similar procedures as rBm-FAR-2.
Stage-specific proteomic expression patterns for the adult B. malayi male (AM) and female (AF) parasites, microfilariae (MF), immature MF (intrauterine stages; UTMF), and the third-stage larvae (L3) were derived from the B. malayi somatic proteomes [24], and normalized using normalized spectral abundance factor (NSAF), where the relative abundance of a protein in a sample was calculated by: NSAF=(SpectraLength)p∑p=1n(Spectralength)p.
The in-vitro derived L4 [25] were lysed in lysis buffer, dialyzed, desalted, and digested with trypsin. The cation-exchange liquid chromatography fractionation of tryptic peptides was analyzed by nanobore reverse-phase liquid chromatography (RPLC-MS/MS). Proteins were identified by searching the spectral data using PEAKS 7 using a combined database of B. malayi (Wormbase WBPS9 ver) and its endosymbiont Wolbachia (wBm).
A Jasco-J1100 CD Spectrophotometer was used for measurements in the far ultraviolet region (UV), from 190 to 260 nm. Spectra were recorded at protein concentrations of approximately 0.1–0.2 mg ml-1 in a cuvette of 0.2 mm path length in a temperature-controlled cell holder at 25°C. Spectra were monitored with a 0.2-nm step with 10 averages per step. Samples were prepared in 20 mM Tris, 20 mM NaCl, pH 6.8 buffer, and all spectra were baseline corrected by subtraction of the spectra for buffer alone. Analysis of the CD spectra was undertaken using the CDSSTR within the Dichroweb program [26, 27].
Recombinant FAR proteins of B. malayi and O. volvulus (rBm-FAR-1, rBm-FAR-2, rOv-FAR-1 and rOv-FAR-2), were used in fluorescence-based lipid binding assays as previously described [5, 12, 28]. The concentrations of the proteins were estimated by absorbance at 280 nm (correcting for any untoward absorbances at 230 and 260 nm), using theoretical extinction coefficients based on their amino acid compositions, and calculated using the ProtParam tool online at http://web.expasy.org/protparam. The protein concentrations of all were calculated to be at approximately 4 mg ml-1. Lipid binding was detected spectrofluorometrically, using all-trans retinol, or the fluorescent fatty acid analogue 11-((5-dimethylaminonaphthalene-1-sulfonyl)amino)undecanoic (DAUDA), which bears the environment-sensitive dansyl fluorophore, or with the intrinsically fluorescent cis-parinaric acid (cPnA), all as described before [5]. DAUDA and cPnA were obtained from Molecular Probes/Invitrogen (Renfrew, UK). All-trans-retinol and oleic acid were obtained from Sigma (Poole, Dorset, UK). The excitation wavelengths were 345 nm, 350 nm, and 319 nm for DAUDA, retinol, and cPnA, respectively, all of which were at final concentrations of approximately 1 μM, 4 μM, and 4 μM, respectively, in 2 ml phosphate buffered saline (PBS) pH 7.2 in a quartz fluorescence cuvette. Binding of a non-fluorescent ligand was detected by a reversal of fluorescence emission enhancement elicited by a test protein upon addition of the ligand to a preformed DAUDA:protein complex. Note that rBm-FAR-2 is larger than rBm-FAR-1, so at equivalent w/v concentrations their molarities will differ. Also, we cannot assume that the proportion of properly folded and active protein in each protein sample is equivalent, or that there is no interference from resident ligand(s) derived from the bacteria in which the proteins were produced despite routine detergent and lipid depletion–our experience is that complete removal of resident lipids from nematode lipid-binding proteins requires methods such as reverse-phase chromatography [29], which was not carried out in the present study. The fluorescence spectra are uncorrected and were analyzed using MICROCAL ORIGIN software.
The serum samples used for the present studies were from a repository from two distinct clinical studies in which plasma samples were collected from two different populations. First, two population studies performed in 1975 and 1992 in the island of Mauke in the Southern Cook Islands, a region endemic for the filarial parasite W. bancrofti. The characteristics of the population studied have been described in detail in previous publications [30, 31]. Second, a clinical study performed between 1995 and 2000 in the Kumba region of southwest Cameroon, an area of hyperendemicity for onchocerciasis. The characteristics of this population have also been described previously [32, 33]. Antibody responses to rBm-FAR-1 and rBm-FAR-2 were tested in children (age 2–9) that were infected (INF; n = 21; microfilaremic and/or positive for circulating filarial antigen) or classified as uninfected or endemic normal (EN; n = 22). Antibody responses to rOv-FAR-1 and rOv-FAR-2 were tested in putatively immune (PI; n = 12; age 5–59), and age and sex match infected (INF; n = 30; 6–41) individuals.
Sera obtained from the individuals were analyzed for IgG1, IgG3 and IgE isotype antibody responses using recombinant Bm-FARs and Ov-FARs using ELISA protocols established for testing antibody responses to filarial recombinant proteins [30–32, 34–36], with some modification. Briefly, each recombinant protein (1 μg/ml) was used to coat the wells of ELISA plates, and sera at a 1:100 (IgG1 and IgG3) or at 1:50 (IgE) dilution were applied. The bound IgG1 or IgG3 antibodies were detected by using a 1:1,000 dilution of monoclonal antibodies against human IgG1 and IgG3 subclasses (Hybridoma Reagent Laboratory, Kingsville, MD). This step was followed by incubation with a 1:1,250 dilution of horseradish peroxidase (HRP)-conjugated rabbit anti-mouse immunoglobulins (Kierkegaard & Perry Laboratories, Inc., Gaithersburg, Md.). Tetramethylbenzidine (Sigma) was used as the chromogen, and the optical density (OD) was read at 450 nm. For IgE responses, 1:1,000 dilution of monoclonal biotin conjugated anti-human IgE antibody (Hybridoma Reagent Laboratory, Kingsville, MD) followed by incubation with 1:1,000 Streptavidin -HRP (Invitrogen) and 0.4 mg/ml of o-Phenylenediamine dihydrochloride (Sigma) dissolved in phosphate citrate buffer with 0.05%H2O2. The optical density (OD) was read at 492nm. A pool of 10 normal human sera from de-identified New York blood donors was used as a negative control, and for setting up the cutoff; mean OD ± 3X SD.
Mongolian gerbils, obtained from Charles River (Wilmington, MA, USA) at 8–10 weeks of age, were maintained on standard rodent diet and water ad libitum. Infective third-stage larvae (L3) of B. malayi were recovered from infected Aedes aegypti mosquitoes using the previously described Baermann technique [37]. All the animals in experimental and control groups received 100 B. malayi L3 subcutaneously in 0.5 ml of RPMI-1640 medium.
The rBm-FAR-1 and rBm-FAR-2 were formulated with alum (Rehydragel LV, General Chemical, NJ) at ratio of 1:8 (w/w) (25 μg of recombinant protein was mixed with 200 μg of alum) for 30 min at room temperature with shaking [21]. The virtually complete binding of recombinant FAR proteins to alum was confirmed by SDS-PAGE. Ten 8–10 weeks old male Mongolian gerbils in each group were immunized intraperitoneally (IP) with 25 μg of rBm-FAR-1 formulated with alum as described above or emulsified with Montanide-720 (Seppic, Paris, France) at ratio of 30/70 (v/v) in a total volume of 100 μl. Gerbils were boosted twice with the same dose at two-week intervals. Another two groups of gerbils were intraperitoneally injected with PBS formulated with the same adjuvants (alum or Montanide-720) as controls using the same regimen. Sera were collected one week after each immunization. IgG responses to rBm-FAR-1 or to rBm-FAR-2 were measured as previously described [38]. Serial dilutions of gerbil serum were made, and total antigen-specific IgG response to recombinant antigens were evaluated for each group.
Two weeks after the third immunization, all gerbils in each group were challenged subcutaneously with 100 infective L3 of B. malayi. Necropsy was performed 42 days post-infection and adult worms were recovered from different body regions of gerbils as described previously [38, 39]. Protective immunity induced by immunization is expressed as the percentage reduction in worm burden as calculated by subtracting the average number of worms recovered in immunized gerbils from the average number of worms recovered from the adjuvant control gerbils divided by average worms recovered from control gerbils and then multiplied by 100.
Analysis of human antibody response data was carried out using the non-parametric Mann-Whitney test to examine statistical significance between the immune responses of infected and EN (uninfected) and putatively immune individuals to FAR proteins. For the immunization experiments in gerbils, every experiment comprised 10 gerbils per group, and statistical analysis between antigen-vaccinated groups and adjuvant control groups were performed using the Mann-Whitney U test in GraphPad Prism 6 (GraphPad Software, San Diego, California USA). Data were expressed as means ± standard deviation. A value of p < 0.05 was considered as statistically significant.
Bm-FAR-1 was the first fatty acid and retinol binding protein identified in B. malayi with α-helix-rich structure and lipid binding activity similar to Ov-FAR-1, its orthologue from O. volvulus [2, 5]. BLAST searching of the B. malayi genome database [40] revealed a paralogous FAR protein, here designated Bm-FAR-2. Both Bm-FAR-1 and Bm-FAR-2 contain the consensus casein kinase II phosphorylation site which has been found in the same position in the amino acid sequence in other FAR proteins [2, 7, 41]. The present study is the first to our knowledge that characterizes a FAR-2 protein from a parasitic nematode.
Phylogenetic analysis shows that Bm-FAR-1 and Bm-FAR-2 fall into different clades of nematode FARs. Bm-FAR-1 is closely related to putative FAR-1 homologues from other filariae such as B. pahangi (Bp-FAR-1, 99% amino acid identity), W. bancrofti (Wb-FAR-1, 97%), and O. volvulus (Ov-FAR-1, 82%) [2]. Bm-FAR-2 shares only 27% amino acid sequence identity to Bm-FAR-1 and possesses an unique N-linked glycosylation site at 53-NFS and 70 additional C-terminus amino acids, but instead is more similar to FAR-2 orthologues from closely related species (88% sequence identity with Wb-FAR-2, and 63% with Ov-FAR-2) (Fig 1A and 1B). Molecular structures of two nematode FARs have been reported, one from Necator americanus (Na-FAR-1 by protein nuclear magnetic resonance (NMR) and x-ray crystallography; PDB accessions 4UET and 4XCP, respectively) [16], and another from Caenorhabditis elegans (Ce-FAR-7, by x-ray crystallography; PDB: 2W9Y) [17]. Notably, the molecular structure of Na-FAR-1 revealed a surface-proximal and a deeper binding pocket for hydrophobic ligands that are accessible to palmitate molecules.
Both Bm-FAR-1 and Bm-FAR-2 share less than 32% amino acid sequence identity with either Na-FAR-1 or Ce-FAR-7. However, these structures proved to be useful for generating theoretical models of Bm-FAR-1 and Bm-FAR-2 using Swiss-Model [19, 20]. The structure-based comparison of the sequences using ESPript [18] reveals that the amino acids in the binding pockets of Bm-FAR-1 are more similar to Na-FAR-1, while those of Bm-FAR-2 are more similar to Ce-FAR-7 (Fig 1B). The two distinct pockets in filarial nematode FARs are evident in the predicted structures (Fig 1B & 1C). The deeper pocket (P1) may bind retinol whereas the exterior pocket (P2) may bind fatty acids [16, 17]. The putative retinol-binding pocket P1 is formed by helices α2 and α6, which are well conserved (Fig 1B). The putative fatty acid binding cavity P2 is formed by helices α4 and α5, which are poorly conserved (Fig 1B). While the putative retinol-binding pocket (P1) is similarly located and oriented in both models, the potential fatty acid binding cavity (P2) is smaller and less accessible in Bm-FAR-2 than Bm-FAR-1 (Fig 1C), which may explain the differences between the two proteins in their binding of one of our fluorescent fatty acid probes (see below).
The normalized spectral abundance levels of Bm-FAR-1 and Bm-FAR-2 in the lysates of different stages of B. malayi showed that the mature microfilariae (MF) and the immature uterine microfilariae (UTMF) only exhibited very low-levels of either Bm-FAR-1 or Bm-FAR-2. Overall, Bm-FAR-1 was more abundant than Bm-FAR-2 across all the stages, with the adult male worm and L4 stages exhibiting the highest abundance (4-fold and ~7-fold respectively) (Fig 2). The similar abundances of both proteins in the adult female (containing both mature and immature MFs) suggest that the source of the proteins is primarily the adult female.
The rBm-FAR-1 protein was expressed in yeast and rBm-FAR-2 was expressed in E. coli. The purified rBm-FAR-1 and rBm-FAR-2 were observed to migrate at their corresponding molecular sizes of ~19 kDa and 27kDa, respectively, on SDS-PAGE (Fig 3). The additional lower molecular size bands for rBm-FAR-2 were reactive to anti-His antibody and so were likely degradation products of the protein (S1 Fig). Both rBm-FAR-1 and rBm-FAR-2 were essentially fully absorbed by the alum-based adjuvant (Fig 3).
Both rBm-FAR-1 and rBm-FAR-2 were examined using circular dichroism, which showed that both proteins are α-helix-rich, with negative molar ellipticity (Ɵ) values at 222 nm and 208 nm, and a positive value at 193 nm [42]. The calculated helicity for both proteins is 75%, while the control protein rOv-FAR-2, the presumed orthologue of Bm-FAR-2 from O. volvulus, contained 67% helicity (Fig 4). The high degree of helical structure for both Bm-FAR-1 and Bm-FAR-2, as estimated by CD, are similar to the predictions from homology modelling and sequence analyses, which indicate a predominance of α-helices (Fig 1B).
The fatty acid and retinol binding properties of rBm-FAR-1 and rBm-FAR-2 were investigated using a fluorescence-based lipid binding assay. Addition of retinol to either rBm-FAR-1 or rBm-FAR-2 resulted in substantial increases in fluorescence emission, indicative of entry of retinol into an apolar protein binding site (Fig 5A). When the fluorescent fatty acid analogue, 11-(dansylamino) undecanoic acid (DAUDA), was used as test ligand, rBm-FAR-1 produced a large increase in DAUDA emission and a substantial blue-shift in the wavelength of peak fluorescence emission from 542 to 483 nm (Fig 5B). This degree of blue shift in DAUDA’s emission is indicative of entry into a highly apolar environment and isolation from solvent water. In contrast, rBm-FAR-2 produced a mild enhancement in DAUDA fluorescence, indicating that rBm-FAR-2 interacts poorly with DAUDA, or that it does bind this ligand but in a way that leaves the attached dansyl fluorophore still in a polar environment (Fig 5B). Addition of oleic acid to preformed DAUDA: rBm-FAR-1 complexes resulted in a significant decrease in the fluorescence intensity, indicating displacement of DAUDA from the binding site into polar solvent, and probable congruence between the binding sites for the two lipids (Fig 5C). Both rBm-FAR-1 and rBm-FAR-2 bound cPnA, which is an intrinsically fluorescent fatty acid that does not bear a bulky attached fluorophore (as is the case for DAUDA) (Fig 5D). The enhancement of cPnA’s fluorescence emission with rBm-FAR-2 was less than that with rBm-FAR-1, which may be due to fundamental differences in their hydrophobic ligand-binding properties, or that the former protein’s population may not have been proportionately as well folded as the latter’s (despite their retinol- and cPnA-binding activities being similar here), or that the environment of their binding sites for cPnA are slightly different. Essentially, however, both rBm-FAR-1 and rBm-FAR-2 bind fatty acids and retinol (and possibly other hydrophobic ligands), but can be discriminated by activity with DAUDA.
In parallel experiments, rOv-FAR-1 and rOv-FAR-2, which are presumed orthologues of rBm-FAR-1 and rBm-FAR-2, respectively, behaved in the fluorescence assays similarly to their counterpart Bm-FARs (S2 Fig).
When the rBm-FARs antigen-specific IgG1 and IgG3 responses in sera from infected (INF) or exposed but uninfected/endemic normal (EN) children living in W. bancrofti endemic region, and the rOv-FARs antigen-specific IgG1, IgG3 and IgE responses in sera from O. volvulus putatively immune (PI) or infected individuals (INF) were performed (Fig 6), it appears that both FAR proteins are immunogenic in all groups tested; no statistical differences were observed between the INF and EN children living in W. bancrofti endemic region (Fig 6A and 6B) or the PI and INF individuals living in onchocerciasis endemic region (Fig 6C and 6D).
However, it appears that both proteins are more immunogenic in the O. volvulus endemic population as the number of IgG1 and IgG3 responders to rOv-FAR-1 and to rOv-FAR-2 antigens was similarly high in both the PI and INF groups (>92% IgG1 responders to both antigens; >93% IgG3 responders to both antigens). The IgE responses were more modest; the number of IgE responders to rOv-FAR-1 and to rOv-FAR-2 antigens in the PI was 58.3% and 33.3%, respectively (Fig 6C and 6D), while the number of IgE responder in the INF was 76.7% and 46.7%, respectively (Fig 6E). In comparison, only the number of the IgG3 responders to rOv-FAR-2 antigen in the infected (INF) or the exposed but uninfected (EN) children from the W. bancrofti endemic region was higher; 90.4% and 72.7%, respectively. The number of IgG1 responders to rBm-FAR-1 and to rBm-FAR-2 antigens was only 50%-66.6%, while the number of IgG3 responders to rBm-FAR-1 antigens was 40.9–66.6%. No IgE responses were detected against any of the rBm-FAR-1 and to rBm-FAR-2 in both populations.
After immunization with 25 μg of rBm-FAR-1 formulated with alum or Montanide-720 for three times, and following challenge with 100 infective B. malayi L3, a significant reduction in worm numbers (68.4%, p = 0.0063) was observed only in the group of gerbils immunized with rBm-FAR-1 formulated with Montanide-720 (Fig 7A). After immunization with 25 μg of rBm-FAR-1 formulated with alum or Montanide-720 for three times, and following challenge with 100 infective B. malayi L3 in two separated experiments, a combined total worm reduction of 68.4% was observed with statistical significance (p = 0.0063) in the group of gerbils immunized with rBm-FAR-1 formulated with Montanide-720 compared to adjuvant control group (Fig 7A). The worm reductions for individual rBm-FAR-1/Montanide-720 immunization were 68.0% (p = 0.0139) and 44.4% (p = 0.0541), respectively, compared to adjuvant control. Though a trend of reduction (38.5%) was observed with the alum formulation, it was not statistically significantly different (p = 0.1458) from its corresponding alum control (Fig 7A). The efficacy outcomes have been assembled from two experiments done in separate times. The differences in protection between the two vaccine formulation cannot be simply accounted for by the overall anti-IgG rBm-FAR-1 antibody titers; the titers in both vaccinated groups after three immunizations and before challenge were similar (geometric mean of 4.81 x 105 and 5.15 x 105, in alum or Montanide-720 formulated rBm-FAR-1 vaccines, respectively) (Fig 7B).
Under the same immunization regimen, gerbils immunized with rBm-FAR-2 formulated with alum or Montanide-720 only produced 16% and 42% of adult worm reduction, respectively, neither was statistically significant (p = 0.402 and p = 0.0612, respectively) compared with the respective adjuvant groups (Fig 6C, Table 1). In both immunized groups high antibody titers against rBm-FAR-2 were nevertheless detected (geometric mean of 4.84 x 105 and 5.94 x 105, respectively) (Fig 6D).
The unusual ligand binding properties and nematode-specific structures set FAR proteins apart from mammalian lipid-bind proteins [5, 7]. The objectives of this study were to assess the structural and ligand-binding characters of two FAR proteins from the filarial parasite B. malayi, and evaluate their potential as vaccine candidates in gerbils. This is the first FAR-2 protein from a parasitic nematode that is being characterized.
The binding properties of rBm-FAR-1 in this study were consistent with previously described results [2] in that rBm-FAR-1 was able to bind retinol and DAUDA. While rBm-FAR-2 also bound retinol, in unexpected contrast to rBm-FAR-1, it did not produce the dramatic change in DAUDA’s emission observed with rBm-FAR-1. Both proteins, however, bound cPnA, an unmodified intrinsically fluorescent fatty acid (Fig 5). The difference between the ligand binding propensities of the two proteins is surprising given previous experience with orthologous proteins in another species [7]. While there is no direct evidence of where in the two proteins retinol and fatty acids localize, the predicted tertiary structures of both proteins reveal differences that may explain the differential binding properties. Bm-FAR-1 and Bm-FAR-2 have similar putative retinol-binding pockets (P1), but Bm-FAR-2 has a smaller and less accessible predicted cavity for fatty acid binding (P2) than Bm-FAR-1. These possible structural differences are therefore consistent with the empirical finding that both proteins bind retinol similarly, but that rBm-FAR-2 is more selective in its binding of fatty acids. In addition to the difference in structure and binding activities, Bm-FAR-1 is more abundant in all the stages of B. malayi than Bm-FAR-2, especially in the adult male adult worms and L4 larval stages, suggesting that Bm-FAR-1 may play quite different role(s) in the maintenance of life cycle than Bm-FAR-2. In A. lumbricoides, a FAR protein named ABA-1 is a major allergen was found to bind retinol and a range of fatty acids, including the pharmacologically active lipids lysophosphatidic acid, lysoplatelet activating factor, and leukotrienes [43]. To investigate whether these proteins are also immunoregulatory, it will be of interest to use the fluorescence-based assay used here to screen both FAR proteins (and their orthologues in O. volvulus and other species) for binding to immunologically- and inflammatory-related lipids such as leukotrienes and prostaglandins, or their precursors such as arachidonic acid.
We found that both B. malayi and O. volvulus FAR proteins are the targets of strong cytophilic IgG1 and IgG3 antibody responses in infected humans, as well as in those individuals classified as being protected (the EN and PI individuals). There is a difference between an antigen being the target of immune responses in infected and naturally immune individuals and their immunoprotective activity. We therefore proceeded to test the immunoprotective activity of the Bm-FAR proteins in gerbils. The results of this indeed suggested that their vaccine potential merits more extensive exploration, but also that Bm-FAR-1 and -2 may perform different functions in parasite survival. Gerbils immunized with rBm-FAR1 elicited statistically significant protection against challenge of B. malayi L3, although this was dependent on the adjuvant used. Gerbils immunized with rBm-FAR-2, on the other hand, were less protected, regardless of which adjuvant was used, despite the IgG antibody titers induced against either protein or adjuvant formulation being similar. The value of FAR proteins in protective immunity has also been observed in immunization trials with a FAR homologue from O. volvulus (Ov-FAR-1) against heterologous challenge of rodent filaria A. viteae in Mongolian gerbils [44], and from the human and animal hookworms [45]. The mechanism underlying the protection and the difference in the vaccine efficacy of the two proteins remains to be determined. The difference in their ligand binding properties and the expression abundance between Bm-FAR-1 and Bm-FAR-2 suggests both FARs may function differently in the maintenance of the parasite’s life cycle and interaction with host. The Bm-FAR-1 protein is more abundant after molting of L3, by twofold, which may indicate that it is more vulnerable to immunological attack than Bm-FAR-2. It cannot be excluded that the immuno-protective difference between rBm-FAR-1 and rBm-FAR-2 may partially results from the different expression system, which will need to be further confirmed in future experiments. In both Bm-FAR-1 and Bm-FAR-2 immunization, Montanide-720 adjuvant induced better protection than did alum. Montanide-720 is a water-in-oil emulsion preparation that induces strong and broad Th1- and Th2-associated humoral and cellular immune responses [46, 47], while alum is known to stimulate a Th2-type humoral and cellular response [48]. The adjuvants used here may therefore elicit different immunoglobulin isotype dominances (e.g. different IgG isotype balances, or IgE) and T cell response biases that may contribute to the difference of protection. Due to the lack of reagents for detecting IgG subclasses and cytokines in gerbils, the nature and biases of immune response induced with different adjuvants remains to be examined. Protection against filarial infection in humans appears to be associated with a combination of IgG1, IgG3 and IgE antibody responses as well as Th1-and Th2 immune responses to larval proteins [49–51].
In conclusion, we demonstrated that Bm-FAR-1 and Bm-FAR-2 are two FAR proteins with a similar but slightly different structure, exhibit different expression abundance and lipid-binding characteristics, and confer different levels of protection against B. malayi infection. We found a similar dichotomy in ligand binding between the orthologous proteins in O volvulus, which may unexpectedly apply widely amongst filarial nematode parasites. B. malayi FARs, especially Bm-FAR-1 may play important roles in the survival of the filarial parasite in the host, therefore are promising vaccine candidates for a prophylactic vaccine against lymphatic filariasis.
|
10.1371/journal.pbio.2000410 | Selection for Mitochondrial Quality Drives Evolution of the Germline | The origin of the germline–soma distinction is a fundamental unsolved question. Plants and basal metazoans do not have a germline but generate gametes from pluripotent stem cells in somatic tissues (somatic gametogenesis). In contrast, most bilaterians sequester a dedicated germline early in development. We develop an evolutionary model which shows that selection for mitochondrial quality drives germline evolution. In organisms with low mitochondrial replication error rates, segregation of mutations over multiple cell divisions generates variation, allowing selection to optimize gamete quality through somatic gametogenesis. Higher mutation rates promote early germline sequestration. We also consider how oogamy (a large female gamete packed with mitochondria) alters selection on the germline. Oogamy is beneficial as it reduces mitochondrial segregation in early development, improving adult fitness by restricting variation between tissues. But it also limits variation between early-sequestered oocytes, undermining gamete quality. Oocyte variation is restored through proliferation of germline cells, producing more germ cells than strictly needed, explaining the random culling (atresia) of precursor cells in bilaterians. Unlike other models of germline evolution, selection for mitochondrial quality can explain the stability of somatic gametogenesis in plants and basal metazoans, the evolution of oogamy in all plants and animals with tissue differentiation, and the mutational forces driving early germline sequestration in active bilaterians. The origins of predation in motile bilaterians in the Cambrian explosion is likely to have increased rates of tissue turnover and mitochondrial replication errors, in turn driving germline evolution and the emergence of complex developmental processes.
| Mammalian germ cells (eggs and sperm) are immortal in the sense that they propagate successive generations. In contrast, somatic (body) cells do not persist to the next generation. Yet neither plants nor basal animals such as sponges and corals have a germline; they simply form gametes from stem cells in adult tissues. The reasons for these differences are unknown. We develop an evolutionary model showing that the germline evolved in response to selection on mitochondria, the powerhouses of cells. Mitochondria retain their own genes, which occur in multiple copies per cell. In plants and basal animals, the mitochondrial genes mutate slowly. Segregation over the many rounds of cell division to form an adult generates variation in mutant mitochondria between gametes, sufficient for natural selection to improve mitochondrial function. In more active animals from the Cambrian explosion onwards, the mitochondrial mutation rate rose strongly. This required the evolution of a dedicated germline, set aside early in development, with lowered mutational input. It also favoured large eggs (starting with thousands of mitochondria) and culling, following overproduction (atresia). Both devices maintain mitochondrial quality. The evolution of germline sequestration had profound consequences, allowing the emergence of complex developmental processes and truly disposable adult tissues.
| In distinguishing between the germline and soma, Weismann argued that the division of labour enabled the specialization of cells in somatic tissues, ultimately permitting greater organismal complexity [1]. In contrast, germline cells alone retain the capacity to provide genetic information for future generations and never form somatic cells [2]. Without the specialization enabled by the germline, complex multicellular animals with post-mitotic tissues such as brain might be impossible. But the division of labour cannot account for the origin of the germline, as all plants and many animals (including tunicates, flatworms, and Cnidaria [3]) have differentiated tissues but do not sequester a germline, instead generating gametes from pluripotent stem cells in somatic tissues (somatic gametogenesis).
The best-known hypothesis for the origin of germline sequestration relates to selfish competition between the cells of an individual. The strict distinction between germline and soma stabilises multicellular cooperation, as the only way for somatic cells to increase fitness is by cooperating with their kin in the germline [4–6]. According to this theory, plants did not evolve a germline because their rigid cell walls restrict cell movement, limiting the systemic effects of any parasitic cell lines [4]. In contrast, animal cells lack a rigid wall, making them more vulnerable to parasitic cell lines that undermine organismal function [4]. Sequestering a germline theoretically limits this competition. However, because cells in multicellular organisms normally derive from a single cell (unitary development), new selfish mutations must arise within a single generation; if these mutants are inherited then all cells in the offspring will carry the selfish mutation, so there are no longer any non-selfish cells to exploit [7,8]. The range of conditions under which selfish competition could give rise to the germline is therefore seriously restricted, if feasible at all [9].
A second line of thinking dating back to Weismann emphasises the protected environment of a sequestered germline [1,10]. By restricting mutations in gametes, the germline enhances germ-cell quality, contrasting with the “disposability” of the soma [11]. Avoiding the accumulation of mutations in nuclear genes associated with greater metabolic work can theoretically favour differentiation of a germline [12]. But testes have a high metabolic rate, and sperm are produced continuously through life (with 30 cell divisions by puberty and 400 divisions at the age of 30 in humans), contributing a large number of new nuclear mutations to offspring [13]. From this point of view, the male germline is equivalent to tissues such as bone marrow, merely specialising in the mass production of cells of a particular type. As all tissue precursor cells are spatially secluded—that is, sequestered—during early development in bilaterians, the male germline is distinct from other tissues only in that it produces gametes.
Focusing on the origin of the germline as a protected environment shifts the problem specifically to female gametogenesis. Across bilaterian animals, from nematode worms to mammals, primordial oocytes are sequestered early in development in a transcriptionally repressed state, with meiosis arrested in prophase I [14]. The mitochondria of primordial oocytes are held in a state of functional quiescence [15] before being amplified up to very large numbers during oocyte maturation, with as many as 106 copies of mitochondrial DNA in mature mammalian oocytes [16]. Mitochondrial DNA is usually inherited uniparentally; male mitochondria are excluded from the oocyte or destroyed on entry to the zygote [17]. Taken together, these traits suggest that mitochondrial function might have played a role in germline evolution.
The possibility that selection for mitochondrial quality could have driven the evolution of two sexes, in which female mitochondria are protected by germline sequestration, has been raised before [15,18–20] but never formally modelled. Female germline mitochondria are proposed to act as “inert templates,” held in a metabolically quiescent state throughout development, whereas male germline mitochondria are damaged through use and so must be destroyed [15,18–20]. There are some difficulties with this hypothesis. First, mitochondria are typically inherited uniparentally (from one “sex”) even in isogamous organisms, where both gametes are motile, and so equally damaged through use [21,22]. Second, two sexes also exist in plants and basal metazoans, which produce oocytes and spermatozoa from the same stem cell lineage that gives rise to adult somatic cells, but do not sequester a germline [23]; no inert template mitochondria are sequestered early in development. Third, in bilaterians that do sequester a germline, most mitochondrial mutations seem to be produced by copying errors rather than reactive oxygen species (ROS) formed by respiration [24–27], so active use is only part of the problem. Finally, while female germline mitochondria seem to be inactive in some systems [15], they appear to be active in others, such as bivalve molluscs [28,29], and may generally be active in mature oocytes and during early development in mammals [30,31]. Overall, germline sequestration is plainly a secondary adaptation—uniparental inheritance and oogamy arose before oocyte sequestration in early development, and the evolution of two sexes cannot simply be a matter of protecting template mitochondria.
Here we investigate how early or later sequestration of gametes could improve mitochondrial quality. We propose that the key to germline evolution lies in the balance between mitochondrial mutations, which undermine fitness, and segregation of mutants, which generates variance, facilitating selection. Both are amplified with each round of cell division with the outcome depending on the relative influence of multiple factors. Our analysis reveals that selection for mitochondrial quality can explain many long-puzzling features of germline evolution, including the widespread stability of somatic gametogenesis in plants and basal metazoans, the secondary association between germline sequestration and complexity, the conditions required for early germline sequestration, and other peculiarities of female germline development, including extreme oogamy, follicular atresia, and the putative germline bottleneck.
We consider an ancestral multicellular organism with isogamy that has evolved uniparental transmission of mitochondria exclusively from the “female” gamete (see Methods for full model details). Gametes are formed by meiosis late in development from the same stem cells that give rise to adult somatic cells, which we call somatic gametogenesis (Fig 1a). We then consider the selective conditions that favour the evolution of an early sequestered germline that has a reduced number of cellular replication cycles before gametes are produced (Fig 1b).
The quality of gametes and fitness of the organism depends on the number of mitochondrial mutations. These build up due to copying errors in mitochondrial genes at a rate μS per cell division (Fig 1c). Early sequestration of a germline restricts the number of cell divisions to gamete production and, therefore, the number of copying errors (Fig 1b). Mutations can also result from “background” damage, μB, per unit time (caused by oxidative damage or UV radiation; Fig 1d), as cells can accumulate mitochondrial mutations even when not actively dividing [32–35]. So, background mutations affect gametes whether they are derived from an early germline or later from stem cells in adult tissues.
Unlike nuclear genes, which are clonally transmitted through mitosis, the mitochondrial population doubles and segregates at random to daughter cells at each cell division. Some daughter cells receive more, others fewer, mitochondrial mutations (Fig 1e). With somatic gametogenesis (i.e., late differentiation of germ cells), segregational variation has the capacity to generate gametes that carry few mutants or are even mutation free. The more rounds of cell division, the greater the degree of segregation, the higher the variance in mutation load, and the larger the proportion of gametes carrying very few or no mitochondrial mutations (Fig 2a). The effect of segregation is dampened with larger numbers of mitochondria per cell, as this diminishes drift (Fig 2b). At higher copying-error rates (high μS), segregation still generates increased variance between daughter cells, but the proportion of gametes with few or no mutations falls to nearly zero (Fig 2c). For simplicity, we ignore selection at the level of mitochondria within cells, as we found that this roughly equates to changes in the mutation rate (as seen previously in [36,37]). For example, removing deleterious mitochondria by mitophagy effectively reduces the mutation rate, whereas faster replication of selfish or impaired mitochondria increases the mutation rate. Likewise, although the severity of mitochondrial mutations is variable, we fix each to have a similar effect for ease of analysis.
Overall, the resulting mitochondrial population in somatic cells is subject to a mutational load, which impacts on cell fitness
ω(m)=1−s(m/M)2,
(1)
where m is the number of mutant mitochondria and M is the number of mitochondria in each cell. This fitness function is concave and assumes that a large number of mutants must accumulate before cell function is significantly undermined (Fig 1f). This is the case for mitochondrial diseases, in which mutant load can reach 60%–80% before symptoms become detectable—the well-known threshold effect [38,39]. We concentrate on a fitness function with s = 1, which is the case for large deletions, a common form of mitochondrial mutation [38,39]. Note that even when s = 1, a large number of mutations need to accumulate before there is a significant effect on cell fitness, and all mitochondria in the cell must mutate for cell fitness falls to zero. Other fitness functions (e.g., linear or convex, S1 Fig) and s < 1 (i.e., weaker selection, S2 Fig) do not affect the general outcome of our modelling; we do not pursue these alternative fitness functions further here, as they do not correspond well with known effects of mitochondrial mutations [38,39].
Another important dimension of fitness in multicellular organisms is the number and quality of tissues and their mutual interdependence. We take the fitness of a somatic tissue to be the mean fitness of its constituent cells, each of which reflects the accumulation and segregation of mutations as described above. We assume epistasis between tissues, so that adult fitness is determined by the quality of the worst tissue, while the magnitude of interdependence is in effect determined by the number of tissues (see Methods). If by chance a disproportionate number of mitochondrial mutations segregate into precursors of one tissue, the fitness of the adult is severely impaired; conversely, if mutations are distributed equally into the precursor cells of all tissues, organismal fitness is improved, as each tissue has equivalent quality. We also analysed an independent tissue epistasis parameter, which can be varied from zero to strong epistasis. The results are equivalent to those generated by varying tissue number, so we do not present them here.
Adult fitness thus depends on the number of mitochondrial mutations inherited by the zygote, the number of new mutations arising during development, how these mutations segregate into cells and tissues, and the degree of tissue-level epistasis. As we will show, this combination determines whether it is better to sequester germ cells early in development in a germline or derive them from adult somatic stem cells late in development. In the model, the specific mutation rates, number of rounds of cell division, mitochondrial numbers, and lifespan were chosen to illustrate the general forces operating and to be computationally tractable (e.g., a reasonably short lifespan, small numbers of cell divisions to adult life or germline sequestration, and relatively high mutation rates per cell division). Parameter values were not chosen to reflect particular species, which vary over many orders of magnitude. Importantly, we find that relative and not absolute values of model parameters matter, and we discuss our findings in this light.
The main benefit of early sequestration of germ cells is to reduce the number of cell divisions before gamete production, so reducing the net input of copying errors (μS) in gametes. This raises the mean fitness of offspring in the next generation. But an early germline comes at the cost of reduced segregational variance. When gametes are derived later in development, there is more chance for segregation to generate gametes with lower numbers of mitochondrial mutations, which facilitates selection in the offspring, improving fitness over generations (Fig 2). The tension between these two forces determines the advantage (or disadvantage) of a germline.
When mitochondrial mutation input through copying errors is low, the benefit of increased segregational variance between gametes tends to outweigh the benefit of avoiding copying errors (μS) through arresting cell division in an early sequestered germline. At low μS, more rounds of cell division increase the proportion of gametes with a reduced mutation load despite the higher net input of copying error mutations (Fig 2a and 2b). Germline sequestration is therefore unlikely to evolve in organisms with low μS (Fig 3a). This outcome depends on the input of copying errors relative to the background mutation rate (μB). Background mutations occur whether oocytes are sequestered early in development or are derived late from somatic tissues. If background mutations dominate (Fig 3a, bottom right corner), a germline is unlikely to evolve. Only when the copying error rate (μS) increases markedly relative to the background mutation input (μB) does selection favour an early sequestered germline, with a marked reduction in the number of cell divisions leading to gamete production (Fig 3a).
Importantly, increasing the number of tissues (Fig 3b) or mitochondria (Fig 3c), which both correspond to greater complexity, does not enhance the likelihood of germline evolution. The reason again relates to segregation. In organisms with a single tissue, mean adult fitness closely follows the mutation load inherited by the zygote, with some variation induced by mutation accumulation (Fig 4a). In contrast, with multiple tissues, mutations can segregate differentially into distinct tissues, and with strong tissue epistasis (organism fitness depends on the mutational state of the worst tissue) will cause a substantial reduction in mean and increase in variance of adult fitness relative to the inherited zygote mutation load (Fig 4b). This creates heightened selective pressure for the late production of gametes as the extra rounds of segregation creates greater variance in mutation load between gametes and allows a stronger evolutionary response compared to competitors with early germline sequestration (Fig 3b). An early germline is less likely to evolve compared to an organism with a single tissue for the same mutation parameters.
Another marker of increased complexity is an increase in the number of mitochondria (M). This has a different effect. With higher M, segregation during development causes a more muted decrease in mean and increase in variance of adult fitness (Fig 4c). This is beneficial for adult fitness but has a knock-on effect of reducing fitness variation among gametes and thus weakens the evolutionary response. The net effect on selection for early germline sequestration is minimal (Fig 3c). Overall, then, greater complexity caused by multiple tissues and high numbers of mitochondria does not by itself contribute to the evolution of early germline sequestration, but, if anything, reinforces the stability of somatic gametogenesis.
Oogamy has typically been interpreted as an outcome of disruptive selection for gamete specialization, with large eggs for provisioning, contrasting with numerous small sperm for success in fertilization [40]. Oogamy is universal in extant metazoan groups and common to multicellular organisms in general, both with and without a dedicated germline, but is most extreme in complex organisms with an early sequestered germline [41]. Our analysis specifies the benefits that flow from the generation of large oocytes with high mitochondrial numbers and the relationship of this to the evolution of a germline.
Oogamy offers a temporary increase in mitochondrial number (M) that feeds through to the zygote. Imagine that oocytes undergo Q additional rounds of mitochondrial replication without cell division, producing mature eggs and, ultimately, zygotes with 2QM mitochondria. After fertilization, the first Q rounds of cell division partition the initial mitochondrial population in the zygote between daughter cells without the need for further replication, until the baseline M mitochondria is restored (Fig 5a). Note that a large oocyte does not alter the mean input of mitochondrial mutations per generation, because replication events lost from early development (when μS = 0) are simply moved later into gametogenesis. Nor does it impact the net input of background mutations (μB), which is simply a function of lifespan unchanged by oogamy.
The temporary increase in M profoundly dampens segregational variance in mutant frequency across the early cell divisions (Fig 5b, left-hand side), which greatly reduces the risk of one tissue inheriting a disproportionate number of mitochondrial mutants. This early suppression of variance is valuable in organisms with multiple tissues (Fig 5c). Once the early divisions of the zygote have allowed cells to regain the standard low level of M mitochondria, the many subsequent cell divisions before the end of development allow segregational variation to be established (Fig 5b, right-hand side). So, oogamy readily fixes in organisms with somatic gametogenesis (Fig 5d).
The spread of oogamy under somatic gametogenesis has a strong inhibitory effect on the evolution of a germline. As the level of oogamy (Q) increases, the fixation probability of an early germline declines, especially when there are multiple tissues (Fig 6a). Oogamy has the downside that reduced segregation early in development carries over into gametes that are sequestered early in development, and this weakens the response to selection. So an early sequestered germline is less likely to evolve in multi-tissue organisms in which oogamy is already established (e.g., Fig 6a, Q = 3).
Yet the observation is that females in active bilaterians invariably combine early germline sequestration with large oocytes, in the extreme packed with several orders of magnitude more mitochondria than are present in normal cells (106 versus ~103) [19]. The model seems at odds with these observations. A solution to this paradox is suggested by the finding that the fixation probability of an early germline rises with additional cell divisions of primordial female germ cells in organisms with multiple tissues (Fig 6b). The extra rounds of cell division in the germline help to restore segregational variance among female gametes, overcoming the suppression of variation caused by oogamy (Fig 6b). There is a balance between restricting μS through early germline sequestration, decreasing variance between tissue precursor cells through oogamy and increasing variance in female gametes through the proliferation of oogonia after germline sequestration. So increasing Q in effect favours higher numbers of cell divisions in an early sequestered germline. It is interesting to note that the need for segregational variance in the germline may produce more primordial germ cells than needed and so be accompanied by atresia, the random destruction of female gametes seen across a vast phylogenetic range of bilaterian metazoa [42, 43].
We have considered the evolution of the germline in terms of purifying selection against defective mitochondria. Mitochondrial quality is undermined by mutations, whereas selection is facilitated by increased variance generated through random segregational drift of mitochondria at each round of mitotic division (Fig 1). The more rounds of cell division, the greater the mutational burden (as the mitochondrial population is doubled at each cell division) but the greater the segregational variance, hence visibility to selection at the organism level (Fig 2). In terms of selection for mitochondrial quality, the switch to an early germline with a reduction in the number of germline cell divisions depends mainly on mutations due to copying errors (μS) relative to the background mutation rate (μB). The balance of these factors explains simply and beautifully why groups with low copying error rates (μS), such as plants [44–46] and basal metazoans [47–49], do not sequester a germline; segregational variance is maximised by forming gametes late from the same stem cell population that gives rise to somatic cells (Fig 2). This provides a better opportunity for selection on adult organisms to reduce the mutation load through multiple generations (Fig 3).
Conversely, in groups with higher rates of copying errors (μS), mutations accumulate faster during each round of mitotic division, making late differentiation of gametes a liability. This leads to benefit flowing from early sequestration of a dedicated female germ cell lineage with a reduced number of germline cell replication cycles, thereby limiting the accumulation of mutations even though this is at the cost of restricting variance. Examples are bilaterians and ctenophores; both groups have high nucleotide substitution rates in their mitochondria [50–54] and apparently evolved germlines independently [55]. Because mitochondria are not generally transmitted via the male germline, these constraints do not apply to sperm, which can be produced in abundance throughout life. So the simple balance of mutation and segregation explains why somatic gametogenesis is stable and widespread across plants and sessile metazoans, whereas more active metazoans evolved early germline sequestration.
The model presented here is an abstract simplification, and the values used for mutation rates, segregational variance, mitochondrial numbers, and the life cycle are not intended to be realistic (see Methods). Our objective was to examine the general principles involved without getting too bogged down in the detail. Parameter values were chosen for computational expediency. In particular, the mutation rates adopted are considerably higher than seen in most natural systems, and the lifecycle was chosen to be much more simple and condensed. This approach does not affect the general conclusions we draw (see S3 Fig and discussion below). There is considerable uncertainty over the actual range of values for most of the parameters we discuss. For example, mitochondrial mutation rates are normally estimated indirectly from nucleotide substitution rates, which tend to be 10–50-fold higher than the nuclear mean in mammals [50–54], and at least 10-fold lower than the nuclear mean in plants [44–46] and basal metazoans [47–49], a range of about three orders of magnitude. Given that purifying selection over generations eliminates a proportion of mitochondrial mutations, the underlying mutation rates are likely to be higher still. The number of mitochondria per cell and per gamete also vary over many orders of magnitude, from 102 [56–58] to 109 [59–61]. While doubts can be raised about estimates over such a broad phylogenetic range, it seems reasonable to assume that the broad features hold true—that the striking differences in rates of sequence divergence and mitochondrial numbers between plants and basal metazoans and bilaterians do reflect real differences in underlying mutation rates and metabolic demands. We believe these wide natural ranges justify our consideration of the consequence of increasing the error rate associated with mitochondria replication (μS) or increasing the number of mitochondria per cell and per oocyte.
We also simplify the life cycle so that early germline sequestration (and tissue segregation) takes place after just a few rounds of cell division, and somatic gametogenesis after 10 rounds of division. In fact, in humans, the oocyte precursor cells undergo about 30 mitotic divisions before entering meiosis [13,62], while the mitotically active cells in the gut [63] or blood [64] may undergo hundreds or even thousands of rounds of cell division. Likewise, somatic gametogenesis in mature sponges, corals, and large land plants may involve considerable rounds of mitotic division [65,66]. By modelling a much larger and perhaps more realistic life cycle with 60 cell divisions, we observe conditions favouring an early germline (set in this case at NG = 10) with a minimal mutation rate approximately two orders of magnitude lower than in Fig 3, without altering the relative contribution of each parameter (S3 Fig). We present these data to reinforce the point that our modelling is best interpreted in light of the relative strength of the parameters involved, not the absolute values. Relatively high mutation rates will tend to drive germline evolution; relatively low mutation rates produce higher quality gametes when coupled with the greater segregational variance generated by somatic gametogenesis.
An important relative difference is the balance between copying errors (μS) and background mutations (μB) (Fig 3). Historically, most mitochondrial mutations were assumed to be caused largely by ROS leak from active mitochondria [67–70], and to a lesser extent by other forms of oxidative stress such as UV radiation [71]. Mitochondrial DNA was claimed to be naked (lacking histones [72]), inadequately repaired [72], and vulnerable to ROS damage from the adjacent respiratory chains [67–70]. More recent work suggests that this one-sided view is far from correct [24–27,72–77]. Mitochondrial DNA is protected by DNA-binding proteins, notably mitochondrial transcription factor A [73]; most of the DNA repair mechanisms that operate in the nucleus have mitochondrial equivalents [74, 75], and the most common ROS, such as superoxide, cause little oxidative damage to mitochondrial DNA [76], acting instead as critical signals in regulatory pathways [77]. In contrast, work with polG mutants in mice suggests that the majority of mitochondrial mutations are copying errors [24–27]. More broadly, the strong bias towards transition mutations in the mitochondrial DNA of bilaterians is thought to be associated with copying errors rather than ROS damage and UV radiation [24,76,78]. But these observations do not rule out a significant contribution of background mutations, as implied by a modest proportion of transversion mutations in mammals [24,76], which tend to be associated with oxidative damage rather than copying errors [24,76,78].
We have accordingly taken both copying errors (μS) and background mutations (μB) to be significant contributors to the mitochondrial mutation rate, but with varying contributions depending on lifestyle and phylogeny. The model suggests that somatic gametogenesis is stabilized by relatively high μB (Fig 3). If a high proportion of transversion mutations are indeed associated with oxidative damage [24,76,78], then this fits well with the observations. Unlike most animal mitochondrial genomes, basal metazoans including sponges, corals, and placozoans all have very low μS [47–49] but a relatively high proportion of transversion mutations [48], equating roughly to a high μB in our model. This combination of low μS with high μB readily explains why these major metazoan phyla lack a germline. Similarly, most plants have low mutation rates (10–20-fold lower than in their nuclear genomes [44–46]) and typically display no bias towards transition mutations in mitochondrial DNA [46]. This again suggests a relatively high proportion of mutations caused by oxidative damage relative to copying errors, equating to relatively high μB. Whether these mutations are produced by ROS leak or by relatively high exposure to UV radiation (given phototropic environments, long lifespans and exposed gamete precursor cells not shielded in ovarian tissue) is not clear. Nonetheless, our predictions are clear: organisms with relatively low μS and high μB will tend to favour somatic gametogenesis (Fig 3), exactly as happens in plants and basal metazoans that lack an early sequestered germline. We are not aware of any other hypothesis that explains so cleanly why these large groups never developed germline sequestration, despite their elaborate tissue differentiation and morphological complexity.
Why did the copying-error rate (μS) increase in the lineages giving rise to bilaterians and ctenophores? The rise could have been linked with an ecological shift from sessile filter feeding in the Ediacaran to motility and predation in the Cambrian. More aerobic activity is linked with more protein synthesis and tissue turnover, so more replication of mitochondrial DNA between each cell division is needed to meet metabolic demands [79–81]. Increasing oxygen levels in the late Neoproterozoic [82] enabled multiple trophic levels and predation for the first time, as energy conservation from aerobic respiration approaches 40% compared with <10% from fermentation, making predation virtually impossible in anoxic worlds [83,84]. In the early Cambrian, predation is thought to have driven an evolutionary arms race, leading to greater size and more physical activity in bilaterians and ctenophores [85]. The evolution of a germline in turn frees mitochondria in somatic tissues from metabolic constraints, enabling greater power and tissue differentiation. A link between mitochondrial mutation rate and germline evolution is corroborated by the Ceriantharia, which, unlike other Anthozoa, have fast-evolving mitochondrial DNA [86], suggesting a relatively high μS; strikingly, their larvae have gonads and are apparently paedogenetic [87], implying that germline sequestration may have evolved in this group. Conversely, the secondary evolution of somatic gametogenesis in other groups such as Ectoprocta and Entoprocta [88] could have been favoured by their sessile lifestyles coupled to a falling μS. Little is known about mitochondrial sequence divergence in these groups, but we predict their μS will be lower than in related metazoans that retain a germline.
An unexpected finding of the model was that greater complexity, measured as the number of tissues, did not enhance selection for a germline. While segregation is helpful in generating variance between gametes, it potentially has a detrimental effect early in development, as it can result in the biased accumulation of mutants in the progenitor cells of particular tissues. Some tissue progenitors will contain more mutant mitochondria than others, and this will depress organismal fitness, given tissue epistasis (Fig 4). That is certainly true of bilaterians, in which fitness is dependent on the mutual functioning of all tissues; for example, brain function depends on heart and lung function for proper oxygen supply. Two effects are relevant here. First, as somatic gametogenesis generates greater variance between gametes and produces a stronger selective reduction in mutation load, it is more strongly favoured with multiple tissues (Fig 3b). This retards the evolution of an early germline. Second, multiple tissues favour the evolution of oogamy. Increasing the number of mitochondria in oocytes (high Q) means that early cell division occurs with a much-reduced effect of segregation, as drift is suppressed by large numbers (Fig 5b). This causes a reduction in mutational variation between cells in early development, thereby improving adult fitness when there are multiple tissues (Fig 5c). But this again retards the evolution of an early germline, because multiple cells divisions are required after the reestablishment of the normal level of M mitochondria per cell to generate a reasonable level of variance between gametes (Fig 6a).
Strikingly, some plants and basal metazoans have smaller numbers of mitochondria in oocytes (i.e., lower values of Q) than are typical of bilaterians [56–58], suggesting that “mitochondrial oogamy” is distinct from the large size of oocytes for provisioning. This lower degree of mitochondrial oogamy can be explained by the lower mutual dependence of adult fitness on organ function in plants and basal metazoans (i.e., less epistasis between tissue fitness). In the model, this is equivalent to a smaller number of tissues, and we find that oogamy is less strongly favoured with fewer tissues (Fig 5d compares four and eight tissues). Nonetheless, even in plants there is a degree of embryonic tissue determination in early development (e.g., in seeds [66]), so moderate mitochondrial oogamy is predicted.
The model reveals that high mutation rates (μS) favour early germline sequestration, whereas tissue differentiation early in development favours large oocytes packed with mitochondria, and this retards the evolution of an early germline. The problem is that, when large oocytes are combined with early germline sequestration, the variance between oocytes is minimized, hindering selection. We suggest that this may help explain follicular atresia, a long-puzzling feature of germline physiology. In human development, >6 million oogonia are generated by mitotic proliferation during foetal development, followed by the seemingly random apoptotic death of all but 500,000 of them (i.e., >90%) by the start of puberty [42]. A similar pattern is common across most bilaterian metazoans [43], including nematodes such as Caenorhabditis elegans, in which >50% of oocytes die [89]. Some kind of selective advantage for atresia has long been sought, largely with the idea of oocyte quality control [90]. But no plausible mechanism has been identified, and it is hard to imagine that >90% of oogonia have low fitness. Seen in terms of maximising variance, however, atresia makes much more sense. The proliferation of oogonia necessarily generates segregational variance between them, some of which now have the chance of carrying few or no mitochondrial mutations (Fig 2). Proliferation generates far more oogonia than are needed for fertilization. In the model, allowing additional rounds of segregation within the germline does indeed promote early germline fixation (Fig 6b). In real life, this “over supply” is rectified by random elimination of nine out of every ten oogonia, with the survivors being “islands” of segregational variance, more different to each other than would be the case if fewer divisions were used to generate an equal number of oocytes. Functional selection could now more realistically take place, for example, in the competition between follicles during oocyte maturation. This may partially account for the reported germline selection against severe mitochondrial mutations in mice [91–93].
Segregation is not the only way to generate variance between oocytes. The idea that has garnered the most empirical [94–96] and theoretical [36,37,97] attention is the so-called mitochondrial germline bottleneck. It is commonly asserted that mitochondrial populations are reduced to low numbers at a single stage of germline development, and some evidence supports this, albeit the low number attained is contested [98]. However, there is rather little evidence that this is anything other than the reduction from the high content of the oocyte back to a normal level. As we have shown, segregation per se has a similar effect, so the need for a bottleneck may have been overstated. Nonetheless, models suggest that, in principle, a very low M at some point during germline development would increase variance between oocytes, possibly limiting the need for proliferation of oogonia. Another possibility is that only a subset of oocyte mitochondria is replicated during oocyte maturation [99], which, again, would increase the variance between oocytes. It would not be surprising if different species had evolved different mechanisms for generating variance between oocytes, or that the need for variance differs amongst groups with divergent ecologies and lifestyles. The concentration on the details of a very small subset of model organisms detracts from understanding how evolution has acted in different ways to related problems.
The rationale we have put forward for the evolution of the germline is that it minimises the accumulation of mutations due to copying errors in mitochondrial replication (μS). The total accumulation of new mutations will also be lowered by adaptations that suppress background damage (μB) in female germ cells. A number of features appear to fulfil this role: protecting oocytes in an internal ovary, provisioning them from follicular cells, and repressing the transcription of mitochondrial genes and active respiration to reduce exposure to ROS, giving rise to quiescent oocytes [15,18–20]. The model shows that the early sequestration of large oocytes is more likely to be favoured when accompanied by specific lowering of μB (Fig 3 and S4 Fig). This is the case in the female germline of most [15,19] but not all [28,29] bilaterians. Even when mitochondria are active it is possible that other specific adaptations could suppress μB, for example, through upregulating antioxidant or DNA repair enzymes [28]. From our point of view, all these features are secondary adaptations that have arisen after the germline has become established. Likewise, once a germline is established, there is no longer a constraint on the replication error rate (μS) among somatic cell lineages. These are now free to become “disposable,” as their lineage will no longer contribute to the heritable germ material.
For the last 30 years, the dominant explanation for the evolution of early germline sequestration has been that it was essential for the emergence of multicellular organisms through the suppression of selfish conflict between the cells that make up an individual. At its heart, this viewpoint lacks a rationale for the evolutionary stability of numerous organisms that lack a germline.
Here we have turned this paradigm on its head, locating the key driving force in selection against mitochondrial mutations. This reflects the interplay of mutation and segregation of mitochondrial mutations in gametes and tissues. With low copying-error rates (μS), the combination of segregation and selection improves mitochondrial quality over generations, favouring somatic gametogenesis in plants and sessile metazoans. Higher μS drives early germline sequestration, which ultimately permits greater tissue differentiation and complexity in groups such as bilaterians and ctenophores. The increase in μS probably related to the rise in oxygen shortly before the Cambrian explosion [82]. We find that the evolution of complex development, with stronger epistasis between tissues, requires suppression of mitochondrial variance between tissue-progenitor cells. This is achieved by large oocytes containing high mitochondrial numbers. That, in turn, risks the loss of segregational variance between oocytes sequestered in the germline early in development. The solution was to restore segregational variance through the mitotic proliferation of oogonia in the germline, followed by atresia, maximizing the variance between the surviving oocytes.
These findings set out a completely novel hypothesis for the origin and evolution of germline sequestration, which predicts the traits of groups from plants and sponges to ctenophores and bilaterians. Unlike other hypotheses, selection for mitochondrial quality can account for the stability of both somatic gametogenesis and early germline sequestration. The requirement for segregational variation in the germline also elucidates the potential risks associated with mitochondrial heteroplasmy, which reduces variance between oocytes while increasing the risk of segregating mitochondrial mutants into particular tissues, lowering adult fitness and contributing to mitochondrial disease. The unusual population genetics of mitochondrial mutation and segregation can uniquely explain the evolution of the germline in complex bilaterians.
Here we review the mitochondrial dynamics at the cell level, showing that random segregation at every cell division generates variation in the mutational load (Fig 2). We start with a cell containing M mitochondria, out of which m0 are mutant. The initial state of an infinite population is then represented by a state vector p(0), with all of its M+1 entries set to zero, except for the m0-th, which is set to one, i.e., pi(0)=δ(i,m0). A single cell cycle consists of mitochondrial mutation, replication, and cell division. Mitochondrial mutation is a Bernoulli trial with success probability μ, while segregation is modelled as a simple random sampling without replacement. After n cell divisions, the population state vector becomes
p(n)=(KJ)np(0).
Here, J is the (M + 1) × (M + 1) matrix with elements representing the transition probabilities due to mutation,
Jq,m=pmut(q;m,M)=(M−mq−m)μq−m(1−μ)M−q.
Similarly, K is the matrix of same dimensions representing the transformation due to random sampling,
Kq,m=pseg(q;m,M)=(2mk)(2M−2mM−k)(2MM).
In Fig 2, we show distributions p(n) for n = 0…10.
Variance in the mutant load after n cell divisions can be expressed analytically in the illustrative case of μ = 0, where only segregational drift is accounted for. Let Xn be a random variable denoting the number of mutant mitochondria within a cell in the developing tissue or embryo after n cell divisions, and xn its actual realization. For sampling from the hypergeometric probability distribution, the population mean equals the initial number of mutants, E(Xn) = x0, while the variance can be decomposed as
Var(Xn)=E[Var(Xn|Xn−1)]+Var[E(Xn|Xn−1)].
For sampling without replacement from the hypergeometric probability distribution, the conditional variance is
Var(Xn|xn−1)=xn−1(M−xn−1)2M−1,
and so
Var(Xn)=E(Xn−1(M−Xn−1)2M−1)+Var(Xn−1)=E(MXn−12M−1)−E(Xn−122M−1)+Var(Xn−1)=M2M−1E(Xn−1)−12M−1E(Xn−12)+Var(Xn−1)=x0M2M−1−12M−1[Var(Xn−1)+x02]+Var(Xn−1)=x0(M−x0)2M−1+(1−12M−1)Var(Xn−1).
Here, x0 is the initial number of mutants within a cell. This is a recurrence relation of the form hn = Hhn−1 with hn = Var(Xn) − x0(M − x0) and H=1−12M−1. With the boundary condition Var(X0) = 0, h0 = −x0(M − x0), the solution is
Var(Xn)=x0(M−x0)[1−(1−12M−1)n].
Variance in the mutant frequency Pn=XnM is then
Var(Pn)=p0(1−p0)[ 1−(1−12M−1)n ].
In oogamous matings, the zygote contains 2Q M mitochondria and undergoes Q divisions without mitochondrial replication. Changes in variance follow a similar trend, but this time the recurrence relation is
Var(Pn)=p0(1−p0)21+Q−nM−1+Var(Pn−1)(1−121+Q−nM−1),1≤n≤Q.
The recurrence relation was used directly to produce Fig 5b. For a large initial number of mitochondria, 12QM→0 and the approximate solution is
Var(Pn)=p0(1−p0)2QM(2n−1), 1≤n≤Q.
To study the evolutionary dynamics of alleles responsible for germline sequestration, we developed a multilevel, agent-based model, implemented as a set of simulation routines in ANSI/ISO C++ (https://github.com/ArunasRadzvilavicius/GermlineEvolution). The model population consists of N = 500 multicellular individuals, with equal numbers of males and females. A single individual is represented as an object containing 2L cells, each containing a number of mutant and wild-type mitochondria (that is, development consists of L cell divisions). The cells are grouped into X tissues, each of which is composed of 2L/X cells. Tissues are differentiated after T cell divisions by assigning each precursor cell into its own tissue, so that X = 2T. We measure time in terms of cell divisions, with a total life span before gamete production and release S = 40. Generations are discrete and non-overlapping. Organism development proceeds by iterating through the population and modifying these objects according to a set of predefined rules. The population is initialized in a random mutational state, and the simulation proceeds as follows:
Mating type, the number of germline cell divisions G, the log-size of the zygote Q, and the degree of uniparental inheritance v are all traits of an organism controlled by a set of loci in the nuclear genome, which is assumed to be diploid. The mating type locus is heterogametic ZW in females, while the other mating types is homogametic ZZ.
We determine the selective advantage of an invading allele in a Monte Carlo simulation, by numerically calculating its fixation probability within a finite population. With the population at equilibrium, the invading allele is introduced at a low frequency f0 = 0.05 and its fate is tracked until either fixation or extinction. This requires 103 − 104 repetitions of the calculation, depending on the mutation rates and the associated levels of noise. This pattern is assessed relative to the fixation probability of a neutral allele, which simply equals its initial frequency f0. The error bars in Figs 5 and 6 represent the 95% confidence interval of the binomial proportion via the Gaussian approximation, that is Δp=±1.96p(1−p)n, where n is the number of trials and p is the measured rate of fixation. An allele is deemed to be evolutionarily advantageous if its fixation probability exceeds the chance of fixation of the neutral allele. The model parameters, such as mutation rate and the strength of epistatic interactions, are not intended to correspond to values measured in specific species, but to show general patterns and to maintain reasonable calculation times (unavoidable in modelling studies).
The same general procedure is applied to determine the fate of modifiers coding for the number of germline cell divisions, the level of mitochondrial oogamy, or mitochondrial exclusion. The germ cell differentiation locus is expressed in both sexes and is assumed to be autosomal, with the invading allele assumed to be dominant. We also examined other dominance states, but the main conclusions of this work remained unaffected. The other two loci are W linked and unaffected by dominance considerations.
|
10.1371/journal.pgen.1007886 | Narya, a RING finger domain-containing protein, is required for meiotic DNA double-strand break formation and crossover maturation in Drosophila melanogaster | Meiotic recombination, which is necessary to ensure that homologous chromosomes segregate properly, begins with the induction of meiotic DNA double-strand breaks (DSBs) and ends with the repair of a subset of those breaks into crossovers. Here we investigate the roles of two paralogous genes, CG12200 and CG31053, which we have named Narya and Nenya, respectively, due to their relationship with a structurally similar protein named Vilya. We find that narya recently evolved from nenya by a gene duplication event, and we show that these two RING finger domain-containing proteins are functionally redundant with respect to a critical role in DSB formation. Narya colocalizes with Vilya foci, which are known to define recombination nodules, or sites of crossover formation. A separation-of-function allele of narya retains the capacity for DSB formation but cannot mature those DSBs into crossovers. We further provide data on the physical interaction of Narya, Nenya and Vilya, as assayed by the yeast two-hybrid system. Together these data support the view that all three RING finger domain-containing proteins function in the formation of meiotic DNA DSBs and in the process of crossing over.
| Errors in chromosome segregation during meiosis are the leading cause of miscarriages and can result in genetic abnormalities like Down syndrome or Turner syndrome. For chromosomes to segregate faithfully, they must recombine with their homolog during the early steps of meiosis. An essential component of the process of meiotic recombination is creating the lesions (double-strand breaks, DSBs) that are required to form a crossover with the homologous chromosome. Crossovers are required to ensure chromosomes segregate properly at the first meiotic division. In this study we have identified two genes, narya and nenya, that are essential in DSB formation. We found that narya arose from a duplication of nenya, and these two genes are functionally redundant. In addition to its role in DSB formation, narya also plays a role in processing DSBs into crossovers. Strengthening our knowledge about the mechanism by which Narya both creates DSBs and processes them into crossovers will lead to a better understanding of the process of meiotic chromosome segregation not only in flies but many other organisms, as these genes have homologs in yeast, worms, plants, mice and humans.
| Homologous recombination is an essential feature of meiosis and is required to ensure proper chromosome segregation. Although several core aspects of meiosis are highly conserved, many of the proteins and structures that mediate meiosis have features that are unique to each model organism. This is most apparent when comparing the process of meiotic recombination in Drosophila to other model organisms.
Meiotic recombination begins with the induction of programmed DNA double-strand breaks (DSBs). In Drosophila (as well as Caenorhabditis elegans) this event occurs in the context of full-length synaptonemal complex (SC) [1,2,3]. Therefore, in flies, synapsis is not dependent on DSB formation, as it is in other model organisms like budding yeast, plants and mammals [4,5,6,7,8]. DSBs are catalyzed by the evolutionarily conserved protein Spo11 (MEI-W68 in Drosophila [9]), the homolog of subunit A of TopoVI DNA topoisomerase [10,11]. Although nine other DSB accessory proteins (Mre11, Rad50, Xrs2, Ski8, Rec102, Rec104, Rec114, Mei4 and Mer2) have been identified in budding yeast (reviewed in [12,13]), only three proteins have been demonstrated to be required for DSB formation in Drosophila (MEI-P22, Trem, and Vilya) [14,15,16]. MEI-P22 has sequence homology to the transducer domain found within the B subunit of TopoVI DNA topoisomerase [17], and therefore may interact directly with MEI-W68 as a complex. Trem is a C2H2 zinc finger domain protein with no known homologs in other model systems [15]. Vilya, a RING finger domain-containing protein, has homology to Zip3-like family members found in several organisms [16]. However, none of the members in other systems appear to affect the formation of DSBs themselves [18,19,20,21,22,23,24,25].
Once DSBs are made, they must be repaired into either crossovers or noncrossovers. This is a multistep process utilizing enzymes and proteins that stabilize crossover intermediates and further promote crossover maturation. Early-acting pro-crossover proteins in most organisms (yeast, plants, nematodes and mammals) include the heterodimer of Msh4 and Msh5 (reviewed in [26]). The Msh4/5 complex is required for stabilizing crossover intermediates and promoting repair through the crossover pathway. Drosophila lacks this complex and instead is thought to use the MEI-MCM complex (REC, MEI-217 and MEI-218) for this function [27,28,29]. In addition to the lack of conservation in early pro-crossover proteins, Drosophila also seems to lack the homologs of late pro-crossover proteins that are required for crossover maturation [29]. Instead of the endonuclease MutLγ (Mlh1 and Mlh3) that is used to resolve crossovers in most organisms, Drosophila appears to use an endonuclease complex consisting of MEI-9, Ercc1, Mus312 and Hdm (reviewed in [29,30]).
Although many of the yeast proteins necessary to create DSBs and determine their fate as crossovers or noncrossovers are not conserved in flies, we recently identified a protein named Vilya, which is required for DSB formation and localizes to the recombination nodule (RN), a protein structure assembled only at sites of crossing over [16]. Vilya appears to be homologous to the Zip3-like protein family that is involved in crossover fate by stabilizing crossover intermediates and aiding crossover maturation. In fact, Vilya appears to link DSB formation and crossover formation in Drosophila. Zip3-like proteins fall into two subgroups: the Zip3-RNF212 group and the HEI10 group, with all members of both groups sharing conserved structural properties (reviewed in [31]). Most of these Zip3-like proteins appear to have dynamic localization patterns that involve either a redistribution of the protein from the SC to sites of recombination intermediates and/or an increase in their concentration at these sites as recombination intermediates are processed into crossovers. Studies in multiple organisms argue Zip3-like proteins act as post-translational regulators at sites of crossing over either through sumoylation or ubiquitination or both [20,22,32,33,34,35,36,37]. Recently, a study in C. elegans identified three paralogs of a previously known member of this group, ZHP-3, which were shown to function in two separate heterodimeric complexes [25]. These complexes are thought to form a signaling network that mediates crossover assurance and crossover interference by functioning both to stabilize crossover intermediates (ZHP3/4) [25,38] and to promote crossover maturation (ZHP1/2) [25], similar to the roles found in mammalian RNF212 and HEI10, respectively.
Here we report on the identification of two paralogs, narya (CG12200) and nenya (CG31053), that encode proteins that are both structurally and functionally related to Vilya and have homology to the Zip3-like family. In D. melanogaster, narya likely arose from a gene duplication of nenya less than 40 million years ago, and the two show genetic redundancy and are required for meiotic DSB formation. Using the CRISPR-Cas9 system to tag the endogenous copy of narya, we find that Narya localizes to DSBs and colocalizes with Vilya throughout pachytene. As we previously showed Vilya to be a component of the RN, this would suggest that Narya (and likely Nenya as well) are also RN components. In addition, as is true for Vilya, the localization of Narya to discrete foci within the SC is dependent on DSB formation, and in the absence of DSBs, Narya localizes uniformly along the SC. Finally, we report the identification of a separation-of-function allele of narya (naryaG4) that links Narya directly to crossover maturation. Therefore, Narya, and most likely Nenya, appear to be the second and third examples after Vilya of proteins linking DSB formation with DSB fate, and likely Narya is the second protein to make up the RN in Drosophila.
Because many organisms have multiple Zip3-like proteins that play a role in meiosis, we conducted a genome-wide search for Zip3-related genes in Drosophila melanogaster. We identified two genes (CG12200 and CG31053) that appeared to encode good Zip3-like candidates. CG12200 (FBgn0031018) is located on the X chromosome at map position 18C7 in the last (6th) intron of CG32533. CG32533 is a gene with unknown function that is predicted to be a helicase. CG31053 (FBgn0051053) is located on the 3rd chromosome at map position 98B6 in the first intron of minotaur (CG5508), a conserved member of the glycerol-3-phosphate O-acyltransferase (GPAT) family. Both CG12200 and CG31053 are predicted to encode proteins that have similar structural properties to Zip3-like family members (including Vilya in Drosophila [16]), such as an N-terminal C3HC4 RING finger domain and an internal predicted coiled-coil domain (S1A Fig). Therefore, we named these genes narya (CG12200) and nenya (CG31053) to complete the Three Rings of Power given by the Elves of Eregion [39].
We used the protein sequences of Narya and Nenya to identify homologous proteins in other model organisms to determine if we could identify either Zip3-like family members or proteins outside of this family that had known roles specifically in meiosis or meiotic recombination. In addition to showing protein homology to Zip3 in budding yeast, Narya, Nenya and Vilya showed homology to all four Zip3-like family members in C. elegans (ZHP-1, ZHP-2, ZHP-3 and ZHP-4) and to RNF212 and RNF212B in several mammalian species. (RNF212B is a protein known to affect the recombination rate in both cattle and sheep [40,41].) All three of the D. melanogaster RING proteins (Narya, Nenya and Vilya) cluster with the Zip3-RNF212 subgroup (S1B Fig) [16].
We then investigated the conservation of vilya, narya, and nenya in the 12 fully sequenced genomes from the Drosophila Genomes Consortium. Using a tBLASTn search, we identified the most likely homolog in each of the 12 Drosophila genomes and determined if the gene locations maintained synteny. While we found evidence of vilya and nenya homologs across the Drosophila genus, we could not identify homologs of narya outside the melanogaster subgroup (Fig 1A). Maximum-likelihood phylogenetic analyses suggest that narya arose as a gene duplication event of nenya less than 40 Mya, prior to the separation of the melanogaster subgroup (Fig 1B). Within D. melanogaster, narya and nenya nucleotide sequences are 69.1% identical to each other, while Narya and Nenya protein sequences share only 49.1% identity and 66% similarity (Fig 1C). However, despite their high level of divergence, narya and nenya are evolving at a similar rate (S1 Table).
Given that narya and nenya are homologous to many of the Zip3 family members, we assessed whether these two genes had roles during meiosis. We had previously created several mutations in narya using TALEN-based mutagenesis where we specifically targeted the RING finger domain [42]. Since RING finger domains are known to mediate protein-protein interactions and are required for mediating E3 ligase activity, we speculated that mutations in this domain would abolish narya function. One such mutation resulted from an indel (insertion of 3 nucleotides/deletion of 13 nucleotides during repair) causing a frameshift at amino acid 42 that eventually truncates the protein to 115 amino acids. This truncated allele, known as naryaJJ6, also lacks the last two conserved cysteines in the RING finger domain and therefore is likely nonfunctional (see Fig 1C).
Using FLP/FRT-mediated recombination with two piggyBac transposons that each flanked the nenya gene [43], we created a chromosomal deletion of nenya (nenyadel). Because nenya is located within the intron of minotaur, a gene known to be required for silencing the piRNA pathway in oocytes [44], we also created an RNAi construct specific for nenya to assay its function in the absence of potential effects created by disrupting the minotaur gene. We used the GAL4-UAS system under the control of the nanos (nos) promoter (Pnos-GAL4::VP16) to induce expression of the nenya RNAi hairpin transgene (hereafter referred to as nenyaRNAi). The nosGAL4::VP16-UAS system results in high levels of expression in the germline throughout most stages of oogenesis, including the germarium where meiosis begins [45,46,47]. qPCR analysis indicated that the nenya transcript levels were reduced by at least 50% in whole ovaries when pValium22-nenyaRNAi was driven within the germline (S2 Fig).
We tested each of these mutant alleles, individually and in combination with each other, for effects on meiotic chromosome segregation (Table 1). There was at most a weak effect on meiotic nondisjunction compared to controls for the homozygous mutants when tested individually. naryaJJ6 showed low (2.2%), but statistically significant, levels of X chromosome nondisjunction when compared to the control (0.0%), while the nenya mutant (nenyaRNAi) repeatedly showed wild-type chromosome segregation (0.3% X chromosome nondisjunction). In addition, there was no significant effect on meiotic segregation when there was only one copy of wild-type narya in the complete absence of nenya (0.5% X chromosome nondisjunction), suggesting that narya is not haploinsufficient as has been reported for members of this group in other species [22,48,49]. In contrast, double mutants (either naryaJJ6; nenyadel or naryaJJ6; nenyaRNAi) showed high levels of X chromosome nondisjunction (49.0% and 32.4%, respectively), indicating that these genes have redundant functions. Supporting this proposal, we were able to rescue the nondisjunction phenotype in the naryaJJ6; nenyadel double mutant with expression of a narya:gfp transgene in the germline (0.0% X chromosome nondisjunction) (Table 1).
In vilya mutants, the increase in meiotic nondisjunction is a result of failed initiation of the meiotic recombination process. To determine if the meiotic nondisjunction we observe in narya nenya double mutants occurs through a similar mechanism, we assayed the presence of DSBs formed in the pro-oocytes during pachytene (Fig 2). To do this, we used a phospho-specific antibody against the histone 2A variant (γH2AV). Phosphorylation of H2AV is an evolutionarily conserved rapid response that occurs at DSB sites [51,52,53]. We found that in the absence of only nenya, DSBs are formed at wild-type levels in early pachytene pro-oocytes (mean 10.2 DSBs, SD ± 0.90 compared to 10.8 DSBs in the same meiotic stage in a wild-type background [54]), consistent with the normal levels of chromosome disjunction (Fig 2A and 2B and S3 Fig). However, in the naryaJJ6; nenyaRNAi double mutant, meiotic recombination failed to initiate in early pachytene cysts, with few, if any, DSBs detected (mean 1.1 DSBs, SD ± 0.78) (Fig 2A and 2B and S3 Fig). Similar results were obtained when analyzing γH2AV foci number in the naryaJJ6; nenyadel double mutant (average 0.74 DSBs, SD ± 0.97 in 27 pro-oocytes), indicating that the level of RNAi knockdown for nenya transcript (less than 50% of wild-type nenya transcript levels within the whole ovary) was sufficient to mimic the genomic nenya deletion with regard to DSB formation function (see Materials and Methods).
We also failed to detect crossovers when assaying crossover formation using genetic markers along the X chromosome (Fig 2C). The failure to detect meiotically induced DSBs using the γH2AV antibody is not due to a general defect in modifying the histones at the DSB sites, as we can detect γH2AV foci during the endoreduplication cycle (S3 Fig). In addition, these effects on DSB formation are unlikely to be caused by defects in synaptonemal complex formation or in the selection of the oocyte by early-mid pachytene, as these processes appeared to be normal in the absence of narya and nenya (S3 Fig).
In the narya TALEN-based mutagenesis described above, we also created a second allele (naryaG4) that deletes five amino acids, including the last cysteine in the RING finger domain, one amino acid prior to it, and the three amino acids that follow it (see Fig 1C). The reading frame in naryaG4 is maintained after the deletion, thus this mutant likely expresses a form of the protein that is missing key residues to form the RING finger domain. We assayed whether naryaG4, which lacks part of the RING finger domain, was able to facilitate DSB formation in the absence of nenya. We found that DSBs were formed in the naryaG4; nenyaRNAi double mutant (mean 7.6 DSBs, SD ± 2.83), unlike in the naryaJJ6; nenyaRNAi double mutant, indicating that DSB formation is not fully dependent on an intact RING finger domain of Narya (Fig 2B).
In the naryaG4; nenyaRNAi double mutant, DSBs were induced at ~70% of the level observed for nenyaRNAi alone (Fig 2B), which led us to reason that we would see a decrease in the amount of nondisjunction (see Table 1) compared to the naryaJJ6; nenyaRNAi double mutant that failed to form DSBs. Therefore, we assayed for both the level of nondisjunction and the presence of crossing over on the X chromosome in naryaG4; nenyaRNAi double mutant females and compared that to the nenyaRNAi mutant and the double naryaJJ6; nenyaRNAi mutant (Fig 2C). As expected, due to the severe reduction in DSBs in naryaJJ6; nenyaRNAi females, we failed to recover any recombinant X chromosomes in their progeny (map distance of 0.0 cM, E0 frequency of 1.0). These females also showed high levels of X nondisjunction (32.4%) compared to nenyaRNAi alone, which makes wild-type levels of DSBs and disjoins homologous chromosomes properly (map distance 39.9 cM, E0 frequency of 0.346, 0.3% X nondisjunction).
We found that while the naryaG4; nenyaRNAi mutant was able to form DSBs (see Fig 2B), those DSBs were not converted into crossovers (map distance of 0.2 cM, E0 frequency of 0.996), and females maintained high levels of X chromosome nondisjunction (39.5%) seen in the naryaJJ6; nenyaRNAi mutant (Fig 2C). Although the frequency of X chromosome nondisjunction in the naryaG4; nenyaRNAi females was greater than what was observed in the DSB-deficient naryaJJ6; nenyaRNAi females, this difference is statistically not significant with the number of progeny scored, and both are consistent with published data for mutants that fail to form crossovers due to the absence of either DSBs or homologous chromosome synapsis (S2 Table) [55].
The failure to form crossovers was not due to a global defect in DSB repair, as we did not detect any delay in removal of the γH2AV mark at mid-pachytene (S4A Fig). We also failed to detect any defect in karyosome structure, such as a fragmented karyosome, that is typical of DNA repair mutants (S4B Fig) [56]. In addition, the fertility did not decrease from that of the naryaJJ6; nenyaRNAi double mutant (each double mutant combination yielded ~19 progeny per female in the recombination assay). These data suggest that the naryaG4 allele is a separation-of-function mutant that maintains the ability to form DSBs, albeit at reduced numbers, but causes a deficiency in the ability to repair those DSBs into crossovers. This also predicts a direct function of Narya in the formation of crossovers, in addition to its separable role in DSB formation.
Since the presence of either narya or nenya is required for DSB formation, and Narya is functionally redundant with Nenya, we next asked whether Narya localized to sites of DSBs. We analyzed the localization of Narya during pachytene by creating a green fluorescent protein (GFP)-tagged version of narya at the genomic locus using CRISPR/Cas9 technology. We tested the naryaGFPcrispr alone and in combination with both nenya alleles to determine if the naryaGFPcrispr allele was completely functional. Females that were homozygous for naryaGFPcrispr in the absence of nenya (either nenyadel or nenyaRNAi) showed little to no meiotic chromosome segregation errors (S3 Table), indicating that naryaGFPcrispr is fully functional.
Immunofluorescence studies on whole ovaries showed that naryaGFPcrispr is highly expressed at the same stage in which DSBs are induced, as both a haze (detected in undeconvolved images) and faint staining along the SC with predominant numerous foci that decrease in number as the cysts progress through pachytene (S5 Fig and see below). Further analysis indicated that NaryaGFP foci colocalized with γH2AV foci, the histone modification created at the DSB site (Fig 3). These results are similar to our observation that Vilya also localizes to DSBs [16]. However, although Vilya, when overexpressed, colocalizes to ~60% of the γH2AV foci, NaryaGFP colocalized with γH2AV foci 93% of the time when expressed at the endogenous level (S6 Fig). In the 10 nuclei analyzed in early pachytene, the average number of DSBs was 13.1 and the average number of NaryaGFP foci was 10.6. In addition, since NaryaGFP is expressed from its endogenous promotor, we could determine that NaryaGFP also localized to the DSBs that are induced in the nurse cells within the 16-cell interconnected cyst. The number of NaryaGFP foci in the oocyte nuclei decreased as the cyst moved from early pachytene stage into early-mid pachytene (Region 2B) (see Fig 4), where the average number of NaryaGFP foci was 5.4 in the 12 nuclei analyzed. The number of Narya foci is similar to the number of VilyaHA foci (4.8 foci) previously found at this stage [16], both of which are consistent with the number of crossovers formed per female meiosis.
Since NaryaGFP associated with DSB sites and the number of NaryaGFP foci decreased as pachytene progressed, we reasoned that these NaryaGFP foci might colocalize with Vilya foci. Therefore, we analyzed the localization of NaryaGFP in ovaries expressing vilyaHA in the germline using the nos-GAL4/UAS system (Fig 4). We found that VilyaHA and NaryaGFP colocalized in SC-positive cells and remained colocalized as both types of foci decreased in number from early pachytene to early-mid pachytene (Region 2A to Region 2B). Examination of single-gallery z-slices throughout an early pachytene nucleus shows the faint localization of NaryaGFP to the SC and the association of NaryaGFP foci with VilyaHA foci (S7 Fig). The maintenance of colocalization in early-mid pachytene (Region 2B, see Fig 4), a stage where VilyaHA localizes to RNs by immuno-EM [16], demonstrates that Narya is a component of the RN.
Previous studies using high-resolution imaging followed by straightening of each of the chromosome arms have shown that at early-mid pachytene, the localization of VilyaHA foci are consistent with both the number and position of crossovers, with each stretch of euchromatic SC between homologous chromosome arms primarily containing one VilyaHA focus [16]. Taken together these results suggest that Vilya and Narya localize to the majority of the DSBs in early pachytene, and as the cyst progresses to early-mid pachytene, both proteins are maintained and concentrated at DSB sites destined to become crossovers. In addition, as we saw in earlier studies with Vilya, at late pachytene (Stage 5) when γH2AV foci are no longer present, there is a change in the localization of NaryaGFP from the discrete foci found at early pachytene to threads of staining exclusively along the SC, where it colocalizes with VilyaHA (Fig 4, see Discussion).
Based on the number and localization of the NaryaGFP foci at sites of DSBs and the fact that the number of these foci decreased as DSBs were repaired into crossovers, we asked what effect DSB formation (Fig 5A) and/or lack of DSB repair (Fig 5B) had on the localization and number of NaryaGFP foci. As is also true for Vilya, NaryaGFP fails to form discrete foci and instead localizes along the SC when DSBs are absent (either in the absence of mei-W68 or in the absence of vilya). However, in the absence of DSB repair, as in an okra (DmRAD54) mutant, NaryaGFP foci form, and the foci number in early-mid pachytene is similar to when DSB repair is normal. These results indicate that Narya displays two types of staining patterns depending on the presence or absence of DSBs. First, in the presence of DSBs, Narya forms discrete foci at DSB sites. Moreover, if there is a failure to repair those DSBs, there is not an increase in number of Narya foci at early-mid pachytene. We interpret this data to mean there is not an increase in the number of designated crossover sites in the absence of DSB repair. Second, in the absence of DSBs, either because the DSBs are undergoing normal repair or fail to form, Narya displays thread-like SC staining.
Since Narya and Vilya colocalize at sites of DSBs and crossovers, and Narya and Nenya are functionally redundant, we wanted to determine if Nenya can physically associate with Narya and/or Vilya. Due to the lack of a functional nenya epitope-tagged transgene or antibodies to any of the RING finger domain proteins, we used the yeast two-hybrid system to help us understand the associations and/or interactions between these three proteins.
We cloned narya, nenya and vilya into yeast two-hybrid vectors and tested their ability to interact with each other in all pairwise combinations. In addition, we tested for the ability of each protein to interact with itself (Fig 6). We found that Narya, Nenya and Vilya interact with each other (Fig 6A) as well as with themselves (Fig 6B).
Previous studies showed that both the RING finger domain of Vilya as well as its C-terminal region are required for its interaction with the DSB accessory protein MEI-P22 [16], so we further investigated the interaction of Vilya with Narya and Nenya by testing whether C-terminal and RING finger domain mutants of Vilya could still bind to Narya and Nenya. Neither Vilya’s RING finger domain nor it’s C-terminal region were required for its interaction with either Narya or Nenya (S8 Fig), indicating that Vilya likely interacts with Narya and Nenya through the middle region of the Vilya protein, perhaps assisted by the coiled-coil domain. Additionally, although Vilya interacts with MEI-P22 as well as Narya and Nenya, neither Narya nor Nenya were able to interact with MEI-P22 by yeast two-hybrid (S9 Fig).
Finally, since Narya interacted with both Nenya and Vilya, we then tested the ability of NaryaG4 to interact with each of these proteins (Fig 6C). We found that Nenya-NaryaG4 and Vilya-NaryaG4 binding were substantially reduced compared to binding with wild-type Narya protein. We also found that NaryaG4 was unable to interact with itself. This inability of NaryaG4 to strongly interact with Nenya, Vilya or itself is not due to the lack of expression of NaryaG4 (S10 Fig). When considered with the results above showing that Narya and Vilya colocalize, these yeast two-hybrid data indicate that all three proteins likely function as part of the RN.
The finding of three Zip3 family members in Drosophila, Narya, Nenya and Vilya, is consistent with studies in other organisms that show that the presence of multiple Zip3 homologs within an organism is not uncommon [25,48,57]. These proteins share common structural features such as a RING finger domain near the N-terminus, and in those organisms that form SC, a predicted coiled-coil domain within the middle third of the protein (reviewed in [31] and [23]). The presence of a RING finger domain suggests that these proteins play roles in either the ubiquitination or sumoylation pathway as E3 ligases [58]. Indeed, members of this family have been shown to be required for sumoylation (e.g., Zip3 [33,37], RNF212 [59]) as well as ubiquitination (e.g., HEI10, mammals [36,59]) or are speculated to be a sumoylation/ubiquitination switch (e.g., HEI10, Sordaria [20]) necessary to stabilize and/or promote crossover formation. However, the mechanism(s) by which the Drosophila homologs act is currently unknown.
Studies in a number of organisms have shown that Zip3-like proteins function as pro-crossover factors during meiosis and localize along the SC as linear arrays of foci and/or as discrete foci at crossover sites (reviewed in [31]). We provide evidence that at least two of the RING finger domain-containing proteins in Drosophila, Narya (this study) and Vilya [16], also localize in this manner. Using an overexpression construct, we previously showed that Vilya localizes along the central region of the SC and at sites of DSBs. Eventually Vilya becomes concentrated at crossover sites, as immuno-EM studies demonstrated that Vilya localizes at RNs. In this study, we analyze the localization pattern of Narya using a CRISPR/Cas9-engineered epitope-tagged version of narya at the genomic locus, which eliminates many of the caveats of using an overexpression system. Although very faint Narya SC staining could be seen when analyzing naryaGFPcrispr, the predominant staining was discrete foci that localized to the majority of DSBs early in pachytene, and those foci decreased in number as pachytene (and DSB repair) progressed (see Fig 4). Vilya has also been shown to localize to a subset of DSBs during early pachytene [16]. The discrete Narya foci observed in both early and early-mid pachytene colocalized with Vilya, indicating that these two proteins are found together within the SC at DSBs as they form and are repaired. These findings indicate that Narya is also found at crossover sites and is a component of the RN. The identification of two Drosophila Zip3-like proteins at sites of maturing crossovers is consistent with studies of all other homologs in that they also localize at or associate with proteins known to be at crossover sites [19,20,22,25,32,37,38,48,60].
The similarities in localization of both Narya and Vilya to other Zip3 family members predict these proteins may play a role in crossover control. However, our previous studies and those described here indicate that, in a fashion that is so far unique to Drosophila, Narya, Nenya and Vilya first function prior to DSB fate determination; which is to say that they are essential for meiotic DSB formation. Our data demonstrate that narya and nenya encode functionally redundant proteins that are necessary for the induction of meiotic DSBs during early pachytene. Only in the absence of both gene products is there an increase in meiotic nondisjunction resulting from the lack of recombination due to the failed induction of DSBs. This severe reduction of DSB induction is not seen in mutants that affect the formation of the SC. Mutants that fail to form SC (c(3)G) or that form fragmented SC (c(2)M) do not eliminate DSBs but reduce their numbers in the pro-oocytes [61]. Therefore, we propose that Narya and Nenya play a direct role in the formation of DSBs. In addition, the absence of vilya, or other genes required for DSB formation (e.g., mei-W68 or mei-P22), results in identical meiotic phenotypes [9,14,16,61]. However, because we are basing the lack of DSBs on the absence of γH2AV signal, we cannot rule out the possibility that narya and nenya, and possibly vilya, instead allow the very rapid repair of DNA lesions thereby reducing the number and/or amount of γH2AV signals, as has recently been found for RNF212 in female mouse oocytes [62].
Previous studies demonstrated that Vilya interacts with MEI-P22, the potential partner of DmSpo11, which is known to be required for DSB formation. In addition to the colocalization of Narya and Vilya throughout pachytene, yeast two-hybrid studies show that Narya, Nenya and Vilya all interact with each other. The direct interaction of Narya or Nenya with Vilya does not appear to require a functional N-terminal RING finger domain of Vilya, which was necessary for its interaction with MEI-P22, or its C-terminal region that is known to be required for DSB formation. This may indicate that it is the middle third of Vilya that is necessary for its interaction with Narya and Nenya. As the middle region of Vilya contains the predicted coiled-coil domain, a domain that can mediate protein-protein interactions, it is highly possible that these proteins interact through their coiled-coil domains. However, the observation that the RING finger domain mutant, NaryaG4, failed to interact with itself, Nenya and Vilya in the yeast two-hybrid assay, may indicate that the coiled-coil domains are not sufficient for interaction and that the RING finger domain may also be required for protein-protein interactions. We should note that the mutations in the RING finger domain of Vilya used in this analysis differed from the mutation in Narya. The Vilya mutations were single amino acid substitutions, whereas the mutation in Narya resulted in a five amino acid deletion. It is possible that the deletion in Narya alters the protein structure, thus disrupting the ability of the coiled-coil domain to interact with other proteins.
Many proteins that localize to the SC contain coiled-coil domains, and our studies here show that while Narya primarily localizes to discrete foci, SC localization is observed at low levels in a naryaGFPcrispr background in early pachytene and Narya is exclusively found along the SC in late pachytene (Fig 4). The SC localization at early pachytene could be due to the propensity of coiled-coil proteins to localize to the SC, or this localization may be required for wild-type levels of DSBs, as most meiotic mutants that fail to assemble SC only induce DSBs at a reduced level [61]. In addition, we show that in the absence of DSB formation, discrete Narya foci fail to form and instead Narya localizes along the SC in a similar staining pattern to that of the SC protein C(3)G. The Narya SC localization occurs in the absence of either mei-W68 or vilya, indicating that although Narya and Vilya colocalize and may interact directly, Narya’s localization to the SC is not dependent on Vilya. The exclusive localization of Narya to the SC during late pachytene in the presence of wild-type DSB repair was similar to the distribution of Vilya in the same genetic backgrounds. In this study, however, we were able to assess the localization of Narya at endogenous levels throughout pachytene, and therefore we are confident that there is a change in the localization pattern from foci in early-mid pachytene to linear staining along the SC in late pachytene. Currently, we do not understand the function of this redistribution. It is possible that Narya, and perhaps Vilya, play a role in the disassembly of SC that occurs post DSB repair.
Based on the relationship of Narya to other Zip3 homologs and its localization and association with Vilya, which is found at RNs, it seems likely that Narya might have a role in processing DSBs into crossovers. However, the fact that narya and nenya are also required for DSB formation makes it difficult to analyze the role of either in crossover formation. An analogous problem arose when studying mutations that affected the function of vilya [16]. In that case, we reasoned that if Vilya could be recruited to exogenous DSBs from its localization along the SC when DSBs were absent, it would provide strong evidence that Vilya had a role in crossover formation. Using X-rays to produce exogenous DSBs, that is precisely what we found. In the absence of mei-W68 (Dm Spo11), but following X-irradiation, Vilya, which in this background is found exclusively along the SC, formed discrete foci at a subset of exogenous DSBs. Here we provide direct evidence that Narya plays an essential role in the formation of crossovers. We obtained an in-frame deletion within the RING finger domain of narya (naryaG4) and analyzed its role in DSB formation and crossing over in the absence of nenya. Unlike the null allele of narya (naryaJJ6), naryaG4 retained its ability to function in DSB formation (Figs 2 and S3). There was a slight decrease in the mean DSB number, and a wider range of DSBs in the nuclei assayed, but a significant number of DSBs (average of 70%) were formed. Surprisingly though, none of the DSBs that were formed were able to be converted into crossovers. The DSBs in naryaG4; nenyaRNAi oocytes were most likely repaired as noncrossovers, given that we did not see either a karyosome fragmentation defect associated with the lack of DSB repair or a more severe fertility defect from the naryaJJ6; nenyaRNAi females. The presence of DSB repair combined with the lack of crossovers resulted in high levels of nondisjunctional progeny. In summary, our data demonstrate that in Drosophila, members of the Zip3 family are required to both form DSBs and repair those DSBs into crossovers, and flies use a mechanism to ensure these processes are directly linked. Future studies will need to be done to determine the precise function of Narya and whether it acts to stabilize crossover intermediates and/or in the maturation of crossovers.
Based on sequence comparison, narya appears to have duplicated from nenya less than 40 million years ago, after the split of D. ananassae from the melanogaster subgroup. Both genes have been maintained in all the sequenced species of the melanogaster subgroup. We provide evidence that narya and nenya encode proteins that are functionally redundant with regard to their role in the early steps of meiosis. The preservation of both genes and their functional redundancy is surprising since genetic redundancy in Drosophila is not prevalent [63,64]. In fact, studies have shown that the vast majority of meiotic genes are not duplicated [65]. In addition to the duplication of nenya found in the melanogaster subgroup, we also found evidence of a gene duplication of vilya in D. ananassae. However, unlike the vilya homolog in D. ananassae that maintains synteny, the duplicated gene is intronless, likely caused by a retrotransposition event. Retrotransposed duplicates do not bring upstream and downstream regulatory regions with them and are often pseudogenized, as they are less likely to be expressed or maintained [66].
It is not obvious why the melanogaster subgroup has maintained two meiotic genes with the same function. As we presently lack any alleles that allow for visualization of Nenya, we can only speculate that Nenya is behaving exactly as Narya. However, we cannot rule out that the functional redundancy of these two genes is due to their roles in DSB formation, and that Narya may be more important than Nenya at the RN in the formation of crossovers. While we failed to detect any meiotic chromosome nondisjunction in the absence of nenya, we consistently observed low levels of chromosome segregation errors in the absence of narya (X chromosome nondisjunction levels ranging from 2–4%, see Table 1). We know based on their sequence alignment that the C-terminal region shows the least conservation. Perhaps future studies will determine whether this domain could be important for independent functions of the two proteins. Our studies here have shown that Narya’s RING finger domain is critical for crossing over but not for its role in DSB formation; it will be interesting to dissect these same domains in Nenya.
Taken together, these studies identify two functionally redundant genes, narya and nenya, that are required for the induction of meiotic DSBs. Both of these genes encode proteins that are structurally and functionally similar to the Drosophila protein Vilya, and all three show similarities to a family of proteins found in many organisms that are required to process meiotic crossover events. We show here that in addition to its role in DSB formation, Narya is required for crossover formation. While Drosophila may lack a subset of both DSB accessory and pro-crossover homologs present in the majority of model systems, flies have clearly found a way to utilize the proteins they do have for both processes.
Drosophila strains were maintained on standard food at 24°C. Descriptions of genetic markers and chromosomes can be found at http://www.flybase.org/. Stocks used in this study include Pnos-GAL4::VP16 [45], PUASp-vilya3XHA [16], vilya826 [16], mei-W684572 [67], naryaJJ6 [42], Pnos-GAL4::VP16 naryaJJ6 (this study), naryaG4 [42], Pnos-GAL4::VP16 naryaG4 (this study), nenyadel (this study), okraAA cn bw/CyO and okraRU cn bw/CyO [56]. vilya refers to the genotype vilya826, mei-W68 refers to the genotype mei-W684572, and okra refers to the genotype okraRU/okraAA.
The rescue construct (below) and all alleles of narya generated in this manuscript were made using the Canton-S stock or the Canton-S narya sequence. The Canton-S narya sequence differs from the narya sequence on FlyBase at 10 bases. Nine of these base changes encode for the same amino acid. One of the base changes result in an amino acid change from alanine at position 69 in FlyBase to glutamic acid in the Canton-S stock. naryaJJ6 and naryaG4 were generated using TALEN mutagenesis as described in [42]. naryaJJ6 deletes 16 bases, adds 3 and makes a nonsense allele after the lysine, and naryaG4 removes 15 bases, causing the deletion of 5 amino acids (deletion CGQVL), but maintains the frame of the gene. nenyadel was generated by FLP/FRT recombination with two piggyBacs (PBac{WH}CG5508[f01088] and PBac{WH}CG5508[f04927]) that both reside in the intron of CG5508, which also contains CG31053 (nenya).
Coding sequences obtained from FlyBase for D. melanogaster vilya, nenya and narya were used as BLAST queries in order to retrieve homologous sequences for additional Drosophila species. The tBLASTn option was used with the expect threshold set to 0.05. Retrieved genes were then examined for shared synteny with D. melanogaster. For narya and nenya in particular, this was done by determining whether they were found within the introns of the homologs of D. melanogaster minotaur or CG32533, respectively. Originally, the narya homolog in D. sechellia could not be definitively determined due to the poor coverage in the area, although partial narya sequence could be found in the first intron of the CG32533 homolog. With the recent release of a new D. sechellia genome [68], full sequence of a syntenic narya homolog was identified.
Nucleotide sequences for identified homologs were aligned using the PRANK+F algorithm [69]. Maximum-likelihood trees were inferred using IQ-TREE [70], with the best-fit model selected by ModelFinder [71]. To infer the relative evolutionary rates of narya and nenya, Tajima’s relative rate tests [72] were performed using MEGA7 [73] on the PRANK-aligned nucleotide sequences.
A naryaGFP knock-in was generated using CRISPR/Cas9 technology. Using the CRISPR Optimal Target Finder (http://tools.flycrispr.molbio.wisc.edu/targetFinder/), two genomic regions were selected for making the gRNAs [CCTTCCACTTGACCCAGTGCCGG and AGATCTTCTCCGCGTTGACTGGG (the PAM sequences are underlined)] and were cloned into the pU6-BbsI-chiRNA vector (gift from Melissa Harrison, Kate O’Connor-Giles and Jill Wildonger; Addgene plasmid #45946) [74] by the protocol outlined at http://flycrispr.molbio.wisc.edu/protocols/gRNA using oligos (IDT) 5’-CTTCGCCTTCCACTTGACCCAGTGC-3’ and the complement 5’-AAACGCACTGGGTCAAGTGGAAGGC-3’ and 5’-CTTCGAGATCTTCTCCGCGTTGACT-3’ and its 5’-AAACAGTCAACGCGGAGAAGATCTC-3’, respectively. Plasmid DNA was isolated using a Qiagen Midi Prep Kit.
The homologous recombination repair template containing the narya gene with a 3’ GFP epitope tag with 1,000 bases of genomic sequence both up- and downstream of the narya gene was generated in the pBS-KS+ vector (Clontech) by the following method. Using the Canton-S stock as the genomic DNA source (gift from Dana Carroll), we first cloned in the region 5’ to the narya gene and the majority of the narya gene using primers 5’-[Phos]GTGGCGCATCGTTGTCAGTC-3’ and internal gene primer 5’-[Phos]CAGAAGGCATATCCGACGGC-3’ using the EcoRV site in pBS that was previously digested and dephosphorylated. The insertion of this fragment was sequenced for directionality so that the 3’ end of the narya gene was positioned closest to the XbaI site in the pBS vector. The pBS vector containing this piece of the genome was digested with StuI (which cuts only within the narya gene) and XbaI (which cuts within the pBS vector). A StuI/XbaI fragment that contained the end of the narya gene at the internal StuI site through a 3’ in-frame GFP tag was amplified from pUASP-attB-naryaGFP (below) using primers 5’-GTATGCGGCCGGATGTTTCGAGTGCA-3’ and 5’-GCGCTCTAGATTACTTGTACAGCTC-3’ and then digesting with StuI and XbaI and used to clone into the vector. The 1,000 bases of genomic region 3’ to the narya gene was then cloned into this vector using primers 5’-GCCGTCTAGATCACTCCAATTACTTG-3’ and 5’-GTACTCTAGACTGCGATCCTCGACAG-3’ and cloned into the XbaI site in the above vector. The insertion of this fragment was sequenced for directionality.
Following the creation of the homologous repair template, which consisted of 1,000 bases upstream of narya, the narya gene with cloned GFP tagged at the 3’ end of the gene and 1,000 bases downstream of narya, the two PAM sequences in the narya gene were mutated using the Quik Change II XL Site-Directed Mutagenesis Kit (Agilent Technology). The base changed in the PAM sequence is in bold above. In both cases, the codon remains unchanged.
250 ng of each gRNA plasmid and 500 ng of the homologous repair template plasmid were injected (BestGene) into y w; nosCas9 (on II at attP40) (gift to BestGene from Shu Kondo). Potential CRISPR/Cas9 hits were screened with primers 5’-GTTGCAGCAGCTGGAGCAGA-3’ and 5’-GGTGAGTGCTCCCCAGATTG-3’, which amplify a region spanning the GFP insertion on the homologous repair template, allowing for PCR fragment size to visualize a repair off the homologous template. Once a CRISPR/Cas9 insertion was identified, the entire homologous region used in repair was sequenced. In this case, only one G0 fly had the correct insertion and was used for further analysis.
pUASp-attB [75] naryaGFP was made by cloning the CDS of CG12200 minus stop codon with primers 5’-GTATGCGGCCGCATGTTTCGAGTGCATTGCA-3’ and 5’-GTATGCGGCCGCCAAGACGAAAGCCTTTAGTG-3’ into a NotI digested pUASp-attB vector that previously had cloned in a venus (GFP) tag at NotI and XbaI. The CDS was sequenced for directionality.
An RNAi hairpin for nenya was identified using http://www.flyrnai.org/cgi-bin/RNAi_find_primers.pl. The sequence identified (GGACATAGATTGCCTTGAAGA) (underlined below) had no predicted off-targets and only shares five bases with narya. The hairpin was cloned using the oligos (IDT) 5’-CTAGCAGTGGACATAGATTGCCTTGAAGATAGTTATATTCAAGCATATCTTCAAGGCAATCTATGTCCGCG-3’ and 5’-AATTCGCGGACATAGATTGCCTTGAAGATATGCTTGAATATAACTATCTTCAAGGCAATCTATGTCCACTG-3’ into the pValium22 vector (gift from Jian-Quan Ni and Norbert Perrimon), https://fgr.hms.harvard.edu/trip-plasmid-vector-sets.
qPCR determined that the level of nenya knockdown, when expressed in the female germline using the nos-GAL4::VP16 driver, was greater than 50% of nos-GAL4/ +; naryaRNAi/ + or Canton-S (wild-type) nenya transcript levels. While the nenya transcript levels are higher than what might be expected given the phenotype, this observation may be explained by the process in which the cDNA was synthesized. Since random hexamer primers were used to amplify cDNA from total RNA, we cannot rule out that the remaining levels of nenya transcript in the presence of RNAi knockdown are not from amplified, unspliced RNA from minotaur in which nenya resides. It is also possible that the remaining levels of nenya transcript are from expression of nenya in the somatic cells of the ovary, since the knockdown was specific to the germline. As well, based on published data of nanos RNA and protein, there are varying levels of expression in egg chambers of different stages within the ovariole [46].
The narya RNAi hairpin (GCAAGATCTCCAAGTTCCAAG), which had no predicted off-targets and differed from nenya sequence at three bases, was used as a non-specific RNAi control. Two qPCR nenya primer pairs were used to determine the relative level of transcript present in nos-GAL4/+; nenyaRNAi/+ ovaries compared to Canton-S control ovaries. Total RNA from ovaries was isolated using the Promega Maxwell RSC Simply RNA Tissue Kit using standard protocol except for increasing the amount of DNase to 10 μL per sample. cDNA was synthesized from total RNA using the Invitrogen SuperScript III First-Strand Synthesis System for RT-PCR using random hexamers. Using the CAS qPCR Setup Robot to prepare the plates, each genotype was run in triplicate using Quanta Biosciences PerfeCTa SYBR Green FastMix ROX reagent. The nenya primer set was 5’-ACGTCGAGCCAACGTTGATC-3’ and 5’-TCGATCGGAATCGCTCGCAG-3’, and the control transcript primer set used was 5’-TGGACAGGTCATCACCATCGGAAA-3’ and 5’-TTGTAGGTGGTCTCGTGAATGCCA-3’ for ACT42A (FBgn0000043).
The frequencies of meiotic nondisjunction and meiotic recombination on the X chromosome were measured by crossing single virgin females of the listed genotypes to y sc cv v f·car / BsY males. This cross allows for the recognition of nondisjunctional offspring from the mother as Bs females (diplo-X exceptions) and B+ males (nullo-X exceptions). Normal segregation results in B+ females and Bs males. Nondisjunction frequency is calculated as the sum of exceptional progeny X 2 (to correct for the inviability of triplo-X and nullo-X exceptional progeny) divided by the sum of all progeny classes (viable plus inviable; denoted as adjusted total progeny scored). For X recombination analysis, only the female progeny (denoted as n) were analyzed for the intervals sc-cv and cv-f. y and v markers were unable to be scored due to the presence of y+ and v+ in the PUASp-nenyaRNAi transgene inserted at attP40.
Yeast transformation, mating and two-hybrid assays were done according to The Matchmaker Gold Yeast Two-Hybrid System User Manual (Clontech). AH109 yeast were used in place of Y2Hgold. The AH109 genotype is as follows: MATa, trp1-901, leu2-3, 112, ura3-52, his3-200, gal4Δ, gal80Δ,LYS2 : : GAL1UAS-GAL1TATA-HIS3, GAL2UAS-GAL2TATA-ADE2, URA3 : : MEL1UAS-MEL1 TATA-lacZ. Y187 genotype is as follows: MATα, ura3-52, his3-200, ade2-101, trp1-901, leu2-3, 112, gal4Δ, met–, gal80Δ, URA3 : : GAL1UAS-GAL1TATA-lacZ. cDNAs were cloned into either the pGADT7 or the pGBKT7 prey and bait vectors using restriction sites within the vector and contained within the PCR primers. The CDS for narya and nenya were obtained from Canton-S, as these genes do not contain introns.
Western blot analysis from yeast haploid cells was performed as described in [16].
Germarium preparation for whole-mount immunofluorescence was performed as described in [16]. Primary antibodies used included affinity-purified rabbit anti-Corolla (animal 210) (1:2000) [76], mouse anti-C(3)G 1A8-1G2 (1:500) [77], anti-Cona (animal 20) (1:500) [78], high-affinity rat anti-HA (clone 3F10, Roche) (1:100), rabbit anti-histone H2AvD pS137 (1:500) (Rockland Inc.), mouse anti-γH2AV (1:1000) (Iowa Hybridoma Bank) [54], monoclonal mouse anti-GFP (1:500) (clone 3E6, Thermo Fisher Scientific) and rabbit anti-GFP (1:500) (AB6556, AbCam Inc.). Secondary goat anti-mouse, rabbit or rat Alexa-488, Alexa-555 and Alexa-647 IgG H&L chain conjugated antibodies were all used at 1:500 (Molecular Probes, Life Technologies, NY).
Images were acquired using a DeltaVision system (GE Healthcare) supplied with a 1x70 inverted microscope with a high-resolution CCD camera. Images were deconvolved using SoftWoRx v. 6.1 or 7.0.0 (Applied Precision/GE Healthcare) software. Image analysis was performed using either SoftWoRx v. 6.1 or Imaris software 8.3.1 (Bitplane, Zurich, Switzerland). Brightness and contrast were adjusted minimally to visualize signals during figure preparation.
|
10.1371/journal.pcbi.1003756 | Blind Predictions of DNA and RNA Tweezers Experiments with Force and Torque | Single-molecule tweezers measurements of double-stranded nucleic acids (dsDNA and dsRNA) provide unprecedented opportunities to dissect how these fundamental molecules respond to forces and torques analogous to those applied by topoisomerases, viral capsids, and other biological partners. However, tweezers data are still most commonly interpreted post facto in the framework of simple analytical models. Testing falsifiable predictions of state-of-the-art nucleic acid models would be more illuminating but has not been performed. Here we describe a blind challenge in which numerical predictions of nucleic acid mechanical properties were compared to experimental data obtained recently for dsRNA under applied force and torque. The predictions were enabled by the HelixMC package, first presented in this paper. HelixMC advances crystallography-derived base-pair level models (BPLMs) to simulate kilobase-length dsDNAs and dsRNAs under external forces and torques, including their global linking numbers. These calculations recovered the experimental bending persistence length of dsRNA within the error of the simulations and accurately predicted that dsRNA's “spring-like” conformation would give a two-fold decrease of stretch modulus relative to dsDNA. Further blind predictions of helix torsional properties, however, exposed inaccuracies in current BPLM theory, including three-fold discrepancies in torsional persistence length at the high force limit and the incorrect sign of dsRNA link-extension (twist-stretch) coupling. Beyond these experiments, HelixMC predicted that ‘nucleosome-excluding’ poly(A)/poly(T) is at least two-fold stiffer than random-sequence dsDNA in bending, stretching, and torsional behaviors; Z-DNA to be at least three-fold stiffer than random-sequence dsDNA, with a near-zero link-extension coupling; and non-negligible effects from base pair step correlations. We propose that experimentally testing these predictions should be powerful next steps for understanding the flexibility of dsDNA and dsRNA in sequence contexts and under mechanical stresses relevant to their biology.
| DNA and RNA are fundamental molecules in the central dogma of molecular biology. Many biological behaviors of double-stranded DNA and RNA – including transcription/translation by proteins and packaging into compact structures – depend on their ability to flex and twist. Single-molecule tweezers now provide accurate mechanical measurements of DNA and RNA helices under force and torque but have not been used to rigorously falsify and thereby advance computational models. Here we present the first such blind challenge, involving recent dsRNA tweezers data that were kept hidden from modelers and a new HelixMC toolkit that resolves challenges in simulating long double helices from base-pair level models. The predictions gave excellent agreement with bending and stretching measurements of dsRNA but failed to recover twisting properties, pinpointing a critical area of future investigation. HelixMC also predicted that poly(A)/poly(T) and Z-DNA–biologically important variants whose elastic responses have not been studied with tweezers–will have distinct mechanical properties. These results open a route to iteratively falsifying and refining computational models of long nucleic acid helices, as is necessary for attaining a predictive understanding of their biological behaviors.
| Nucleic acids play central roles in biological processes including transcription, translation, catalysis and regulation of gene expression [1], [2]. Double-stranded RNA and DNA (dsRNA and dsDNA) stretch and twist when interacting with proteins [3], [4] and when forming compact structures such as nucleosomes [5] and packaged viruses [6], [7]. Understanding such deformations is critical for a fundamental understanding of nucleic acids in their biological contexts and for efforts to rationally engineer nanostructures built from dsRNA and dsDNA helices. High precision experimental data are becoming increasingly available from measurements using optical and magnetic tweezers [8]–[20] that measure end-to-end lengths and linking numbers of kilobase-length single molecules upon variation of solution condition, sequence, applied force and torque. In principle, these data offer rigorous challenges that can falsify or validate – and thereby advance – models of nucleic acid flexibility. However, such direct comparison of model predictions and experimental observables remains incomplete.
On one hand, fits to analytical equations based on worm-like chain (WLC) or elastic rod models are in common use for interpreting single-molecule manipulation data [14], [21]–[24], but they lack the power of predicting new experimental results and involve numerous approximations (see below). On the other hand, high-resolution approaches that integrate all-atom energy functions and crystallographic knowledge [25]–[32] offer the prospect of predictive calculations, but the computational costs to simulate kilobase-scale helices remain prohibitively large. Coarse-grained models, such as the base-pair level models (BPLMs) pioneered by Olson and colleagues [33] as well as models that use reduced representations for each base (rather than base-pair) [34]–[36], provide mesoscopic “bridges” between simple analytical models and atomic-level simulations. In this work, we focus on BPLMs as they have fewer degrees of freedom than single-base level models, enabling efficient calculations, and their parameterization can be more easily refined by the growing data of crystallographic structures [33], [37]–[47]. It is worth noting that BPLM is only expected to be applicable to duplexes at low-to-medium tension. Structural transitions involving breaking of base-pairs or formation of non-canonical base-pair interactions, typical at very high tension, are better modeled with single-base level models [48]–[52].
Despite continuing advances, BPLM simulation methods have not yet been used to make direct comparisons with single-molecule experiments. BPLM simulations have focused on helices up to hundreds of base-pairs, significantly smaller than the kilobase lengths probed in single-molecule experiments at which helix bending and twisting may play significant roles in the measured properties. In addition, BPLM calculations have been primarily developed for B-DNA duplexes; growing crystallographic knowledge for dsRNA helices has not been integrated into the BPLM framework. Finally, accurate methods for computing and constraining the twist, writhe, and link of discrete, open-ended helices have not been established until recently [53]–[56] and have not been integrated into BPLM modeling.
Here, we describe a blind prediction challenge, where developers of modeling algorithms (FCC, RD) predicted unreleased data on the mechanical properties of dsDNA and dsRNA helices measured by a team of experimenters (Lipfert et al., unpublished data). More specifically, the torsional properties and stretch modulus of dsRNA have not been previously reported (only the bending persistence length of dsRNA was measured previously [13]; the stretch modulus of dsRNA was published during the modeling [20]). This challenge motivated the development of a software package HelixMC, first presented in this work, to close the methodological gaps described above and thus enable simulations of force vs. extension, effective torsional persistence vs. force, link vs. force, and extension vs. link experiments. The goal of calculating actual experimental observables necessitated several systematic studies to check widespread but poorly tested modeling assumptions, including simulation-based validations of the Moroz-Nelson formula for torsional persistence length [21], [22]. Most importantly, the rigorous comparison between blind predictions and data revealed how current BPLMs largely succeed in modeling stretching and bending but apparently miss physics necessary for understanding dsDNA and dsRNA torsional properties. Finally, HelixMC predictions for previously unmeasured properties of two biological important variants, poly (A)/poly (T) dsDNA and Z-DNA, delineate future experiments that will allow incisive evaluation and revision of current modeling approaches.
Before presenting the results of the blind prediction, we present an overview of the simulation system and algorithm. Detailed descriptions are given in the Methods section. BPLMs [33], [37]–[46] abstract the entire duplex into multiple base-pairs stacking on top of each other. The coordinate transformation between two neighbor base-pairs (i.e. a base-pair step) is conventionally described with six standard step parameters (shift, slide, rise, tilt, roll, and twist). The internal interactions between neighbor base-pairs can therefore be described using the distribution of these parameters drawn from the Protein Data Bank (PDB) in six-dimensional (6D) space. Typically, these 6D distributions are approximated with 6D multivariate Gaussians to allow continuous sampling of the conformation space. We also tested an alternative scheme which samples directly from existing parameters in the database, without assuming Gaussianity.
The duplexes, represented in BPLM, are then simulated with a Metropolis Monte Carlo (MC) method, with stretching forces and torsional constraints incorporated into the energy function. By default we simulated dsDNA/dsRNA of 3,000 base-pairs at room temperature (298K). At the end of each cycle of Monte Carlo updates, the helix extension and the linking number are recorded. For direct comparison to single molecular tweezers analysis, these data from simulations at different forces and torsional constraints are then used to compute global mechanical properties including bending persistence length, stretch modulus, torsional persistence length and link-extension coupling, by fitting to analytical equations based on the elastic rod model.
Single-molecule tweezers experiments allow accurate measurements of the extension and the linking number of long molecules under externally applied stretching forces and torques. Typical experiments include force vs. extension, effective torsional persistence vs. force, link vs. force, and extension vs. link measurements. The published literature on dsDNA mechanical measurements is extensive (see e.g. [10], [11], [18], [19]), but magnetic tweezers data directly probing the torsional properties of dsRNA had not been published at the time of this study (only the bending of dsRNA has been previously studied [13]). Instead, a comprehensive experimental portrait (Lipfert et al., unpublished data) had been acquired by one of us with colleagues but was not publicly released. This situation therefore permitted blind prediction tests of the BPLM approach. Our modeling challenges were to simulate the different experimental setups, to test the applicability of phenomenological formulae used for curve-fitting, and to make quantitative predictions with estimated errors for the following standard constants: bending persistence length A, stretch modulus S, torsional persistence length C, and link-extension coupling g.
Drawing on extensive prior work [33], [54], [56], we were able to simulate dsDNA (for validation of the algorithm) and dsRNA (for blind prediction) under applied force using HelixMC. Fig. 1 gives example simulation frames with random sequences, with BPLMs parameterized on crystallographic data with diffraction resolutions better than 2.8 Å and without proteins. (Other BPLM variants are described below.) For both dsDNA and dsRNA, higher stretching force leads to longer end-to-end extensions and smaller fluctuations orthogonal to the stretching direction, qualitatively consistent with theoretical predictions and experimental observations.
Measurements of the mean end-to-end extension as a function of force give quantitative data for how nucleic acid helices bend, and we first tested if HelixMC recovered the bending persistence length seen in experiments for dsDNA. The simulated data fit well to standard models used in interpreting tweezers experiments, including the extensible worm-like chain (WLC) model proposed by Bouchiat et al. [23] (Fig. 2A; A = 54.7±0.6 nm), the inextensible WLC model [23] (A = 53±1 nm), and an alternative extensible WLC fitting model developed by Odijk [57] (A = 55±1.0 nm); see Table S1 and Fig. S1. The agreement of all three fits to each other and to more direct estimates of A by averaging the base-pair step transforming matrix [33] (A = 53.0±1.0 nm) confirmed the robustness of A as a comparison metric between experimental and simulated data. To bracket systematic error, we further performed simulations using BPLMs with a high-resolution subset of crystallographic data (2.0 Å vs. 2.8 Å diffraction resolution cutoff), without using a Gaussian approximation for the BPLM distributions, and symmetrizing the base-pair step parameters; these variations gave less than 10% changes in A (Table 1 and S2). We did however find that inclusion of protein/DNA crystallographic structures, which include more distorted helical conformations, led to reduction of A by 30% to 39 nm. Given this level of systematic error, the agreement of the HelixMC calculation and the experimental value for dsDNA (A in the range of 44–49 nm at near-physiological salt concentrations [16], [20], [58], [59]) was reasonable.
The agreement for dsDNA suggested that the prediction of the dsRNA bending would be similarly accurate. The HelixMC prediction for dsRNA was 66 nm, greater than the value for dsDNA, with a systematic error of ∼30%, again based on an alternative BPLM parameterization including protein/RNA crystallographic models (Table 2 and Fig. 2B). Experimental dsRNA tweezers measurements gave values of A = 57±2 nm (Lipfert et al., unpublished data) and 59±3 nm [20], greater than the value for dsDNA and in quantitative agreement with the HelixMC value.
In addition to enabling fits of the bending persistence length A, force/extension curves give estimates of the stretch modulus S, particularly at high force where the helix is pulled straight without bends. For dsDNA simulations with several variations, the HelixMC calculations gave estimates of S = 2000 pN. As with the bending behavior, inclusion of protein/DNA structures produced lower stretch modulus values, corresponding to more flexibility (S = 1500 pN; Table 2). These calculations overestimated the experimentally measured value for dsDNA of S in the range of 900–1400 pN [20], [60], [61], slightly beyond our estimated error.
The HelixMC prediction for the stretch modulus of dsRNA was S = 980 pN, with a systematic error of 25%. This estimate was also supported by using an alternative model to fit the simulation stretch modulus (Table S3 and Fig. S1). Given the dsDNA results above, we expected this HelixMC value to overshoot the experimental measurement. Nevertheless, beyond this error in absolute values, we strongly expected that dsRNA would give a relative stretch modulus significantly lower than dsDNA. Unlike the nearly straight axis curve of dsDNA, the base-pair centers of dsRNA trace a ‘spring-like’ axis curve, twirling in circles of radius 8 Å. We developed a novel “springiness” hypothesis, that this “spring-like” property of dsRNA would render it more pliable to stretching, analogous to a spring's lower stretch modulus compared to a straight wire (Fig. 3). Indeed, the experimental measurements for the dsRNA stretch modulus was 350±100 pN (Lipfert et al., unpublished data), more than two-fold less than for dsDNA, in agreement with our prediction. An independent experimental dsRNA measurement released at the time of modeling gave a similar value lower than dsDNA (500–683 pN) [20]. Additional simulation-based tests of the ‘springiness’ hypothesis are described in Supplementary Results and Table S4, S5.
The development of magnetic tweezers with increasingly sophisticated geometries has enabled torsion-sensitive measurements of dsDNA [16], [62]–[64] and, most recently, measurements on dsRNA that were included in our blind challenge. Before describing the blind comparison, we present HelixMC simulations that were necessary to shed light on puzzling prior results on dsDNA torsional stiffness. Measurements based on topoisomer distributions of closed dsDNA circles, fluorescence polarization anisotropy of intercalated dyes, and x-ray scattering of tethered gold nanoparticles give lower values for torsional persistence length (C = 25–80 nm [47], [65]–[68]) than measurements from optical and magnetic tweezers experiments (C = 100–120 nm [12], [16], [17], [21], [59]) from several different laboratories and with different tweezers geometries. One potential resolution to these discrepancies is that the apparent torsional stiffness of dsDNA is enhanced beyond its intrinsic value due to tethering constraints that attenuate torsional fluctuations in single-molecule experiments [44]. However, testing this hypothesis has been complicated by a prior inability to integrate link (number of helix turns) in base-pair-level simulations. Additional concerns have stemmed from the poor quality of fits to infer C from single molecule experiments with the analytical Moroz-Nelson formula [21], [22], which assumes the Fuller writhe expression and negligible self-avoidance effects.
To address these problems, we reasoned that the direct simulations enabled by HelixMC would reveal any systematic overestimation of intrinsic torsional persistence length due to tethering constraints or to the inaccuracy of the Moroz-Nelson model. First, we simulated link fluctuations in dsDNA helices as a function of force, analogous to experiments in references [12], [17], and computed the effective torsional persistence length Ceff by dividing the contour length of the polymer by the variance of the link (Table 2 and Fig. 4A–B). We first observed that the asymptotic value of Ceff (29–40 nm) in our simulation was within error of the ‘intrinsic’ value computed from a normal mode analysis (37.5 nm [43]), suggesting that C is not overestimated due to the tethering setup in single molecule experiments. We also tested the effects of x-y constraints (perpendicular to the direction of pulling) that might dampen torsional fluctuations, although such constraints are negligible in magnetic tweezers setups (and would also be expected to have a suppressive effect on bending fluctuations). Applying a harmonic x-y restoring force with strength of 0.025 pN/nm gave no significant change in Ceff (Fig. S2), disfavoring tether constraints as an explanation for the high C anomaly. Second, to test the use of the Moroz-Nelson formula, we fit these simulation data to the Moroz-Nelson model, and found excellent agreement with the same C values as described above. The rarity of self-clashing conformations (Supplementary Results and Table S6) and validity of the Fuller writhe formula above 0.4 pN further supported the use of this analytical fit. As a final crosscheck, we also computed the torsional persistence length using the slope of torque vs. number of turns in independent link-constrained simulations at 7 pN, analogous to an alternative experimental approach [16], [59], [62] (Fig. 4C–D, Supplementary Methods). This second simulation method gave torsional persistence length values that agreed well with the first method (within 1%, Table S7), confirming the robustness of the simulation method and Moroz-Nelson fits for inferring C in a way that matches experimental procedures.
Given the checks above, the discrepancy between the simulated dsDNA torsional persistence length C = 28.8 nm and the value in single molecule experiments C = 109 nm cannot be easily explained by systematic errors in the modeling. Furthermore, the deviation of experimental measurements from the Moroz-Nelson formula [16], [17] does not appear to be due to inaccuracies in this phenomenological model, given the successful fits of the model to simulated data. The discrepancies in C value and fitting curve strongly indicate either missing physics in modeling dsDNA in both the BPLM and simpler elastic-rod frameworks or currently unknown systematic errors in the experiment (see below, Discussion). Given these issues, we expected that our blind prediction for the torsional persistence length of RNA (C = 53 nm) might be an underestimate of the value measured from magnetic tweezers experiment. Indeed the experimental value was two-fold higher, with C = 100 nm. However, as with the dsDNA measurements, the Moroz-Nelson formula fit these experimental measurements relatively poorly (Lipfert et al. unpublished data), suggesting that some basic assumption of the BPLM approach is violated (see Discussion below).
The first measurements of helix mean end-to-end distance versus mean linking number for dsDNA highlighted gaps in theories of DNA elasticity [14], [15]. We thus expected that our final blind challenge, to predict analogous experiments for dsRNA, would provide a highly stringent test for HelixMC and the BPLM approach.
Before presenting the blind comparison, we describe simulation-based tests of assumptions made in the experimental inference of the link-extension coupling g (also described as twist-stretch coupling). In previous work, the coupling has been estimated from two different kinds of experiments: (1) stretching the polymer at different forces and observing how the linking number changes in the process [14], [69], and (2) setting up a constant stretching force and observing the polymer's extension as increasing numbers of turns are introduced [14], [15]. In both cases, bending fluctuations at low force (<15 pN) should, in principle, cause deviations from the linear relationships assumed to fit the experimental data (Supplementary Results, Fig. S3, S4). Nevertheless, linear relationships have been empirically observed for link and force (in experiment type 1) and of link and extension (in experiment type 2, but not in experiment type 1) for experiments on dsDNA. Furthermore, linear fits from these independent types of experiments gave consistent results (g = −90±20 pN·nm and −70±20 pN·nm, respectively); due to the convention in use, the negative sign corresponds to over-winding of the double helix upon extension (Table 2 and Fig. 5). This empirical relation was indeed confirmed in our simulations. We discovered linear correspondences between dsDNA link and extension in both types of simulated experiments, despite non-linear relationships of the underlying variables. The simulated dsDNA data gave couplings of g = −130 pN·nm and −150 pN·nm, respectively, for the two types of experiments, with systematic errors of ±30 pN·nm, based on alternative BPLM parameterizations (Table 2). The dsDNA calculations were therefore in agreement with experimental values within the estimated errors.
For dsRNA, the HelixMC-predicted g value was −120 pN·nm (from simulations of both types of experiments), with errors of ±40 pN·nm based on alternative BPLM parameterizations (Table 2 and Fig. 5). This predicted dsRNA value is the same, within error, as the dsDNA simulations. Nevertheless, separation of the link into twist and writhe components in the simulation suggested a different physical picture of link-extension coupling to dsRNA than for dsDNA. The simulated writhe vs. force slope is negative for dsRNA but nearly zero for dsDNA. This effect can be again attributed to the “springiness” of dsRNA axis curve, which carries an intrinsic writhe. Stretching dsRNA unwinds this writhe, while stretching dsDNA has little impact on its already straight axis curve. This behavior would result in a positive link-extension coupling g value, opposite in sign to dsDNA. However in the HelixMC dsRNA simulations, the helix twist, the other component of link, rises with extension and overpowers the writhe decrease to produce a net negative link-extension slope, matching the sign of dsDNA simulations.
The dsRNA tweezers experiments gave a value of g = +47±14 pN·nm, different from the value given by blind prediction (−120 pN·nm). This discrepancy is well beyond the error associated with different BPLM parameterizations, providing strong evidence against the current BPLM framework for modeling the torsional flexibility of dsRNA. Since the link-extension slope for RNA is a result of cancellation between a positive twist-extension correlation and a negative writhe-extension correlation, the predicted slope is quite sensitive to changes of many of the parameters of the underlying Gaussian potential (Supplementary Results, Table S8, S9). Indeed, by modification of the parameters, we were able to recapitulate the experimentally measured link-extension coupling, as discussed extensively in the experimental paper associated with this work (Lipfert et al., unpublished data). However we note here that this reparameterization is not unique, because the number of parameters (15, for a 6D covariance matrix) is far greater than the number of experimental measurements (four, i.e. bending persistence, stretch modulus, torsional persistence and link-extension coupling).
To understand the sequence-dependence of the mechanical properties being studied, and to propose future tests of the BPLM approach, we performed additional simulations of poly(A)/poly(T) and poly(G)/poly(C) for both DNA and RNA (which has U instead of T). Stretches of these homopolymer sequences play critical roles in accessibility of chromatin to RNA polymerase and transcription factors [70], [71]. We also performed simulations on Z-form DNA, which has been hypothesized to occur during DNA transcription to absorb torsional stress [72]. The results are listed in Table 2. For sequence-dependent simulations, we found that for poly(A)/poly(T) DNA, using the default dataset, all the measured mechanical properties increased by 1.5- to 3- fold compared to the random-sequence simulations. However if we used BPLM parameters from the 2.8_all dataset, which includes protein-binding DNA structures, the poly(A)/poly(T) results were not significantly different from the random-sequence results. The difference of predicted stiffness can be explained by the different underlying base-pair step parameters (Supplementary Results, Table S10). We also found smaller but measurable differences between other sequence-specified and random-sequence simulations, and between sequence-specified simulations performed with different base-pair step parameter sets. Further experimental comparisons between sequence-specific and random-sequence DNA/RNA will provide stringent tests of these predictions and to help discriminate which dataset (if any) is more accurate in modeling the sequence-dependence of the mechanical properties.
Simulations of Z-DNA gave dramatically higher bending and torsional persistence lengths (175 nm and 125 nm, respectively) compared to random B-DNA (55 nm and 29 nm, respectively). Again, this higher stiffness is encoded in the underlying step parameters (Supplementary Results, Table S10). Furthermore, the link-extension coupling is estimated to be near zero; this value arises from a complicated cancellation of twist and writhe, and is difficult to explain with simple arguments. Our simulation results agree with data obtained by Thomas and Bloomfield [73] indicating Z-DNA to be much stiffer than B-DNA, with a bending persistence length of 200 nm. However, previous studies on Z-DNA using light scattering, electron microscopy and fluorescence anisotropy have led to inconsistent results, with bending persistence length ranging from 21 to 200 nm and an extremely low torsional persistence length of 7 nm [73]–[75]. These studies did not agree on whether Z-DNA is stiffer then B-DNA. Additional single-molecule tweezers experiments on Z-DNA appear necessary to resolve these issues, and would provide stringent tests of the BPLM approach.
We have presented a set of fundamental tests of how well base-pair level models predict the flexibility of double-stranded nucleic acids, motivated by a desire for improved rigor in this field and by recent single-molecule measurements of dsRNA helices that were blinded to the modelers. A new software package HelixMC that integrates rigorous treatment of twist, writhe, and link allowed direct simulations of dsDNA and dsRNA tweezers experiments with base-pair level models. By fitting the simulated observables with the same analytical models used in experimental measurements, we were able to make direct comparisons of simulation and theory for properties including the bending persistence length, stretch modulus, torsional persistence length and link-extension coupling. We obtained predictions that match some experimental observations, particularly in the ratios of dsRNA to dsDNA values for mechanical properties like bending persistence length. However, we observed quantitative discrepancies for torsional persistence length at high force and the incorrect sign of the link-extension coupling constant for dsRNA. An extensive set of simulations checked that assumptions such as the effects of tethering, the Moroz-Nelson model of torsional persistence length, the curation of the database used to parameterize the BPLM, and the fitted relation of force and link could not account for these discrepancies.
The discrepancies between the BPLM model and tweezers measurements could be due to at least five reasons. First, electrostatic repulsion may account for some discrepancies, but it is difficult to see how corrections needed to increase the torsional stiffness of simulations by three-fold would not also substantially increase the simulated bending stiffness beyond the current values, which agree well with experiments. Experiments with different ionic conditions (particularly highly screening conditions) would help bound these effects. A second possibility is that the base-pair step distributions observed in crystallized nucleic acids do not reflect the fluctuations of nucleic acids in solution [47]. In this case, however, neither a simple overall scaling nor the parsimonious adjustment of a few parameters suffices to bring simulated data into agreement with experiments. Large changes in multiple BPLM parameters are required, in different directions for dsDNA vs. dsRNA and beyond the systematic deviations seen in different curated crystallographic databases, especially to account for a sign change in dsRNA link-extension coupling while retaining the experimental value for dsDNA link-extension coupling (Supplementary Results and Table S8, S9). A third explanation might involve thermal fluctuations involving bulges or non-Watson-Crick pairs, as have been resolved recently albeit with rare population [76]; the population of these alternative structures could be potentially enhanced during torsional stress. Due to the energetic cost of such fluctuations, we would predict that they would lead to a strong temperature dependence of torsional properties. Fourth, the conformation of each base-pair step may affect neighboring base-pair steps. Recent Au-SAXS scattering experiments and crystallographic analyses have suggested the importance of such correlations [47], [77]. Preliminary tests with multi-base-pair fragments in HelixMC indicate that such correlations may have up to 2-fold effects on predicted tweezers-measured properties (Supplementary Results and Fig. S5, S6).
A final explanation for the discrepancy involves the applied tension in single molecule tweezers experiments. On one hand, the tweezers data at low force (<5 pN) are used to infer the bending persistence length A and low-force effective torsional persistence lengths Ceff. These parameters are sensitive to both bending as well as intrinsic torsional persistence length via fluctuations captured by the Moroz-Nelson model. In this low force regime, BPLM gives predictions for both parameters with less-than-two-fold discrepancies, for both dsDNA and dsRNA. On the other hand, forces higher than 4 pN are required to suppress bending fluctuations and thereby to isolate stretch modulus S, intrinsic torsion persistence length C, and link-extension coupling g. For these values, the BPLM predictions do not agree with dsDNA or dsRNA measurements. Indeed, there is a more fundamental discrepancy: while the Moroz-Nelson model accounts for the predicted torsional persistence length vs. force from BPLM calculations over a wide range of model parameters, the experimental measurements of Ceff at forces >2 pN cannot be fit by this analytical model. These high-force discrepancies could be rationalized by a model in which tensions greater than 1 pN favor structural states that are more pliant to stretching but torsionally stiffer than the ensemble of conformations seen in crystallized dsRNA and dsDNA. Nucleic acids in solution under constant tension or strong torque, as might be provided by solution-based tweezers [78] or circularization, may enable bulk experimental methods like NMR or Au-SAXS to test this model. It is also possible that single-molecule tweezers experiments on alternative polymers such as poly(A)/poly(T) or Z-form DNA (simulated above) will agree well at all forces with BPLM predictions and thereby offer a baseline for comparison to the mixed sequence dsDNA and dsRNA cases. Alternatively if atomic-level computational methods could predict the structure of the putative weakly stretched state and design sequences or atomic modifications that favor it, the HelixMC toolkit should be able to integrate predictions for long helices that can then be precisely tested through future tweezers experiments.
The BPLM framework has been described in detail in previous studies [33]. Briefly, each base pair in the nucleic acid is represented by a vector representing the base-pair center and by a coordinate frame representing the orientation of the base-pair [42]. The degrees of freedom of the system are the base-pair steps, defined by the transformation of coordinates from one base-pair to the next base-pair. Each step is described by six parameters (shift, slide, rise, tilt, roll and twist) [79]. The transformation of the step parameters to Cartesian coordinates follows the Calladine and El Hassan Scheme (the CEHS definition) [80], which is also the convention used in the 3DNA package [81], [82]. The ‘technical details’ section of the 3DNA manual offers comprehensive examples of this scheme.
In HelixMC, the origin and the frame of the first base-pair is placed at the origin of the global coordinate system. That is, the base-pair center is placed at the coordinate origin; the normal vector of the base-pair is aligned with the z-axis; and the long-axis of the base-pair lies on y-axis. In terms of experimental setup, this placement is analogous to fixing one end of the nucleic acid to a surface (i.e. the xy-plane in our simulation), an approach routinely employed in magnetic and optical tweezers studies.
Once the origin and the frame of the first base-pair are set, the coordinates of the entire helix can be computed from the six base-pair step parameters. In HelixMC, the conformation of helix is stored and updated in this space of the step parameters, instead of in the Cartesian space. This is similar to describing protein conformations with the internal torsion angles instead of using the Cartesian coordinates of the atoms. For each base-pair step, we assumed the six step parameters form a multivariate normal distribution, of which the parameters were derived by surveying the existing RNA crystal structures (see below). This assumption is equivalent to assuming that positions and orientations of adjacent base-pairs are constrained by a six-dimensional harmonic potential [33].
In this work, the BPLM system was simulated using the Monte Carlo (MC) algorithm. A typical MC run consists of tens of thousands of cycles. A sample, which includes the current extension and linking number of the helix, was extracted at the end of each cycle (i.e. number of cycles equals to number of samples in the simulation). For each cycle, the base-pair steps of the entire helix was updated sequentially starting from the first base-pair step. For each update, a proposed move was generated by modifying only the conformation of the target base-pair step, while keeping the conformation of the rest of the helix intact. Note that the term “conformation” here refers to the six step parameters of each base-pair step, not the Cartesian coordinates of the base-pairs. Because we assumed the step parameters follow a multivariate normal distribution, this proposed conformational move can be efficiently achieved by drawing a random sample from the distribution.
The standard Metropolis criterion [83] was then used to whether to accept the proposed MC move:(1)Here ΔE equals the energy after the proposed move minus the energy of the initial conformation, T is the temperature and kB is the Boltzmann constant. Because the internal interactions between the base-pair steps are included in the multivariate Gaussian sampling, the ΔE in Eq. (1) only reflects the applied torque and force, as described next. For cases where external forces and torques are absent (free helix), the ΔE is always zero and the acceptance rate is 100%. For cases with external forces and torques, since each update is applied to one base-pair step only, the new proposed conformation is usually similar to the previous conformation. Therefore the acceptance rates are reasonable in the force and torque range used in this work (8% (40 pN) to 55% (1 pN) for dsDNA, Table S11).
We performed two types of simulations. In the first type of simulation, a stretching force along the z-direction was applied to the free end of the nucleic acid (the other end was fixed to the origin), and no torsional constraint was applied to the system. The energy of the system due to the applied force was(2)Here F is the applied stretching force, and z is the helix extension. This simulation was equivalent to the measurement of force-extension curves in typical single-molecule magnetic tweezers or constant-force optical tweezers experiments [8], [13], [62], [84]–[86].
In the second type of simulation, the nucleic acid was subjected to a fixed stretching force and was required to maintain a link (which is equivalent to the bead rotation) close to a target value through a harmonic potential. The energy of the system was:(3)Here krot is the stiffness of the torsional trap (200 pN·nm by default), Lk is the helix link, and Lkt is the target link of the trap. This type of simulation corresponded to torsion-trapped tweezers experiments [14]–[17].
In both types of simulations, we computed the base-pair center and the coordinate frame of the terminal base-pair as well as the overall link of the helix after each full-helix MC update. The number of base pairs in the simulated double helices was set to 3,000 (3 kbp) in this work unless stated otherwise.
At the beginning of the simulation, we initialized the helix by assuming that all base-pair steps have step parameters equal to their average values in the input parameter database. We then performed by default 120 cycles of full-helix MC updates to relax the helix under the specified stretching force (but no link-constraint). For link-constrained simulations, we performed further relaxation steps analogous to the torsional trap experiments, which involve slowly rotating magnets of the torsional traps to bring the helix from zero-turn state to a highly twisted state. We first turned on the link constraint, but set initial target link equal to the current link of the helix. Then we performed the following cycles:
After this “trap-ramping” step, we further relaxed the helix under the specified force and link constraint for 50 cycles. These relaxation steps ensured that the state of the helix at the beginning of the simulation was random and representative of the specified force and link constraint, without memory of the initial conformation.
In the HelixMC package, all the parameters discussed above, including the number of base-pairs and the applied external forces and link constraint, can be modified by user inputs. The details of the setup of the HelixMC calculations reported in this work are given in Supplementary Methods. We set the number of samples collected during our simulations to ensure that the standard errors of the average extensions and links were below 0.2% (Table S12).
Computing torsional properties and modeling torque in HelixMC required the integration of mathematical formulae developed in a number of separate papers by different authors. To document our final approach, we describe these equations and their connections here in some detail.
The observed bead rotation in a single-molecule tweezers experiment is mathematically described by the link (also known as the linking number). The original definition for the link of circular dsDNA is based on a closed continuous ribbon model [87]–[91]. A ribbon is defined by two mathematical objects: an axis curve, which is a smooth non-self-intersecting closed curve following the axis of the polymer; and a set of ribbon vectors, which are unit normal vectors everywhere along the axis curve that are perpendicular to the axis curve and pointing to reference points on the polymer [91]. To compute the link, we followed previous work by Britton et al. [56] to convert the BPLM to a ribbon model (Fig. 6A). Here we defined the axis curve to be the line connecting the base-pair centers (black vectors, also known as the base-pair centerline), and the ribbon vectors to be the long-axis of the base-pair (red vectors). This discretization scheme leads to a polygonal axis curve where multiple straight lines are joined by sharp bends (at the base-pair centers), and the ribbon vectors are defined only at each bend. While this discretization is simple and easy to manipulate numerically, it leads to two problems that forbid direct applications of the formulations for the closed continuous ribbon model to the BPLM. First, the discretization leads to an axis curve with discontinuous first derivatives at each bend. Therefore the tangent vectors at these bends are ill-defined, and the corresponding ribbon vector is in general not perpendicular to both the axis curve segments connected to the bend. This behavior invalidates the original assumption that the axis curve is smooth and the ribbon vectors are always perpendicular to the axis curve. Second, the BPLM we studied here is for open duplexes, different from the closed curve assumption in the conventional treatment.
By the Călugăreanu theorem (also known as the White's formula, or the Călugăreanu-White-Fuller theorem), link equals the sum of writhe and twist [87]–[90]. Intuitively, writhe represents the degree of coiling of the ribbon axis curve, and twist represents the amount of internal twist stored in the ribbon due to the local rotations of ribbon vectors. The sum of coiling and internal twist gives the overall bead rotation of the ribbon. In the following sections, we discuss separately how to compute the writhe and twist for such an open, polygonal ribbon.
Before discussing the writhe calculations for the BPLM, we first review the original definition of writhe, which described the coiling of the axis curve. The writhe of a smooth closed ribbon can be computed using the Gauss linking integral:(4)Here r1 and r2 are the Cartesian coordinates of the axis curve, r12 = r1−r2 is a vector connecting points r1 and r2, and we compute writhe (and, below, link and twist) in units of radians. Note that writhe only depends on the axis curve of the ribbon. Fuller proposed a simplified version of this integral [91]:(5)Here ez is a unit vector aligned with z-axis, and t is the tangent vector of the axis curve. The Fuller writhe simplifies the original double integral into a single integral but is only correct modulo 4π.(6)Here the expression “a≡b (mod n)” means(7)Mathematically speaking, a and b are said to be congruent modulo n.
The calculation of writhe of BPLM in this work is based on previous studies on polygonal open curves [53]–[55]. In the section below, we will derive the formulas for computing writhe in BPLM, mainly following the approach developed by Rossetto and Maggs [54].
The twist for a smooth ribbon can be computed as(21)Here t is the tangent vector of the axis curve, and l is the normalized ribbon vector. Unlike writhe, twist is a local identity, well defined on a curve segment of arbitrary length. Therefore twist is well defined for a smooth open curve. In addition, twist is additive. For our polygonal ribbon, the overall twist of the ribbon equals the sum of the twists of all the line segments. As an example, consider a straight line segment parallel to z-axis of length L (Fig. 6C). The ribbon vector starts as l0, varies smoothly and ends as l1. Using the fact that the tangent vector t = ez and the ribbon vectors are perpendicular to t, Eq. (21) can be evaluated as(22)Here we used the property that l×dl is parallel to ez. Geometrically, this integral is twice the area on unit circle swept by l throughout the integration. Therefore the twist of a straight line segment is just the angle (in radians) between the vectors l0 and l1. This result is consistent with the conventional definition of twist parameter in a base-pair step.
However, applying the above result for straight line segments to our polygonal ribbon is nontrivial, because here the ribbon vectors are not necessarily perpendicular to the straight line segments. A naïve strategy would be to simply sum the twist parameters of all base-pair steps in the helix to obtain the overall twist, but this sum turns out to be inconsistent with the ribbon twist considered in the Călugăreanu theorem. It thus cannot be added with writhe to produce a link that corresponds to the actual experimental observable of, e.g., bead rotation in a magnetic tweezers experiment. As pointed out by Britton and colleagues [56], the ribbon twist of dsDNA (‘twist’ discussed below refers to the ribbon twist, unless stated otherwise) is different from the conventional definition of twist parameter for a base-pair step, necessitating a new procedure to calculate twist for base-pair steps.
The main challenge in computing twist for the discrete chains of the nucleic acid helix is that the ribbon vector at each base pair, li, is not in general, normal to the continuous axis curve traced by base pair centers ri, as is assumed in the mathematical treatment of ribbons. Our strategy therefore is to first define at each base pair a ‘reference’ ribbon vector bi that obeys this mathematical convention, and to compute a reference twist. We will then compute additional twist contributions by li using its angle with bi. Fig. 6A illustrates the polygonal ribbon model. The choice of , where ti−1 and ti are unit vectors pointing into and out of ri, guarantees normality of bi to the axis curve. Then we can compute the reference twist based on the above result for straight line segments:(23)Here N is the total number base-pairs in the model, Tw1 and TwN−1 is the twist contribution of the first and the last base-pair steps (N−1 base-pair steps in total), where b's are not defined. βi is the signed angle between the reference ribbon vector bi and bi+1. Note that because both bi and bi+1 are orthogonal to ti, βi is also the dihedral angle bi - ti - bi+1 (Fig. 6A, inset).
The use of alternative reference ribbon vectors to compute the twist can be justified with the following thought experiment. Imagine holding the two ends of a continuous ribbon, and then change the ribbon vectors by rotating the ribbon in the middle. As long as the two ends stay fixed, such changes of ribbon vectors do not affect the overall number of turns of the ribbon (i.e. the link). In addition, the writhe stays constant because it only depends on the axis curve, which is unmodified in this process. By the Călugăreanu theorem, we can conclude that the twist, which equals the link minus writhe, remains unchanged. Therefore in a continuous ribbon we may modify any ribbon vector except the two ends without affecting the overall twist. However for a discretized ribbon (as in our BPLM), such modifications of ribbon vectors may change the twist by 2nπ, where n is an integer (Fig. 6D). In general, we have the following modulo congruence relation between the true twist and reference twist (see Eq. (7) for definition of modulo congruence):(24)To address the modulo 2π ambiguity, we must take into account whether the original ribbon vectors li sweep out additional turns around the axis curve relative to the reference ribbon vectors bi. Here we calculate the local twist of each base-pair step as:(25)Here αi is a signed angle between li and bi; Ti is folded into the range [−π, π) upon the modulo 2π operation. For the terminal base-pair steps, we first attach virtual segments to both ends, pointing towards −z and +z respectively, to obtain the corresponding bi, then Eq. (25) can be employed to compute T1 and TN−1 (illustrated in Fig. 6D). The overall twist can then be calculated by summing all the Ti:(26)
As an additional consistency check, Eq. (26) satisfies Eq. (24), as shown below.(27)The factors α1 and αN correspond to the twist contribution from two ends of the helix; all internal factors cancel.
Our twist definition is similar to the definition proposed by Britton and colleagues [56]. The main difference is in the definition of α angle. The previous proposal defined the tangent vector at base-pair i as , then projected the original ribbon vector li to the plane defined by , to obtain the new ribbon vector di. Then α is defined as the angle between di and bi:(28)While this definition is mathematically correct and gives results equivalent to our definition, the expression can become ill-defined when there is a sharp bend in the ribbon. Fig. 6E illustrates such cases. In the first example, the original ribbon vector l1 is parallel to the tangent vector , therefore the projection gives a null vector d1, making α ill-defined. In the second example, we take l1′ as the ribbon vector, then the projection gives d1′ parallel to b1, leading to a zero α even though l1′ and b1 are quite distinct from each other. In both cases, our definition just sets α equals to the angle between l1 and b1 and the angle between l1′ and b1, leading to a well-defined result. While this type of sharp bend does not occur in natural dsDNA, in our system the line segment of the last base-pair step and the added virtual segment can form such sharp bends, making the previous definition unsuitable for HelixMC.
The definition in Eq. (25) has two additional convenient properties. First, if the axis curve of the dsDNA/dsRNA is perfectly straight and pointing towards +z, and the base-pairs are all parallel to the xy-plane, the calculated ribbon twist equals to the sum of the base-step twist parameters [56]. Second, in our system setup, if the normal vector of the last base-pair aligns along +z, the computed link corresponds exactly to the bead rotation observed in single molecule tweezers experiments.
The multivariate Gaussian distributions for sampling are constructed using the base-pair step parameters from crystallographic models in the PDB. To ensure the quality of the data in the default parameter sets, we used models derived from data with resolutions better or equal to 2.8 Å. Protein-binding DNA/RNA structures were excluded from the dataset since protein binding may affect the deformability of the nucleic acids. We also tested several other selection schemes to estimate the systematic error, including using higher resolution cutoff (2.0 Å) or including protein-binding structures. We then used the 3DNA software [81], [82] to extract the base-pair step parameters for the canonical Watson-Crick base-pairs (i.e. not including G-U wobble base-pairs and other non-canonical base-pairs). Parameter sets with twist ≤5° (due to Z-DNA conformations), with rise ≥5.5 Å (due to ligand intercalation), or with any value more than four standard deviations away from the mean were discarded as outliers. For the dsDNA datasets, we noticed that there were two major clusters in the data, corresponding to the A-form and B-form dsDNA (except for the ‘DNA_2.8_all’ dataset, where the protein binding rendered A-DNA and B-DNA inseparable by clustering). We used the k-means algorithm to separate the two clusters and only used the B-DNA parameters in sampling (Table 1, Supplementary Methods and Fig. S7). For Z-DNA, the dataset is composed of two distinct base-pair step distribution for the GC steps and the CG steps. The population distribution for the Z-DNA dataset is shown in Fig. S8. Detailed statistics and population distribution of the curated dataset are shown in Table 1, S13 and Fig. 7. Fig. S9 shows the correlation plots between each base-pair step parameter for the default dataset.
For each type of base-pair step (16 in total, e.g. 5′-AT-3′/5′-AT-3′, 5′-CA-3′/5′-TG-3′, etc.), a multivariate Gaussian was fitted based on the corresponding six base-pair step parameters, enabling sequence-dependent simulations. We note here that one can also categorize the base-pair parameters into 10 independent sequence-specific categories using the symmetry of the base-pair steps [40], [80]. This symmetrization is not the default option in HelixMC; symmetrization gives minor changes in the predicted mechanical properties (Supplementary Methods and Table S2).
For simulations with random sequence, in each update we randomly picked a distribution from the 16 types of step parameters and drew samples from it. In this sampling scheme, we effectively averaged the 16 types of parameters, so all base-pair steps follow the same parameter distribution [40]. The distribution can be further approximated with a single multivariate Gaussian (Supplementary Results). The approximation leads to a reduced number of parameters in the model, and therefore facilitates the understanding the effect of each parameter on the observed mechanical properties. However, we note that this sampling scheme may lead to unrealistic base-pair step combinations (for example, 5′-AT-3′/5′-AT-3′ followed by 5′-GC-3′/5′-GC-3′), therefore the sequence of the RNA is not always well defined in each simulation snapshot. To justify that our sampling scheme indeed gave reasonable estimates of the mechanical properties of a random RNA, we also performed simulations with a single randomly generated RNA sequence (Table S14). The obtained mechanical properties using a single random sequence agreed within simulation error to our default random sequence simulation.
In addition, we observed that some of the population distributions in our dataset did not appear Gaussian (Fig. 7). To test the validity of the Gaussian approximation, we also tested a different sampling scheme, by randomly picking parameter sets existing in the database without assuming Gaussianity, and obtained nearly undistinguishable results (see Results). All the curated parameter sets and sampling schemes used in this work are available and further documented in the HelixMC package.
The bottleneck steps of HelixMC have been optimized in C (using Cython); a typical single-point HelixMC calculation for a DNA/RNA helix of experimental length (few kilo-base-pairs) takes minutes to hours on a standard desktop computer (Table S15). HelixMC is coded in Python in an object-oriented fashion that allows easy modification and extension, is free and open-source (http://github.com/fcchou/helixmc), and enables fast and accurate predictions with available computational power.
We numerically tested the validity of the link calculations above by comparing simulated link values to the bead rotations (Fig. 8A). Here the “bead rotation” is defined as the angle between the y-axis of the global coordinate and the projection of the ribbon vector of the last base-pair to the xy-plane. This is equivalent to attaching a virtual bead along the ribbon vector of the last base-pair and observing its rotation, analogous to recent single-molecule tweezers experiments [17]. For better comparison, in Fig. 8A we folded the computed link into the range of [−π, π), and found that the experimentally observed bead-rotation indeed corresponds to the link. The match between the link and the bead rotation was close but not exact (RMSD of 4.5°), because the normal vector of the last base-pair did not point exactly to +z during the simulation; this discrepancy induces negligible error in computed helix mechanical properties.
As discussed above, the Fuller formula is faster but gives exact writhes only in certain conditions. The formula breaks down if the helix path fluctuates so that segments point away from the applied force (towards −z). Fig. 8B shows a plot of exact writhe vs. Fuller writhe in a simulation of 3 kbp dsDNA at 0.1 pN stretching force. It is apparent that in this setting the Fuller formula is only correct modulo 4π (two turns; spacing between parallel lines). To test under which conditions the Fuller formula was accurate, we computed its RMSD error to exact writhe across simulations. For force-extension simulations, Fuller writhe is effectively exact if the force is larger than 0.4 pN for dsDNA and 1 pN for dsRNA (Fig. 8C). For link-constrained simulations, the Fuller formula holds in the current simulated link-range, but breaks down when the target link exceeds ±40 turns for DNA and ±20 turns for dsRNA (corresponding to supercoiling densities of 0.022 and 0.012; Fig. 8D).
As with experimental measurements, the simulated data were summarized through fits to the elastic rod model, which assume that the total energy of the helix without external force and torque can be expressed using the above parameters by an integral along the helix axis curve s:(29)Here L is the helix contour length, kB is the Boltzmann constant and T is the temperature. The constants are bending persistence length A, B = S/kBT the stretching stiffness (where S is the stretch modulus), torsional persistence length C, and D = g/kBT is the unit-less link-extension coupling (here we use the convention in ref. [14], where g has units of pN·nm). The three quantities β, z and θ describe the deformations per unit length of a short rod segment. β is the bending deformation that measures how the tangent vector changes along the rod, z is the extensional deformation that measures the change in the length of the segment, and θ is the torsional deformation that determines how the each segment is rotated around the rod axis with respect to adjacent segment. The analytical equations used to fit experimental measurements, derived from this model, are compiled in the Supplementary Methods.
|
10.1371/journal.pbio.2006548 | Visual attention is not deployed at the endpoint of averaging saccades | The premotor theory of attention postulates that spatial attention arises from the activation of saccade areas and that the deployment of attention is the consequence of motor programming. Yet attentional and oculomotor processes have been shown to be dissociable at the neuronal level in covert attention tasks. To investigate a potential dissociation at the behavioral level, we instructed human participants to move their eyes (saccade) towards 1 of 2 nearby, competing saccade targets. The spatial distribution of visual attention was determined using oriented visual stimuli presented either at the target locations, between them, or at several other equidistant locations. Results demonstrate that accurate saccades towards one of the targets were associated with presaccadic enhancement of visual sensitivity at the respective saccade endpoint compared to the nonsaccaded target location. In contrast, averaging saccades, landing between the 2 targets, were not associated with attentional facilitation at the saccade endpoint. Rather, attention before averaging saccades was equally deployed at the 2 target locations. Taken together, our results reveal that visual attention is not obligatorily coupled to the endpoint of a subsequent saccade. Rather, our results suggest that the oculomotor program depends on the state of attentional selection before saccade onset and that saccade averaging arises from unresolved attentional selection.
| The premotor theory of attention postulates that spatial visual attention is a consequence of the brain activity that controls eye movement. Indeed, attention and eye movement share overlapping brain networks, and attention is deployed at the target of an eye movement (saccade) even before the eyes start to move. But is attention always deployed at the endpoint of saccades? Here, we measured visual attention before accurate saccades and before saccades that landed in between 2 targets (averaging saccades). While accurate saccades were associated with a selective enhancement of visual sensitivity at their endpoint, no such enhancement was found at the endpoint of averaging saccades. Rather, visual sensitivity was evenly distributed across the 2 saccade targets, suggesting that saccade averaging arises from unresolved attentional selection. Overall, our results reveal that attention is not always coupled to the endpoint of saccades, arguing against a simplistic view of the premotor theory of attention at the behavioral level. Instead, we propose that saccadic responses depend on the state of attentional selection at saccade onset.
| To process information from our rich visual environment, we evolved with attentional mechanisms allowing us to discriminate which flow to account for and which to ignore [1,2]. For example, we can extract salient saccade targets from a cluttered visual scene to later examine their contents with precise foveal vision [3–6]. This link between attention and saccadic eye movements led researchers to propose that spatial visual attention is directly dependent on the oculomotor system [7,8], introducing what they called the “premotor theory of attention.”
This influential theory relies on 2 main hypotheses. The first hypothesis states that visual attention is operated by the oculomotor system itself. Indeed, overlapping neuronal activations have been observed in visual attention tasks involving the deployment of attention with (overt) or without (covert) eye movements in functional magnetic resonance imaging (fMRI) [9]. These activations include cortical and subcortical areas such as the Frontal Eye Field (FEF), the parietal cortex, and the Superior Colliculus (SC).
At the behavioral level, there is indeed evidence for a concurrent encoding of spatial attention and saccade programming [10]. For example, various studies demonstrated that visual attention, measured as a local improvement in visual sensitivity, is allocated to the saccade target before the eyes start to move [11,12]. Nevertheless, some other studies suggested that saccade preparation does not necessarily entail a shift of attention towards the saccade goal, casting some doubt in regard of the coupling between attention and oculomotor control [13–16].
The second hypothesis of the premotor theory of attention implies that the deployment of visual attention is always preceded by an activation of the oculomotor system. Under this hypothesis, covert attention involves the preparation of a saccade that is canceled before the eyes move. In line with this hypothesis, subthreshold microstimulation of the FEF or the SC, which did not systematically lead to a saccade, resulted in attentional benefits measured both behaviorally and electrophysiologically at the stimulated movement field position [17–20]. However, because microstimulation effects cannot be solely restricted to the motor cells within the stimulated areas, these results did not demonstrate that the deployment of visual attention is preceded by a premotor activation alone. Instead, it was shown that motor cells within FEF or SC stayed completely silent during a covert attention task [21–23], while visual and visuomotor cells displayed sustained attentional effects. In other words, attention is not always preceded by motor activity, at least not within these recorded oculomotor centers.
To shed light on this controversy and to test this second hypothesis at the behavioral level, one can imagine measuring visual sensitivity at the intended saccade goal and at the endpoint of the saccade. Under such conditions, measured sensitivity should correlate with the activity of both the visual and motor cells within oculomotor centers. Taking advantage of the fact that saccades tend to undershoot the target, Deubel and Schneider [12] found that attention was restricted to the intended saccade goal rather than to the saccade endpoint. However, using saccadic adaptation to decrease the saccadic gain, some authors found the exact opposite effect, with attention allocated to the adapted saccade endpoint rather than to the intended saccade goal [24,25]. Knowing that oculomotor centers have several overlapping large receptive fields within the range of these effects [26,27], it is hard to link these contradictory behavioral findings to the neurophysiology described above.
Here, we thus propose to use a paradigm leading to a larger spatial dissociation between the intended saccade goal and the saccade endpoint, such as the global effect [28–31]. Indeed, the global effect is associated with systematic and large saccade endpoint deviations towards the center of gravity of 2 saccade targets [28,32,33], or of a saccade target and a distractor [34,35], shown at 2 positions separated by up to 60° of rotation [34]. Although the global effect was originally described as reflecting a low-level averaging of neuronal activity (and therefore respective saccades are often called averaging saccades) within the oculomotor centers [28,36,37], different behavioral observations later suggested a dependency on higher-level attentional processes. First, it was shown that averaging saccades can be elicited by second- and third-order saccade targets [38,39], suggesting that the global effect cannot merely reflect low-level oculomotor processes. Next, it was shown that specifying the location [40,41], the identity [42,43], or the probability of a saccade target to appear at a certain location relative to a distractor [44] systematically reduced the occurrence of averaging saccades. Monkeys make averaging saccades when the FEF or the SC are simultaneously microstimulated at 2 sites [45–48] and when 2 targets are shown in close proximity [49,50]. At the neuronal level, it was first proposed that a single peak of motor cell activity associated with saccades ending in between 2 targets precedes an averaging saccade [51,52]. Later work suggested instead that averaging saccades follow 2 peaks of activity associated with saccades directed towards the 2 saccade targets [53,54]. Recently, Vokoun and colleagues [55] used voltage imaging of slices of rat SC to record population dynamics in response to dual-site electrical stimulation. They observed that the simultaneous stimulation of 2 nearby sites in the intermediate layers led to a merged peak centered in between them in the superficial layers. Moreover, they proposed that such merged activation feeds back into the visual system, leading to the perception of a target at the averaging saccade endpoint.
If this proposal of a feedback of merged activation from the superficial layers of the SC into the visual system was true, we would expect to find a presaccadic enhancement of attention at the endpoint of averaging saccades, a result that would be in line with the premotor theory of attention. Van der Stigchel and de Vries [56] directly tested this proposal, instructing participants to move their eyes towards a saccade target presented simultaneously with a distractor and measuring presaccadic attention at these positions as well as in between them. They observed both averaging saccades as well as saccades directed towards the target and the distractor, allowing them to compare the deployment of attention at the intended saccade goal and at the saccade endpoint. Unfortunately, they reported no main effect of the saccade landing direction as well as no interaction between the saccade landing direction and the position of their attention probes when analyzing visual discrimination performance as a function of the saccade endpoint. Therefore, contrary to many reports [11,12], the saccade landing position had no significant effect on the deployment of attention in their paradigm, preventing any conclusion about whether or not attention is deployed at the endpoint of averaging saccades.
Other studies suggested that attention is not necessarily allocated to the saccadic endpoint [11,44] or argued that saccades towards the center of gravity within extended target configurations are based on the computation of a central reference point via spatial pooling [57,58]. However, none of these studies measured visual attention at the averaging saccade endpoint to determine whether averaged oculomotor programs are associated with attentional averaging. Here, we measured visual attention at various locations in space, including the averaging saccade endpoint, in a free-choice saccade task that entailed the presentation of 2 nearby saccade targets. Our design therefore allowed us to investigate whether attention is allocated at the endpoint of averaging saccades. More specifically, given the spatial resolution of our design, we could distinguish the following 3 possible outcomes related to the deployment of visual attention before averaging saccades: (a) attention is deployed at the exact location of the saccade endpoint, (b) attention spreads across an extended area including the saccade endpoint, and (c) attention is deployed at 2 discrete saccade target areas flanking the saccade endpoint but not at the endpoint itself. We observed a presaccadic enhancement of visual sensitivity at the endpoint of accurate but not averaging saccades, ruling out an obligatory coupling of attention to the endpoint of a subsequently executed saccade (against [a]). Contrary to the idea of an extended spread of attention around the center of gravity, averaging saccades were associated with moderate enhancement of visual sensitivity at the 2 saccade targets (against [b]). Our results instead suggest that the oculomotor program depends on the state of attentional selection before saccade onset, with attention being deployed at the 2 discrete targets (favoring [c]) and saccade averaging resulting from uncompleted attentional selection.
Our goal was to determine whether the presaccadic deployment of attention is obligatorily coupled to the saccade endpoint. To do so, we probed visual attention at various locations while participants prepared a saccade towards 1 of 2 potential saccade targets, presented either transiently or continuously and separated by an intertarget angular distance of either 90° or 30° (Fig 1A). Just before the saccade, a discrimination target was shown randomly across trials at 1 of the 2 potential saccade targets (ST1 and ST2), at the position in between the saccade targets (BTW), or at 1 of 21 equidistant control positions (CTRL).
Fig 1B shows the normalized frequency of saccade landing endpoints observed across participants within the 90° and 30° condition, irrespective of the duration of the saccade targets (i.e., transient and continuous combined). While saccades were equally distributed over the 2 saccade targets in the 90° condition (Fig 1B, top), a substantial proportion of saccades ended in between them in the 30° condition (Fig 1B, bottom). To further analyze our data, we looked at the distribution of saccade landing directions either binned in evenly distributed angular sectors of 5° (Fig 2A and 2B) or 15° (centered on the 24 stimuli streams, Fig 2C and 2D). In the 90° condition (Fig 2C), 41.0% ± 1.0% of the saccades ended within the sector including ST1 (most counterclockwise saccade target) and 41.8% ± 1.9% within the sector including ST2 (most clockwise saccade target). Note that an average of 4.0% ± 0.9% of saccades ended within the sectors adjacent to the saccade targets. In the 30° condition (Fig 2D), 33.6% ± 2.4% of the saccades ended within the sector in between the 2 saccade targets (BTW), while 29.95 ± 1.6% of the saccades ended within the sector of ST1 and 32.0% ± 1.8% within the sector of ST2. Therefore, when participants had to select between 2 equidistant saccade targets separated by an angular distance of 30°, they executed an averaging saccade (ending in the BTW sector) in about one-third of the trials. For further inspection, saccade endpoint distributions as a function of saccade latency are provided for each participant in S1 Fig. In order to determine potential differences between the 2 intertarget angular distance conditions (90° and 30°), we first looked at saccade latencies and amplitudes. We found slightly longer saccade latencies (90°: 192.2 ± 1.7 ms versus 30°: 188.2 ± 2.2 ms; p = 0.0012) and larger amplitudes (90°: 10.0 ± 0.1° versus 30°: 9.7 ± 0.1°; p = 0.0002) in the 90° as compared to the 30° condition. Saccade latency did not differ as a function of the saccade landing position (ST1, ST2, or BTW) both in the 90° and 30° condition (all p > 0.05, Fig 2E and 2F). In the 90° condition, amplitudes of saccades towards ST1 (10.1 ± 0.1°) and ST2 (10.0 ± 0.1°) did not differ significantly from each other (ST1 versus ST2: p = 0.7902), whereas amplitudes of saccades towards BTW (7.9 ± 0.2°) were significantly smaller than those of saccades towards ST1 and ST2 (both p < 0.0001) (see Fig 2G). In the 30° condition, amplitudes of saccades towards ST1 (9.7 ± 0.1°) and ST2 (9.8 ± 0.1°), as well as towards ST1 and BTW (9.7 ± 0.1°), did not differ significantly from each other (ST1 versus ST2: p = 0.2216; ST1 versus BTW: p = 0.5998), whereas amplitudes of saccades towards ST2 were significantly larger than those of saccades towards BTW (ST2 versus BTW: p = 0.0118) (see Fig 2H). Note that the proportion of averaging saccades did not vary as a function of saccade latency. Comparing trials of the 30° condition separated in 2 equal groups of early (167.1 ± 1.8 ms) and late (209.3 ± 3.2 ms) saccade latencies, we found a comparable proportion of averaging saccades (early BTW: 35.1 ± 3.0% versus late BTW: 32.1 ± 2.2%; p = 0.1632). This effect is most likely the consequence of the instruction given to the participants to saccade as fast as possible, such that early and late averaging saccade latencies differed by less than 40 ms (early BTW: 168.2 ± 2.0 ms versus late BTW: 207.4 ± 3.1 ms; p < 0.0001). However, we found that the mean absolute saccade endpoint deviation relative to the BTW location slightly increased as a function of saccade latency (see A-B in S2 Fig and A-B in S2 Fig for individual participant data for both the 90° and 30° conditions). Thus, saccade averaging was more pronounced for short-latency saccades. Overall, for each intertarget angular distance, we observed either no differences or only some nonsystematic differences of a few milliseconds and a few minutes of arc. Although saccade latencies and amplitudes did not differ much between these conditions, the saccade landing-direction distributions reflect 2 distinct oculomotor modes as a function of the intertarget angular distance. Saccades were mostly accurate in the 90° condition, whereas we observed both accurate and averaging saccades in the 30° condition.
Our paradigm allowed us to measure both the oculomotor behavior and the presaccadic allocation of attention through the presentation of a discrimination target at 1 of 24 possible positions. We first verified that the presentation of the discrimination target itself did not systematically influence oculomotor behavior. We did not find any differences with respect to saccade latency and amplitude when comparing trials with and without the presentation of a discrimination target (3.5% of trials were without discrimination target, both p > 0.05). This result validates that the distractor streams and, in particular, the presentation of a discrimination target did not bias the deployment of attention. Fig 3A and 3B shows visual sensitivity as a function of the discrimination target position rotated as to align the 2 saccade targets around the geometrical angle zero in both the 90° (Fig 3A) and 30° (Fig 3B) condition. Irrespective of the duration of the saccade targets, we found higher sensitivity for discrimination targets shown at the saccade targets than at the control positions (corresponding to the average across all positions except for ST1, ST2, and BTW) in both the 90° (ST1: d’ = 2.2 ± 0.3 versus CTRL: d’ = 0.3 ± 0.1, p < 0.0001; ST2: d’ = 2.2 ± 0.4 versus CTRL, p < 0.0001; ST1 versus ST2, p = 0.8964; Fig 3A) and the 30° (ST1: d’ = 2.2 ± 0.3 versus CTRL: d’ = 0.3 ± 0.1, p < 0.0001; ST2: d’ = 2.1 ± 0.3 versus CTRL, p < 0.0001; ST1 versus ST2, p = 0.6026; Fig 3B) condition. These effects contrast with the low sensitivity observed for discrimination targets shown in between the saccade targets (BTW) in the 90° (BTW: d’ = 0.2 ± 0.1 versus ST1, p < 0.0001; BTW versus ST2, p < 0.0001) and especially in the 30° (BTW: d’ = 0.6 ± 0.2 versus ST1, p < 0.0001; BTW versus ST2, p < 0.0001) condition.
Thus, despite the fact that saccades landed in between the saccade targets in a third of the trials in the 30° condition, the overall sensitivity at this position stayed rather low. One should, however, note that sensitivity was still increased at this position compared to the control positions in the 30° condition (30°: BTW versus CTRL, p = 0.0010), whereas this was not the case in the 90° condition (90°: BTW versus CTRL, p = 0.7732). On the other hand, such slight facilitation observed in between the saccade targets in the 30° condition relative to the control positions was only observed for trials in which the targets were shown transiently (BTW: d’ = 0.8 ± 0.2 versus CTRL: d’ = 0.3 ± 0.1, p < 0.0001) but not continuously (BTW: d’ = 0.5 ± 0.2 versus CTRL: d’ = 0.3 ± 0.0, p = 0.10880). It is important to note that the discrimination target temporally overlapped with the saccade targets in the continuous but never in the transient condition. The observed difference between the 2 conditions therefore suggests that the appearance of a discrimination target at BTW was masked by the continuous presentation of the saccade targets. Altogether, the results above demonstrate that presaccadic attention was mainly allocated towards the saccade targets, and to a much smaller extent towards the position in between. This last result, however, cannot be attributed to a large spread of attention extending to more than 1 of the tested directions because we did not observe a consistent benefit at the 2 other positions adjacent to the saccade targets in the 30° condition (ST1 + 15°: d’ = 0.4 ± 0.1 versus CTRL: d’ = 0.3 ± 0.1, p = 0.0914; ST2 − 15°: d’ = 0.4 ± 0.1 versus CTRL, p = 0.0336; here, CTRL excludes ST1 + 15° and ST2 − 15°, respectively, in addition to ST1, ST2, and BTW) nor at the 4 adjacent positions of the saccade targets in the 90° condition (ST1 ± 15°: d’ = 0.3 ± 0.1 versus CTRL: d’ = 0.2 ± 0.1, p = 0.5742; ST2 ± 15°: d’ = 0.3 ± 0.1 versus CTRL, p = 0.3200; here, CTRL excludes ST1 ± 15° and ST2 ± 15°, respectively, in addition to ST1, ST2, and BTW).
At that stage, one cannot exclude the possibility that attention is always drawn towards the saccade endpoint before both accurate and averaging saccades because we found higher sensitivity for both the saccade targets—and, in the 30° condition, also for the position in between them—compared to the control locations. Although we found higher sensitivity at the saccade targets than in between them, this may just reflect the combined effect of the saccade preparation and the presence of visual cues (the saccade targets themselves). To estimate the effect of saccade preparation, we thus needed to specify our results depending on where the saccade ended within each trial. To do so, we redefined the position of the discrimination targets relative to the saccade direction. Fig 3C and 3D shows visual sensitivity as a function of the discrimination target position relative to the saccade direction. We found higher sensitivity for discrimination targets shown at the saccade targets when compared to the position in between them or to the control positions in both the 90° and 30° conditions, for trials in which accurate saccades were made towards ST1 (all p < 0.0001) or ST2 (all p < 0.0001). The same effects were found for averaging saccades in the 30° condition (all p = 0.00010). In addition to the facilitation effect of the saccade target presentation, we found that, irrespective of the intertarget distance (90° or 30°), sensitivity at ST1 was improved when an accurate eye movement was made towards ST1 (90°: ST1: d’ = 3.2 ± 0.5 versus ST2: d’ = 1.7 ± 0.4, p < 0.0001 [see blue lines and bars in Fig 3C and 3D]; note that in the 30° condition, sensitivity at ST1: d’ = 2.9 ± 0.4 was only marginally superior to those observed at ST2: d’ = 2.1 ± 0.5, p = 0.0740). The same selective improvement was observed at ST2 before the execution of accurate saccades towards it (90°: ST2 versus ST1, p < 0.0001; 30°: ST2 versus ST1, p = 0.0002 [see red lines and bars in Fig 3C and 3D]). In particular, preparing an accurate eye movement towards 1 of the 2 saccade targets improved sensitivity when comparing trials in which the discrimination target was shown at the saccaded location (e.g., DT at ST1 and saccade made towards ST1) to trials in which the discrimination target was shown at the same position when it was not the saccaded position (e.g., DT at ST1 and saccade landing at ST2 or BTW) in both the 90° (Fig 3E; ST1+2 saccaded: d’ = 3.0 ± 0.4 versus ST1+2 nonsaccaded: d’ = 1.7 ± 0.4, p < 0.0001) and the 30° (Fig 3F; ST1+2 saccaded: d’ = 2.7 ± 0.4 versus ST1+2 nonsaccaded: d’ = 2.0 ± 0.3, p = 0.0080) condition. Crucially for averaging saccade trials, for which the intended saccade goal (ST1 or ST2) and the saccade endpoint (BTW) were dissociated (see green lines and bars in Fig 3D), we found a rather low sensitivity for discrimination targets shown in between the saccade targets (BTW: d’ = 0.4 ± 0.2), highly reduced when compared to discrimination targets shown at the saccade targets (ST1: d’ = 2.2 ± 0.4 and ST2: d’ = 2.2 ± 0.4, both p < 0.0001). Furthermore, and contrary to above (Fig 3B), it was not different from the sensitivity gathered across the control locations (CTRL: d’ = 0.3 ± 0.1, p = 0.4026), both when the saccade targets were shown transiently or continuously (both p > 0.05). Thus, contrary to accurate saccades, the execution of averaging saccades did not lead to any improvement at the saccade endpoint. Moreover, a visual inspection of sensitivity as a function of the saccade latency shows a relative independence of these measures, suggesting that, irrespective of the saccade latency, attention was not deployed at the averaging saccade endpoint (see C-D in S2 Fig and A-B in S4 Fig for individual participant data in the 90° and 30° conditions). Visual sensitivity was significantly reduced at the intermediate location (BTW) before averaging saccades compared to saccades that landed at 1 of the saccade targets (Fig 3F; BTW saccaded: d’ = 0.4 ± 0.2 versus BTW nonsaccaded: d’ = 0.7 ± 0.2, p < 0.0001). This sensitivity reduction can, however, be mainly attributed to a masking effect of the continuous presentation of the saccade targets (BTW saccaded: d’ = 0.3 ± 0.3 versus BTW nonsaccaded: d’ = 0.7 ± 0.2, p = 0.0088) because it was not found for saccade targets presented transiently (BTW saccaded: d’ = 0.7 ± 0.2 versus BTW nonsaccaded: d’ = 0.7 ± 0.3, p = 0.9664).
These findings demonstrate, contrary to what is predicted by the premotor theory of attention, that the preparation of averaging saccades does not lead to a deployment of attention at the corresponding saccade endpoint. Instead, we found that averaging saccades were associated with an equal distribution of attention towards the 2 saccade targets (ST1: d’ = 2.2 ± 0.4 versus ST2: d’ = 2.2 ± 0.4, p = 0.8402). One interpretation of these effects could be that averaging saccades result from an unsuccessful or at least uncompleted presaccadic attentional selection among the 2 saccade targets, with resources equally distributed between them. On the other hand, it is possible that, despite landing in between the targets, presaccadic attentional selection was successful before averaging saccades but directed half of the time towards the most clockwise saccade target and half of the time towards the most counterclockwise saccade target. If this were the case, across trials, one would also expect to find an equal and moderate enhancement of sensitivity for discrimination targets shown at the saccade targets.
To disentangle these 2 interpretations, we analyzed trials in which a corrective saccade followed the execution of an averaging saccade. We reasoned that if averaging saccades resulted from a successful trial-by-trial presaccadic attentional selection of 1 of the 2 saccade targets, they should be followed by corrective saccades directed equally often towards both targets. Moreover, they should be associated with an attentional benefit at the goal of the corrective saccades. Contrary to these predictions, we observed corrective saccades in only 48.1% ± 5.8% of the averaging saccade trials. Corrective saccades were not all clearly directed towards the saccade targets (see A-B in S5 Fig), ending either in the angular sector of the most counterclockwise saccade target (ST1: 48.3% ± 3.1% of all the corrective saccades following an averaging saccade), the most clockwise saccade target (ST2: 38.3% ± 2.5%), or in between them (BTW: 11.9% ± 2.8%). They were, moreover, not equally often directed towards each of the saccade targets (ST1 versus ST2, p = 0.0288), probably reflecting a bias of our participants. As shown in C in S5 Fig, we did not find any significant benefit at the endpoint of the corrective saccades following an averaging saccade, when comparing trials in which discrimination targets were shown at the endpoint of the corrective saccade (ST1+2 correctively saccaded: d’ = 2.8 ± 0.5) to trials in which a discrimination target was shown at the same position when it was not the endpoint of the corrective saccade (ST1+2 correctively nonsaccaded: d’ = 2.5 ± 0.8, p = 0.68300). Moreover, no significant benefit could be found when the corrective saccades following an averaging saccade ended still in between the saccade targets (BTW correctively saccaded: d’ = 0.7 ± 1.1 versus BTW correctively nonsaccaded: d’ = −0.1 ± 0.6, p = 0.4698). Taken together, these results suggest that averaging saccades result from an unsuccessful or uncompleted presaccadic attentional selection among the 2 saccade targets.
Finally, we wanted to exclude the possibility that the poor discrimination performance at the endpoint of averaging saccades was a result of the rather coarse saccade direction binning used in our analysis (±7.5° of rotation around ST1, BTW, ST2, and the distractor locations). We chose this binning procedure to end up with 24 equal saccade direction bins centered on the locations at which we measured visual sensitivity. Nevertheless, one might argue that we thereby classified a substantial proportion of saccades as averaging saccades (landing within the BTW bin) despite the possibility that they were actually biased towards 1 of the saccade targets and landed in the outer areas of the bin. To validate our analysis, we analyzed visual sensitivity as a function of the saccade direction using smaller bins (±2.5°). As evident in S6 Fig, in which we contrast the data for these 2 binning procedures, the smaller binning did not systematically alter our results. Crucially, we still found low visual sensitivity at BTW even for the proportion of saccades landing precisely at the most central bin (i.e., within ±2.5° around the center of BTW).
We observed a clear oculomotor dissociation between trials in which 2 equidistant saccade targets were shown at 2 different angular distances from each other. While only accurate saccades were found for an intertarget angular distance of 90°, we observed both accurate and averaging saccades when the same targets were separated by 30°. Combined with a measure of presaccadic visual sensitivity, this dissociation allowed us to determine the influence of saccade preparation on the deployment of attention when the intended saccade goal and the saccade endpoint were spatially associated (accurate saccades) or clearly dissociated from each other (averaging saccades). Accurate saccades were associated with a strong and systematic presaccadic enhancement of visual sensitivity at the saccade endpoint when compared to the nonsaccaded locations for intertarget angular distances of both 90° and 30°. In contrast, we did not observe a presaccadic enhancement of visual sensitivity at the endpoint of averaging saccades. Rather, averaging saccades were associated with an equal deployment of attention at the 2 saccade target locations. Our corrective saccade analysis indicated that this result cannot be explained by a trial-by-trial presaccadic attentional selection of 1 of the 2 saccade targets. Overall, these effects rule out the proposal that the deployment of attention is strictly derived from the upcoming oculomotor program. Rather, they reflect a spatial dissociation between the deployment of visual attention and the averaging saccade endpoint. More specifically, these results rule out an account in which attention is precisely allocated to the saccade endpoint (alternative [a] in Introduction) or spreads over an extended region including the saccade endpoint before averaging saccades (alternative [b] in Introduction). Our data instead favor an account in which attention is equally allocated at 2 discrete saccade target locations before averaging saccades (alternative [c] in Introduction). Contrary to the idea that the activation of the oculomotor system precedes spatial attention, we propose that the oculomotor program depends on the state of attentional selection before the saccade, with averaging saccades arising from an uncompleted attentional selection process.
Findlay [28] referred to the "global effect" as the phenomenon of directing the eyes towards the center of gravity of 2 presented targets [29]. To his view, this phenomenon reflects a coarse or global processing of a visual scene before rapidly generated eye movements. His account thus predicts that in our experiment, visual sensitivity should be coarsely distributed over the 2 saccade targets as well as over their adjacent locations before the execution of averaging saccades. Our precise measure of presaccadic visual sensitivity allowed us to determine the spatial specificity of attentional deployment during saccade preparation. Contrary to the notion of a global processing (including the locations at the saccade targets and in between) before averaging saccades, we observed a precise allocation of attention limited only to the saccade targets (limited to at least approximately 2.6°, the distance between 2 of our adjacent stimuli). Therefore, before an averaging saccade, the visual system indeed seems to have precise access to the saccade target configuration, reflecting an enhancement of local rather than global visual information processing [59]. Such a discontinuous deployment of attention was also found in various tasks entailing the presentation of multiple targets [60–62]. Our results can also rule out other models of averaging saccades based solely on low-level oculomotor processing [36,37,63]. We report here that when an accurate saccade is prepared towards 1 of 2 identical saccade targets, the subsequent movement correlates with an attentional benefit at the saccade endpoint, whereas averaging saccades resulted in the absence of a selective attentional benefit at 1 of the 2 targets as well as in between them (i.e., at the saccade endpoint). In this regard, our results match with previous studies showing a reduction in the occurrence of averaging saccades when attentional selection of the saccade goal is made easier by specifying its location or its identity [40–44]. Similarly, a model relying on attentional selection could also explain why averaging saccades are less often observed in delayed saccade tasks [40,64], as they also give more time for the attentional selection to complete [43]. Early studies have often reported that averaging saccades are associated with faster saccade latencies as compared to accurate saccades [28,34]. Yet, recently, Weaver, Zoest, and Hickey [65] proposed that the spatial and temporal components of saccade programming are relatively independent from each other. They argued that attentional mechanisms can affect oculomotor behavior only when acting upon it before the onset of the movement. It might well be that our instructions to saccade as fast and as accurately as possible reduced the saccade latency range and thereby reduced potential differences between the latencies of accurate and averaging saccades. Furthermore, given that participants were engaged in a dual task, the attentional task might have slowed down saccade execution, leading to averaging saccades even at longer latencies. We propose that the type of saccade executed on a given trial was determined by the speed at which attentional selection was processed. Accordingly, accurate saccades were presumably executed whenever attentional selection of a target was readily resolved before saccade onset.
Another account of the global effect is that averaging saccades reflect a time-saving strategy [40], in which an averaging saccade followed by a correction movement allows for faster oculomotor action than a deliberately delayed accurate saccade. Given that participants saccaded accurately towards one of the targets with a similar latency as found for averaging saccades in two-thirds of the trials in our paradigm, our results speak against such a strategy. Although we observed some corrective saccades that ended nearby the saccade targets and therefore increased the accuracy of initial averaging saccades, they came with a cost of about 200 ms, rendering such strategy inefficient. Moreover, if participants would have strategically planned 2 successive saccades (an averaging saccade followed by a corrective saccade), we would expect to find attentional benefits at both saccade endpoints as reported in sequential saccade tasks [62,66]. Contrary to this prediction, we found neither an attentional enhancement at the endpoint of averaging saccades nor at the endpoint of corrective saccades compared to the positions not reached by corrective saccades.
Therefore, our results argue against earlier accounts of the global effect and propose that averaging saccades reflect a compromise between the dynamics of attentional selection and the instructions to move the eyes as fast as possible. Our proposal is based on the results of a combined measure of visual attention and averaging saccades. Similar to a previous report [56], we found an overall enhancement of visual sensitivity at the 2 saccade targets, when the data were not split depending on the saccade direction. In order to conclude on the deployment of attention before averaging saccades, however, one needs to specify visual sensitivity depending on the saccade direction. Crucially, and contrary to Van der Stigchel and de Vries [56], we indeed found an influence of the saccade direction (i.e., endpoint) on the allocation of attention when taking into account saccade direction. Within a paradigm producing both accurate and averaging saccades, we observed a presaccadic shift of attention [11,12], reflected by selectively enhanced sensitivity at the endpoint of accurate saccades. The replication of this presaccadic attention effect comes as a prerequisite to drawing conclusions on the effect of averaging saccades, for which, instead, we found no attentional benefit at the saccade endpoint. Van der Stigchel and de Vries [56] concluded that there is no attentional shift towards the endpoint of averaging saccades. However, they also reported no main effect of the saccade landing direction as well as no interaction between the saccade landing direction and the position of their attention probes when they analyzed their data as a function of the saccade endpoint. Their results are therefore inconclusive, or even speak in favor of an attentional global effect. Moreover, when we combined all trials irrespective of the saccade direction, we found a slight increase of sensitivity at the position in between the 2 potential saccade targets when they were presented transiently but not when they were presented continuously. Because Van der Stigchel and de Vries [56] used a continuous presentation of a saccade target and a distractor, their results most likely reflect a masking effect of their stimuli on the discrimination target rather than an absence of attentional modulation. Here, we clearly dissociated attention allocated to the intended saccade goal from attention allocated to the endpoint of the saccade and found no benefit at the averaging saccade endpoint. This result is theoretically consistent with the idea that attention is not restricted to the endpoint of a saccade [11,44] and provides behavioral evidence against the main hypothesis of the premotor theory of attention, which postulates that the deployment of visual attention is derived from oculomotor programming [7,8].
We illustrate our results in a theoretical framework (Fig 4), inspired by both behavioral and neurophysiological findings, linking visual attention and oculomotor programming [67]. This theoretical framework neither provides a strict model nor a computational framework. It aims at putting our results in the context of the current view on saccade programming and yielding new testable predictions. We propose that our attentional effects rely on a top-down modulation [5,19] of feature-selective areas of the visual cortex by the priority maps [68]. Initially, the onsets of the saccade targets strongly activate neurons with corresponding receptive fields within columns of the feature and priority maps (Fig 4A). Their activity will then decay until the saccade target-selection process begins. We propose that, before an accurate saccade, one of the saccade targets is selected, such that oculomotor cells centered on the saccaded location become more active in comparison to those encoding the nonsaccaded target location (Fig 4B). Because our 2 targets were physically identical, saccade target selection probably occurs within the priority maps and propagates via a top-down mechanism to the corresponding feature map columns [5,69–71]. Oculomotor cells within the priority maps are connected to the areas of the brainstem circuitry controlling the horizontal (e.g., pons and medulla) and vertical (e.g., rostral midbrain) components of an eye movement [72,73]. Given that only 1 saccade can be executed at a time, a winner-takes-all integration of the motor output [47,74,75] from the priority maps is typically assumed such that the most active population will determine the subsequent saccade vector. The exact nature of this integration is, however, beyond the scope of this study. Thus, in our framework, an accurate saccade towards the selected saccade target (i.e., the saccade target that is represented as the most active population at the level of the priority maps) is triggered by the saccade generator, and the activity state within the feature maps leads to higher sensitivity at the saccade endpoint before the eyes start to move (Fig 4B).
Following the same rationale, we propose that averaging saccades arise from an unresolved saccade target-selection process. Given the behavioral nature of our data, we can only speculate about the neural correlates of averaging saccades at the level of the priority maps in this experiment. We will, however, discuss our results in the light of 2 alternative accounts concerning the representation of averaging saccades at the level of the SC. While Edelman and Keller [54] found evidence for a bimodal distribution of collicular activity before averaging saccades at express latencies, an earlier study by Glimcher and Sparks [52] argued for an intermediate unimodal distribution in case of regular-latency averaging saccades.
Because averaging saccades were executed at regular latencies in this experiment, they might indeed have been associated with a unimodal distribution of activity at an intermediate collicular site (early oculomotor selection—Fig 4C) at saccade onset. According to this view, averaging saccades were initially reflected by 2 equally enhanced collicular populations coding for the 2 saccade targets. This bimodal distribution of activity propagates to the feature maps, leading to an equal enhancement of visual sensitivity at the 2 saccade targets. However, the initial bimodal collicular activity distribution then progresses into a unimodal distribution centered at an intermediate collicular site to subsequently allow for the execution of a single saccade. Such a scenario is in line with evidence from a recent study performing dual-site electrical stimulation in the intermediate layers of the SC [55]. If the absence of attentional deployment at the averaging saccade endpoint observed here was indeed associated with a single active population located at an intermediate site of the SC, our results would clearly refute the premotor theory of attention.
Alternatively, averaging saccades may result from a bimodal collicular activity distribution at saccade onset (late oculomotor selection—Fig 4D). In this case, the collicular sites of enhanced activity would match with the observed attentional benefits at the 2 saccade targets, and oculomotor averaging across the active collicular populations would be achieved by integration downstream of the SC. This conception could be considered compatible with a weak version of the premotor theory of attention because one could argue that the output from the SC—which is likely the last node for visuomotor transformation—is simultaneously recruited to guide attention and eye movements. However, while the final oculomotor program was averaged, attention clearly was not in this experiment. Thus, attentional and oculomotor programming are necessarily dissociable at some processing level. One possible option to account for the observed dissociation at the behavioral level is to assume that the brainstem circuitry and the attentional system deploy different algorithms to read out the collicular code.
Disentangling the 2 options discussed above (early versus late oculomotor selection) would constitute an important step in the understanding of the link between attention and action and would require simultaneous behavioral and neural recordings. In regard to the neural recording, one should, however, carefully distinguish between the different classes of neurons (fixation, visual, motor, and visuomotor), which appear to reside along a continuum with variable response properties depending on the experimental conditions [76].
According to our view, attentional selection is not completed at the onset of averaging saccades, as reflected by the equal and moderate attentional benefits at the saccade targets. This proposal is supported by electrophysiological recordings showing that averaging saccades are associated with 2 distinct peaks within the intermediate layers of the SC [53,54]. A similar, general conception of oculomotor programming was expressed by He and Kowler [44], who proposed a 2-stage process in which a single mechanism resolves attentional selection before the oculomotor program is computed at a later stage based on attentional weighting. Our results, moreover, go against a recent proposal that a merged activation within the superficial layers of the SC would feed back into the visual system [55] because this should have led to some attentional enhancement in between the saccade targets before an averaging saccade.
Our framework leads to some predictions in regard to the global effect. First, it predicts that any experimental manipulation modifying the difficulty of saccade target selection will directly impact the occurrence of averaging saccades. For example, specifying the location, the identity, or the probability of a saccade target appearing at a certain location will decrease the task difficulty, thereby increasing the speed of the attentional selection process and reducing the occurrence of averaging saccades [40–44,77]. Also, it predicts that, at a given latency, an easy saccade task should lead to fewer averaging saccades as compared to a more difficult one. Using a simple 2-saccade target task, it was shown that monkeys make averaging saccades only for express but not for normal saccade latencies [50], whereas they execute averaging saccades even for normal saccade latencies in a task rendered harder by a visual search display [49]. Similarly, Viswanathan and colleagues [78] showed that—at a saccade latency for which no consistent global effect was found with a distractor shown nearby a prosaccade target—a clear global effect was evident with the same distractor shown nearby an antisaccade target. These results are in line with our first prediction, as antisaccades are associated with a slower attentional selection [79]. Second, our framework predicts that one should not find any incremental presaccadic attentional benefit at one of the competing saccade targets before an averaging saccade, irrespective of the observed saccade latency. Future studies could directly test this prediction by measuring neuronal activity associated with the saccade targets before an averaging saccade. Third, we proposed 2 alternative explanations that could account for the observed behavioral dissociation between attention and the saccade endpoint before averaging saccades at the neuronal level. Both accounts question the validity of the premotor theory of attention in a saccade task rather than in a covert attention task [21–23].
Combining a measure of presaccadic visual sensitivity with a free-choice saccade task, we spatially dissociated attention allocated to the intended saccade goal from attention allocated to the saccade endpoint. We report here that attention is not obligatorily coupled to the endpoint of the oculomotor program, providing evidence against the strict view that oculomotor processes precede attention. Instead, we propose that saccadic responses depend on the state of attentional selection at saccade onset.
This experiment was approved by the Ethics Committee of the Faculty for Psychology and Pedagogics of the Ludwig-Maximilians-Universität München (approval number 13_b_2015) and conducted in accordance with the Declaration of Helsinki. All participants gave written informed consent.
Thirteen participants (aged 20–28, 7 females, 12 right-eye dominant, 1 author) completed the experiment for a compensation of 50€. The study was run over 2 experimental sessions (on different days) of 12 blocks of approximately 150 minutes each (including breaks). All participants except for 1 author (LW) were naive as to the purpose of the study, and all had normal or corrected to normal vision.
Participants sat in a quiet and dimly illuminated room, with their head positioned on a chin and forehead rest. The experiment was controlled by an Apple iMac computer (Cupertino, CA). Manual responses were recorded via a standard keyboard. The dominant eye’s gaze position was recorded and made available online using an EyeLink 1000 Desktop Mount (SR Research, Osgoode, Ontario, Canada) at a sampling rate of 1 kHz. The experimental software controlling the display and the response collection as well as the eye tracking were implemented in Matlab (The MathWorks, Natick, MA), using the Psychophysics [80,81] and EyeLink toolboxes [82]. Stimuli were presented at a viewing distance of 60 cm, on a 24-in Sony GDM F900 CRT screen (Tokyo, Japan) with a spatial resolution of 1,024 × 640 pixels and a vertical refresh rate of 120 Hz [83].
Each trial began with participants fixating on a central fixation target forming a black (approximately 0 cd/m2) and white (approximately 57 cd/m2) “bull’s eye” (0.4° radius) on a gray background (approximately 19.5 cd/m2). When the participant’s gaze was detected within a 2.0°-radius virtual circle centered on the fixation point for at least 200 ms, the trial began. At that time, 24 distractor streams appeared equally distributed along a 10°-radius imaginary circle centered on the fixation target (see Fig 1A). Distractor streams consisted of flickering stimuli (40 Hz), alternating every 25 ms between a vertical Gabor patch (frequency: 2.5 cpd; 100% contrast; random phase selected each stream refresh; SD of the Gaussian window: 1.1°; mean luminance: approximately 28.5 cd/m2) and a Gaussian pixel noise mask (made of approximately 0.22°-width pixels with the same Gaussian envelope as the Gabors). After a random fixation period between 300 and 600 ms (in steps of 1 screen refresh: approximately 8 ms), the fixation target switched off together with the onset of 2 saccade targets. Saccade targets, ST1 and ST2, were gray circles (approximately 39 cd/m2; 1.1° radius; 0.2° width) surrounding 2 randomly chosen streams with an intertarget angular distance of 90° or 30°. They were either presented transiently (50 ms) or continuously (until the end of the trial). When presented transiently, the saccade targets had always disappeared from the screen at the time the discrimination target appeared on the screen. When presented continuously, on the other hand, the saccade targets always temporally overlapped with the presentation of the discrimination target. Our motivation to include these 2 saccade target durations was to check for a potential masking effect of the saccade targets on the discriminability of a discrimination target. Participants were instructed to select 1 of the saccade targets by moving their eyes towards it as fast and as accurately as possible. In 96.5% of all trials, between 75 and 175 ms after the saccade target onset (a time determined to maximize discrimination target offsets in the last 200 ms before the saccade), 1 of the 24 distractor streams was replaced by a discrimination target stream in which a tilted Gabor was played (25 ms, rotated clockwise or counterclockwise by 12° relative to the vertical). The discrimination target could appear at any of the 24 distractor streams with equal probability, and subjects were explicitly informed about this fact at the beginning of the experiment. In 3.5% of all the trials, we did not present any discrimination target, in order to evaluate its influence on saccade metrics (note that all other analyses are based on the discrimination-target-present trials). At 500 ms after the saccade target onset, all stimuli disappeared, and participants were instructed to report the orientation of the discrimination target using the keyboard (right or left arrow key). Incorrect responses were followed by a negative feedback sound. On trials in which no discrimination target was shown, participants’ responses were followed by a random feedback sound.
Three participants were excluded from the analysis because their performance stayed at chance level irrespective of the position of the discrimination target. The remaining 10 participants completed between 6,972 and 7,055 trials of the saccade task. Correct fixation within a 2.0°-radius virtual circle centered on the fixation point was checked online. Trials with fixation breaks were repeated at the end of each block, together with trials during which a saccade started (i.e., crossed the virtual circle around the fixation target) within the first 50 ms or after more than 350 ms following the saccade target onset (participants repeated between 46 to 395 trials across all blocks).
In our experiment, we did not indicate the location of the discrimination target. Therefore, the perceptual task required participants to base their decision on multiple potential locations. One might therefore argue that the low sensitivity at the intermediate location BTW was observed because participants did not take the intermediate location into account as a decision variable for the perceptual task. In order to validate that our results reflect attentional effects and were not selectively biased by varying decision criteria across the different locations, we ran a control experiment, in which the position of the discrimination target was revealed by the presentation of a report cue at the end of each trial. Consequently, participants knew which location to base their discrimination judgment upon in this control experiment, which was—except for the presentation of the report cue—identical to the main experiment. Participants were instructed to give their discrimination judgment only after the report cue had appeared. The report cue (a black circle; approximately 0 cd/m2) was presented right after the offset of the distractor streams and stayed on the screen until the trial end. Overall, we tested 8 participants (4 participated in the main experiment) on an equal amount of blocks and trials as in the main experiment. S7 Fig shows the results of this control experiment in the same format as those of the main experiment (see Fig 3).
Before proceeding to the analysis of the behavioral results, we scanned offline the recorded eye-position data. Saccades were detected based on their velocity distribution [84] using a moving average over 20 subsequent eye-position samples. Saccade onset and offset were detected when the velocity exceeded or fell below the median of the moving average by 3 SDs for at least 20 ms. We included trials if a correct fixation was maintained within a 2.0° radius centered on the fixation target, if a correct saccade started at the fixation target and landed at a distance between 7° and 13° from the fixation target (±30% of the instructed saccade size), and if no blink occurred during the trial. Finally, only trials in which the discrimination target offset was included in the last 200 ms preceding the saccade onset were included in the analysis (mean ± SEM discrimination target offset relative to the saccade onset for the selected trials: −50.2 ± 1.3 ms). In total, we included 53,117 trials in the analysis (78.2% of the online-selected trials; 75.7% of all trials played) corresponding to an average of 106.0 ± 2.1 trials (115.9 ± 3.3 no-discrimination-target trials) and 105.3 ± 1.8 trials (125.0 ± 4.4 no-discrimination-target trials) per discrimination target location and participant, in the 90° and 30° conditions, respectively.
Corrective saccades were defined as the saccades directly following the offline-selected main saccades sequence and landing at a distance between 7° and 13° from the fixation target. Corrective saccades were included only if they started before the participant’s behavioral response and within the first 500 ms following the main saccade sequence. In total, we obtained 14,714 corrective saccade trials in the analysis (21.7% of the online-selected trials; 21.0% of all trials played).
Before proceeding to any behavioral analysis, we first rotated the trial configuration as to align the 2 saccade target locations (ST1: +45°, ST2: −45° and ST1: +15°, ST2: −15° for the conditions in which they were separated by 90° and 30°, respectively) symmetrically around the geometrical angle 0 (BTW). We then determined the sensitivity to discriminate the orientation of the discrimination targets (d’): d’ = z(hit rate) − z(false alarm rate). To do so, we defined a clockwise response to a clockwise discrimination target (arbitrarily) as a hit and a clockwise response to a counterclockwise discrimination target as a false alarm. Corrected performance of 99% and 1% were substituted if the observed proportion correct was equal to 100% or 0%, respectively. Performance below the chance level (50% or d’ = 0) were transformed to negative d’ values [83]. We analyzed sensitivity as a function of the discrimination position in space irrespective of the saccade landing direction (Fig 3A and 3B) but also as a function of the discrimination target position relative to the saccade landing direction (Fig 3C–3F). To do so, we redefined the position of the discrimination target relative to the saccade direction binned across 24 even, angular sectors of 15° (±7.5° from each distractor stream center angle). This binning was chosen to match with the locations at which we tested visual attention.
We initially computed single-subject means and then averaged these means across participants for each of the compared conditions to get the presented results. For all statistical comparisons, we drew (with replacement) 10,000 bootstrap samples from the original pair of compared values. We then calculated the difference of these bootstrapped samples and derived 2-tailed p-values from the distribution of these differences.
Individual raw data and averaged processed data can be found in the Open Science Framework (OSF) online repository at https://osf.io/762up/.
|
10.1371/journal.pcbi.1002930 | High Prevalence of Multistability of Rest States and Bursting in a Database of a Model Neuron | Flexibility in neuronal circuits has its roots in the dynamical richness of their neurons. Depending on their membrane properties single neurons can produce a plethora of activity regimes including silence, spiking and bursting. What is less appreciated is that these regimes can coexist with each other so that a transient stimulus can cause persistent change in the activity of a given neuron. Such multistability of the neuronal dynamics has been shown in a variety of neurons under different modulatory conditions. It can play either a functional role or present a substrate for dynamical diseases. We considered a database of an isolated leech heart interneuron model that can display silent, tonic spiking and bursting regimes. We analyzed only the cases of endogenous bursters producing functional half-center oscillators (HCOs). Using a one parameter (the leak conductance ()) bifurcation analysis, we extended the database to include silent regimes (stationary states) and systematically classified cases for the coexistence of silent and bursting regimes. We showed that different cases could exhibit two stable depolarized stationary states and two hyperpolarized stationary states in addition to various spiking and bursting regimes. We analyzed all cases of endogenous bursters and found that 18% of the cases were multistable, exhibiting coexistences of stationary states and bursting. Moreover, 91% of the cases exhibited multistability in some range of . We also explored HCOs built of multistable neuron cases with coexisting stationary states and a bursting regime. In 96% of cases analyzed, the HCOs resumed normal alternating bursting after one of the neurons was reset to a stationary state, proving themselves robust against this perturbation.
| It is often not appreciated that different activity regimes can coexist with each other in a given neuron so that a transient stimulus can cause a persistent change of activity. Such multistability of the neuronal dynamics has in fact been shown in a variety of neurons and can play either a functional role or present a substrate for neurological diseases. We explored the propensity for multistability in a database of a leech heart interneuron model, testing each case (parameter set) in a database for multistability. We found a large proportion of multistable cases, especially the coexistence of silent and bursting regimes. This was a surprising result, since these cells pace the heartbeat of the leech, and the coexistence of silence and bursting could disrupt the functional pattern, threatening the viability of the leech. Analysis of networks of mutually inhibitory multistable neurons, however, showed robustness in maintaining functional activity, suggesting that the mutually inhibitory coupling can act as a protective mechanism against failures induced by multistability.
| Recent studies of neuronal networks of identifiable neurons have shown that the same neuron type can significantly vary in membrane properties from animal to animal. The biophysical characteristics of the single neurons performing the same task can be orders-of-magnitude different [1]–[4]. This fact testifies to the great flexibility and robustness demonstrated by nervous systems. It is also captured by mathematical models analyzed with brute-force databases. With a database a population of models is considered so that those parameter sets (cases) which satisfy constraints derived from experimental data are identified as functional. Thus, following this approach, we obtained a set of cases producing functional activity although the underlying ionic current compositions were different. The apparent simplicity of the product of the brute-force database approach is moderated by complications posed by multistability.
Single neurons can produce a plethora of regimes of activity including silent, spiking and bursting regimes depending on their membrane properties. What is less appreciated is that these regimes can coexist with each other. Multistability has been reported for different neurons in a number of experimental and modeling studies [5]–[18]. It can play either a functional role or present a substrate for dynamical diseases. A comprehensive database of a neuronal model should attempt to describe all possible observable regimes of activity to assess the functionality of each case of the model.
Neuronal models exhibit a variety of activities depending on the set of parameters chosen. Parameter set databases of computational models are powerful tools used to understand how different components of neuronal dynamics interact to produce functional activity. Brute-force database approaches classify these dependencies and infer the roles played by intrinsic membrane and synaptic currents in the normal and pathological dynamics of the neuronal system of interest. These applications have proven their effectiveness in finding suitable parameter regimes that fit experimental measurements and recorded activities, and have shown that a large variety of the parameter regimes can satisfy experimental constraints [2], [19]–[25]. Here we show that it is feasible to expand such a database by appending information about stable and unstable stationary states. To obtain this information, we applied techniques from the bifurcation theory. These techniques allowed us to systematically reveal cases of multistability of bursting and stationary states.
Doloc-Mihu and Calabrese [25] have built a database for a model of an isolated leech heart interneuron and a model of the half-center oscillator consisting of a reciprocally inhibitory pair of interneurons (HCO). An individual leech heart interneuron is represented as a single isopotential electrical compartment with eight Hodgkin-Huxley type intrinsic membrane conductances [26]. In addition, the model of the HCO includes two types of inhibitory synaptic currents, spike mediated and graded [26]. By varying a set of 8 key parameters (the leak reversal potential and maximal conductances of synaptic and several membrane currents) in all combinations possible (brute-force approach) the authors of [25] systematically explored the parameter space and analyzed more than 10 million simulations (cases) of the model.
In this study, we wanted to assess how prevalent multistability of bursting and silent regimes is in a population of functional cases of the leech heart interneuron. Application of the theory of dynamical systems allows mapping the transitions between the regimes by using bifurcation diagrams [10], [17], [18], [27]. Through the use of numerical continuation of stationary states, we incorporated information about stable and unstable stationary states into the database developed in [25]. This novel approach does not resort to direct integration of differential equations, circumventing the arbitrariness in the choice of initial conditions, besides being less computationally demanding. Our methodology allowed us to systematically reveal cases of multistability in the database, and it turned out that the number of such cases is surprisingly large. Coexistence of silent regimes with functional bursting regime of leech heart interneurons poses a threat to viability of the animal since they pace the heartbeat. We also investigated how alteration of leak current and network interactions could resolve this potential problem.
Doloc-Mihu and Calabrese [25] built an extensive database of a model of a half-center oscillator (HCO). This model was developed to represent dynamics of the elemental oscillator which produces the basic rhythm of the central pattern generator controlling heartbeat of the medicinal leech [26]. It consists of two reciprocally inhibitory identical neurons. The canonical model replicates the electrical activity of the oscillator interneurons of the leech heartbeat central pattern generator (CPG) under a variety of experimental conditions [26]. Doloc-Mihu and Calabrese [25] analyzed activity regimes of the HCO model implemented in Genesis2.3, a software for simulation of neuronal dynamics [28]. Each individual leech heart interneuron (HN) was modeled as a single isopotential electrical compartment with Hodgkin and Huxley type intrinsic membrane currents. It has 8 voltage-gated currents: 1) INa- fast Na+ current, 2) IP - persistent Na+ current, 3) ICaF- rapidly inactivating low-threshold Ca++ current, 4) ICaS - slowly inactivating low-threshold Ca++ current, 5) Ih- hyperpolarization-activated cation current, 6) IK1 - delayed rectifier-like K+ current, 7) IK2 - a persistent K+ current, and 8) IKA - a fast transient K+ current. The model also included two types of inhibitory synaptic currents between the two interneurons: graded ISynG, and spike-mediated ISynS. The differential equations describing the model are given in the appendix of Hill et al. [26].
Doloc-Mihu and Calabrese [25] used a brute-force approach to explore the parameter space of the HCO model, by systematically varying a set of eight key parameters: the leak conductance and reversal potential, the maximal conductances of the spike-mediated and graded synaptic currents, and , and the maximal conductances of IP, ICaS, Ih, and IK2. Five distinct values were used for the leak reversal potential (−70 mV, −65 mV, −60 mV, −55 mV, −50 mV) and eight values were used for the maximal conductances, which were set to 0%, 25%, 50%, 75%, 100%, 125%, 150%, and 175% of the canonical values described in Hill et al. [26]. All possible combinations of the values of the varied parameters were tested. This brute-force approach generated a grid in the parameter space of the HCO model consisting of 10,321,920 cases, which had at least one of the two types of synaptic currents present (≠0 nS and/or ≠0 nS). A special set of the database represents decoupled neurons ( = 0 nS and = 0 nS). This set consisted of 163,840 cases of the isolated neuron model. All simulations of HCOs were started from the same initial conditions such that one of the two cells was firing while the other was hyperpolarized. These initial conditions were picked as a point on a bursting trajectory of the canonical HCO, obtained as the last point of an activity trace generated by integrating the canonical model for 200 s. For each case of the isolated neuron model, two trajectories were generated starting from the two different initial states described for the pair of neurons in the HCO model.
Trajectories obtained for each case were analyzed and a plethora of different regimes were found and classified as subsets of the database [25]. The classification of the trajectories is mostly based on the analysis of spike times. Spikes were detected by finding the maximum value of the membrane potential above a threshold of −10 mV. Any trace that didn't contain spikes was classified as silent (sub-threshold oscillations would be considered as a silent regime). Time series with all inter-spike intervals (ISI) smaller than 1 s were classified as tonic spiking. If at least one ISI was larger than 1 s, bursting descriptive statistics were calculated determining interburst interval, burst duration, and period of bursting. If the coefficient of variation of bursting period was higher than 5%, the trace was classified as irregular. Furthermore, Doloc-Mihu and Calabrese [25] defined HCO models as functional if both cells exhibited regular bursting activity with a small variability of the burst period (coefficient of variation of the bursting period smaller than 5%) and relative phase in the range 0.45–0.55 (balanced activity).
We focused our analysis on a subset of the cases in the database which conformed to a number of constraints on the activity. These cases were classified as the robust bursters ([25], Figure 1). When disconnected (i.e., as isolated neurons), they exhibited robust bursting activity for at least one of the two initial conditions. This bursting activity was not plateau-like or irregular. The coefficient of variation of the period was less than 5%. While connected into a HCO, they produced a functional alternating bursting pattern for at least some values of synaptic conductances. In addition, the coefficient of variation of the interburst interval was smaller than 10% (formulae Figure 1). The database subset of such isolated neuron cases (robust bursters) had a total number of 2,387.
In this study, numerical integration of the model equation implemented in C was performed with an implicit Runge-Kutta method of 5-th order, Radau IIa [29], suitable for stiff systems of ordinary differential equations. The relative error tolerances were 10−9 for each state variable, and the absolute tolerance was 10−12. Unless otherwise mentioned, the trajectories were integrated for 400 s (model time); the initial 200 s of activity were discarded to remove the transient part of the trajectory, and the last 200 s were analyzed.
For each robust burster case, we performed a single parameter bifurcation analysis of the stationary states. The leak current conductance was used as the controlling parameter. The continuation was initiated from the stable hyperpolarized stationary state obtained by direct integration of the system with a large value (20 nS). The coordinates of all codimension 1 bifurcations of stationary states (saddle-node and Andronov-Hopf bifurcations) as well as the stability of each stationary state in the curve were recorded (Figure 2). Numerical continuation of stationary states was performed with the PyDSTool Python package [30], using 4,000 control parameter steps, with minimum and maximum step sizes of 0.02 nS and 10−12 nS respectively.
We determined the ranges of for which the robust bursters exhibited robust bursting activity. Starting from the original value from the database, we iteratively incremented by 1 nS steps and integrated the model, until a transition from bursting to silence, tonic spiking, or irregular bursting was detected. Once this transition was found, the initial conditions were reset to the coordinates of the endpoint of the last accepted bursting trajectory, and the process repeated with a tenfold smaller step size to achieve a certain precision of the critical parameter value. This procedure was iterated until the precision of 10−4 nS had been achieved. Then, the whole procedure was repeated with negative steps, to determine the boundary of robust bursting towards smaller values. Thus a range of values was determined for which the robust bursting regime was observed. All analysis and auxiliary scripts were developed in-house, using Python 2.6 along with the SciPy package [31], and were run on an Intel Core i7 platform cluster. The calculation of steady state curves (numerical continuation) for 2,387 robust bursters took approximately 2.5 cpu days, while calculating the ranges of supporting bursting (direct integration) took approximately 3.5 cpu days.
The ranges of supporting the coexistence of attracting regimes were determined by checking for overlaps of the ranges supporting robust bursting activity and the ranges supporting stable stationary states, obtained from the continuation curves. Similarly, ranges supporting multistability of stationary states were determined by finding the overlaps of the corresponding ranges.
For building the database extension concerning the stationary states, we used home-written Java scripts (Java 1.5). We used the Java language for several reasons: to have scripts that can be run as is on different operating systems, to be able to query a large table (millions of records), and to be consistent with the previous work on the HCO model [25]. The database was built as two standalone database tables using MySQL (www.mysql.com). Now, for the cases considered in this article one database table has entries describing each determined stationary state by its coordinates in the state space and information on stability. This database table also has entries describing types of multistability determined. The second database table contains all the bifurcation values of the bifurcations determined in the first table. This information was recorded into the tables by using the same identification (unique) number as it had in the original HCO/single neuron database. In this way, one could query the three databases simultaneously, or could just work only with the latter.
To determine the effects of multistability on neurons in a half-center oscillator, we used the HCO model composed of two identical neurons connected through both spike mediated and graded synaptic currents [10], [26]. This is the simplest network that retains the topology and synaptic currents observed in the leech heartbeat central pattern generator, and allows direct checking of whether the presence of stable stationary states can disrupt rhythmic pattern generation. The simulation protocol consisted of setting the maximal synaptic conductances and to 150 nS and 30 nS respectively, and then of integrating the HCO model for 100 s. Then one of the cells had its initial conditions reset (not clamped) to a stable stationary state (two possibilities in the case of tristable cells), and the synaptic interactions were removed for 5 s, by setting their maximum conductances and to 0 nS. After these 5 s, the synaptic currents were restored and the system was integrated for additional 100 s. Bursting activity was analyzed separately for both the initial and final intervals.
Application of bifurcation analysis allowed us to obtain information about stationary states for each case considered from the database. These results are not sensitive to the choice of the initial conditions. We used single parameter bifurcation diagrams: the stationary state curves. It is difficult to obtain such steady state information by direct integration from random initial conditions, as a brute force gridding of the model's high dimensional state space would involve prohibitive computational power requirements. We chose the leak current conductance as the continuation parameter because it is present in vast majority of biophysically meaningful models of neurons. In our previous work [10], [17], [26], we have shown its importance in shaping multistable behavior in leech heart neuron models.
In Figure 2A, we sketch the naming convention for the intervals of the bifurcation parameter supporting stable stationary states on the branches of the stationary state curve. We classified stationary states as either hyperpolarized (hyp) or depolarized (dep) ones, if the membrane potential was negative or positive to −35 mV, respectively. For some cases we found that there could be more than one interval of each type. We numbered them in a sequence so that the hyp1 interval includes the first point of the curve. For example, the case # 288298 exhibits stable hyp2 stationary state (Figure 2B). The approach was based on the assumption that for any case of the database, making the leak conductance sufficiently large will cause the model to exhibit a stable stationary state. For the cases considered (robust bursters) it was sufficient to set = 20 nS. To establish the first point (hyp1) on the stationary state curve we integrated the model with this large value of . We used bifurcation analysis software to follow this stationary state as the parameter was decreased. In the vast majority of cases, the hyp1 state lost stability at some smaller value of via an Andronov-Hopf bifurcation (AH1). After that the stationary state was continued on the curve as an unstable one. We numbered hyp and dep states in a sequence, following the curve.
As an example, let us consider the bifurcation diagram of case #1292494 (Figure 2C,D). We continued the stable stationary state exhibited at = 20 nS. It lost stability at an Andronov-Hopf bifurcation (AH1). The bifurcation analysis software allowed us to continue the now unstable stationary state. The interval hyp1 includes only stable stationary states. Thus, it is located on the diagram to the right from AH1 bifurcation value of . As we traced the stationary state further, it disappeared at a fold bifurcation marked by LP1. At this bifurcation point two unstable stationary states met and disappeared. On the diagram, the curve turned at this bifurcation value, and proceeded to the right, towards larger values of . At the value of marked on the diagram by LP2, another fold bifurcation occurred and the curve of the stationary state made a second turn. The curve still represented unstable stationary states. Notice that, after the second Andronov-Hopf bifurcation (AH2) on the curve, the stationary state curve reached a stable interval (dep1). A stable stationary state from this interval represents a neuron with excitability block. The stable stationary states persisted even after turned negative until it coalesced with an unstable stationary state at a saddle-node (LP3) bifurcation. Tracing the stationary states after turned negative does not have biophysical meaning, but, interestingly, this unstable branch reached back to the positive values, where it coalesced with a new stable stationary state branch (dep2) at yet another fold bifurcation (LP4). This highly depolarized stationary state has never been observed in models or experimental studies of the leech heart interneurons. A simple analysis of this bifurcation diagram reveals overlap of dep1 and dep2 intervals, determining a range of multistability supporting both stable stationary states (Figure 2C).
The multistability of two excitability blocks with different values of the rest potential was demonstrated by a perturbation which triggers a switch from one regime to the other. First, by using information from the stationary state curves in Figure 2C, we chose a value from the range supporting both stationary states (denoted by a red brown bar in Figure 2C). The bifurcation curve provides the initial conditions corresponding to the dep2 stationary state. For the demonstration, we selected initial conditions in a vicinity of the actual dep2 stationary state. A switch from the dep2 to the dep1 state was induced by a brief pulse of hyperpolarizing current (Figure 3).
There was also a case that presented a range supporting stable stationary states in the middle branch (hyp2) of the continuation curve, as seen in the expanded section of the bifurcation diagram for case #288298 displayed in Figure 2B. The interval of hyp2 is delimited by two Andronov-Hopf bifurcations. If the system is integrated from initial conditions close to this region, it will give rise to subthreshold oscillations damped towards the hyp2 stationary state.
Our methodology also allowed us to systematically detect multistability, when the ranges of values corresponding to stable stationary states and to robust bursting activity overlapped. For this analysis, the cases from the database which differ only in belong to the same bifurcation diagram and are treated as one case thus giving rise to 2,223 cases of unique robust bursters instead of the original 2,387 cases. We considered five distinct regimes: the hyperpolarized stable stationary states hyp1 and hyp2, the depolarized ones dep1 and dep2, and bursting (Figure 2A), and screened each case for all possible coexistences of these regimes.
In the bifurcation diagram for case #1292494 (Figure 2C,D), we can locate three ranges supporting multistability, denoted by colored bars below the continuation curve: the red brown bar - coexistence of dep2 and dep1 stationary states, the magenta bar -dep1 and bursting, and the cyan bar -hyp1 and bursting. Multistability is also prominently present in case #861497, as can be seen in the bifurcation diagram shown in Figure 4. In the diagram, the yellow bar indicates the coexistence of three regimes, bursting, hyp1 and dep1 stationary states (tristability), while the orange bar corresponds to the coexistence of hyp1 and dep1 stationary states (bistability) and the magenta bar locates the coexistence of bursting and the dep1 stationary state (bistability).
We illustrated tristability detected in Figure 4 by integrating case #861497 from three different sets of initial conditions (Figure 5). The first set of initial conditions led to bursting (Figure 5A). The second set of initial conditions led to slow oscillations damped into the hyp1 stationary state (Figure 5B), and the third set of initial conditions produced damped spiking oscillations towards the dep1 stationary state (Figure 5C).
Applying this procedure to the whole unique robust burster subset (2,223 cases), we obtained the following five multistability scenarios upon variations of (Figure 6):
In summary, tristability was found in 48 out of 2,223 unique robust bursting cases (2%). Adding up all multistable cases, we found that 2,016 cases (91% of the unique robust bursters) exhibited multistability under some range of values. To assess the significance of these results we introduced a measure describing the sensitivity of multistability to variation of the leak conductance. For each case exhibiting multistability, we calculated the prevalence of multistability of bursting and silent regimes as the percentage of the whole range of values supporting bursting that supported multistability of the stationary states and robust bursting. A histogram of the prevalence of the multistability showed a large peak around 17% (Figure 7). The histogram also had a smaller peak at 100%. Interestingly, all cases with tristability had the prevalence of 100%. The mean value of the prevalence of the multistability was 19.6% with the standard deviation 17.7%. If we drop outliers from consideration by excluding the cases with prevalence two times larger than the mean value, we obtain the mean value corresponding to the large peak. This adjusted mean value was 15.5% with standard deviation 5.6%. This analysis shows a high prevalence of multistability of bursting and silent regimes in the dynamics of the leech heart interneuron model.
Besides determining the ranges supporting multistability for a wide range of values, our methodology allowed us to check for its presence in any case extracted from the robust burster subset of the database, just by determining whether the original value belonged to the range supporting coexistence of bursting and other regimes. This value is marked in Figures 2B,D,F and 4B as a solid vertical blue line. For the case #1292494, the original belonged to a range which supported the robust bursting as the only attractor available (indicated by the navy blue bar), thus no multistability was observed. For the case #637432 the original value of belonged to the range supporting multistability of bursting and hyp1 (the range is indicated by the cyan bar, Figure 2E,F), while in the case #288298, the original value belonged to the range supporting two regimes, bursting and hyp2 stationary state (indicated by the green bar, Figure 2B). In the case #861497 the original value belonged to a range supporting hyp1, dep1 and bursting (indicated by a yellow bar) (Figure 4). Thus, for the last two cases, multistability was already present in the robust burster subset of the database.
Applying this procedure to all robust burster cases, out of the 2,387 models 421 (18%) were multistable for the original value in the database. This result is consistent with the results of the analysis of the prevalence of multistability of bursting and silent regimes. Out of those 421 cases, 361 exhibit bistability between bursting and the hyp1 stationary state; 47 cases show bistability of bursting and the dep1 stationary state; and one case shows bistability of bursting and the hyp2 stationary state (Figure 6). Moreover, 12 cases demonstrated tristability: coexistence of bursting, the hyp1 and dep1 stationary states (Figure 4B and Figure 5).
The leech heart interneurons analyzed in the robust burster database are units of half-center oscillators (HCO) that control the leech heartbeat. The possibility of one of the cells going into a silent state presents a potentially dangerous situation for the leech. As stated previously, the analyzed cases show high prevalence of multistability of bursting and silence, with 89% of them exhibiting multistability for some range of values. Hence, we developed the protocol described in the methods section to examine whether HCOs constructed of multistable neurons regain functional alternating bursting pattern after a perturbation setting one of the cells exactly into its stable rest state.
In Figure 8, we present four examples: two with the functional pattern restored (A,B) and two with both cells caught in the depolarized rest states (C,D). Either bistable or tristable cases could be assembled into HCOs resistant to perturbation: a bistable case #637432 (Figure 8A) exhibiting coexistence of the hyp1 stationary state and bursting (Figure 2E,F) and a tristable case #861497 (Figures 4,5,8B) demonstrating coexistence of the hyp1 and dep1 stationary states and bursting. For the first case (Figure 8A), when the synapses were blocked and the bottom cell set into the hyp1 stationary state (red part of the time series), the top cell continued to burst according to its endogenous dynamics, since it was not being inhibited. When the synaptic connections were restored (green part of the time series), the upper cell was in the hyperpolarized phase of bursting, but already on its way towards initiation of the next burst. There were no synaptic currents to perturb the bottom cell towards its bursting regime, so it stayed close to the hyp1 stationary state. When the top cell began its burst (still not inhibited by the bottom cell), the synaptic currents perturbed the bottom cell out of the basin of attraction of the stationary state, restoring its bursting activity and the functionality of the system as a half-center oscillator.
For the second case (Figure 8B), where two tristable cells formed the HCO, when the synaptic coupling was restored the top cell was already firing a burst due to its own dynamics. Thus, the bottom cell was immediately pushed out of the dep1 stable stationary state into bursting, and the half-center oscillator recovered its alternating bursting pattern. Similar activity was obtained when we reset the cell to the hyp1 stationary state, as in the bistable example (Figure 8A).
The half-center configuration was not always effective in restoring the functional pattern after this perturbation (Figure 8C). In the case #820216 a single neuron is tristable. It exhibits the hyp1 and dep1 rest states and bursting regime. In this example, both cells ended in the dep1 depolarized stable rest state after the perturbation. The perturbation set the second cell to the dep1 rest state during a burst. At the beginning of the perturbation the first cell was inhibited. Having the inhibition from the second cell removed the first cell experienced a rebound and ended in the depolarized rest state too. At the end of the perturbation both cells were found in their rest states and stayed there afterwards. This scenario was also observed in the case of an asymmetric HCO (Figure 8D). We modified the HCO from the previous example by changing of one cell from 6 nS to 5.9 nS; the other cell remained at 6 nS. With this value the altered cell was bistable since the hyp1 stationary state lost stability via an Andronov-Hopf bifurcation at = 5.951 nS. This difference in the values between the leak conductance caused strong asymmetry of the bursting pattern but the perturbation still caused the switch into a dysfunctional silent regime.
To check whether the recovery of functionality was prevalent among other half-center oscillators built from multistable units, we explored the activity of HCOs constructed with the 421 cases that were multistable in the original database, as described above. Not all of those models displayed functional activity for the initial conditions given in [25], so our analysis was restricted to the 353 (84%) that were initially functional for at least one of the initial conditions. Applying the protocol to this subset, we observed that in 96% of the cases the HCOs recovered functional bursting after the perturbation, proving themselves robust against the presence of multistability involving stationary states. In the remaining 4% of cases, the HCOs were switched by perturbation into dysfunctional regimes where both cells were stuck in rest states, or at least one of them produced damped oscillations around a rest state, tonic spiking, or irregular bursting.
The same identified neurons and their synaptic connections in a circuit show a high level of variability from preparation to preparation and yet produce surprisingly similar, appropriate patterns of activity in accordance with their function [1]–[4]. This fact shifts the paradigm in modeling from searches for a canonical model, which would be perfectly tuned to experimental data, towards construction of populations of well-tuned cases of a model [24], [25], [32]. Application of a brute force database approach showed that, indeed, multiple sets of parameters produce a model exhibiting similar, functional activity, i.e., an activity with characteristics satisfying all constraints measured experimentally [20], [21]. With this study we demonstrate that multistability of oscillatory and silent regimes can be prevalent in dynamics of neuronal models and has to be taken into account.
A brute force database is a powerful tool for assessing and cataloging regimes of neuronal activity [2], [19]–[25], [33]. It sweeps through multidimensional parameter space, obtaining and categorizing the activities of the model for each parameter set, i.e., case. The goal is to describe the activity of each case with maximal completeness and store it into a database. Since the dynamics of neurons and networks can be multistable [6], [7], [10], [15]–[18] to achieve this goal one has to describe all possible regimes of activity. In the face of a large number of cases to be analyzed, which easily reaches an order of magnitude of 10 million, it is not feasible to use more than a few initial sets of state variables (initial conditions). This turns the investigation of multistability of the dynamics of each case into a computational challenge. This computational challenge raises a question: How common is multistability? If only a negligible number of models show multistability, it might appear that it is not worth the effort to investigate the database for multistability.
By using methods developed in the bifurcation theory, we extended an existing database of a leech heart interneuron, providing information on stationary states. To demonstrate a proof of these methods we analyzed a set of important cases in the database – the cases representing a single neuron which produces robust endogenous bursting when isolated and functional bursting when assembled into a half-center oscillator. For each of these cases we upgraded the database with entries describing all stationary states found, stable and unstable. The number of stationary states varied from case to case from 1 to 5. This analysis also provided new information on possible regimes of activity. We found that the model could have excitability block, represented by stable depolarized stationary states, at two different levels of membrane potential. The highly depolarized one occurs around +20 mV and is reported here for the first time.
A high percentage of the cases considered showed multistability of activity regimes. We found that 18% of the cases exhibited the coexistence of stationary states and bursting. Furthermore, considering these cases under leak conductance variation, 91% of them exhibited multistability in some range of leak conductance. These results complement studies which have established the prominent role of leak currents as a target of modulation [34]–[36]. The results show that modulation of the leak current can lead not only to transitions between silent, bursting and spiking regimes but also into and out of the multistable dynamics.
The prevalence of multistable cases in the database of a leech heart interneuron is intriguing since these neurons pace an animal's heartbeat, and multistability of an oscillatory regime and a stationary state poses a threat to the functionality of the central pattern generator, which is required to produce a persistent, steady pattern. These results raise questions concerning how a neuronal network exhibiting co-existence of functional and dysfunctional regimes can maintain normal operation in the face of naturally occurring variation in parameters and external perturbations.
We suggest that the half-center oscillator configuration could serve as a network-based protective mechanism against instances of dysfunctional multistability. By analyzing multistable models coupled via mutual inhibition, we verified that in the majority (96%) of the robust bursting cases, the functional regime is robust against tested types of perturbations of initial conditions and synaptic coupling, showing that multistability involving stationary states is not necessarily disruptive to HCO functional bursting activity. This conclusion is consistent with our previous findings that half-center configuration brings robustness to the CPG against deviations in the dynamics of the single neurons involved [10].
The main results presented here were obtained for symmetrical half-center oscillators with two identical neurons. The alternating bursting pattern was sensitive to differences in properties of the two cells, exhibiting a difference between the burst durations of the two cells (Figure 8D). It seems obvious that there exist no two identical neurons in living neuronal networks. This notion leads us to further questions concerning homeostatic mechanisms preserving functional patterns under variation of cellular properties and persistence of the dysfunctional regimes [37], [38]. In other systems of coupled endogenously bursting cells, heterogeneity of cellular properties in the network was shown to eliminate some dynamical regimes. This effect is particularly prominent in cases of electrically coupled cells like the β-cells of the islet of Langerhans [39]. In our model of the leech heartbeat HCO, this mechanism does not eliminate the high risks of the multistability with rest states (Figure 8D). The rest states observed in single cells are preserved in HCOs, if silent cells cannot interact. If a perturbation places both cells into their stable rest states without the possibility of synaptic interaction the HCO would remain silent. This consideration applies to asymmetric HCOs as well. Here we considered a perturbation less severe but one more likely to happen in a real system: only one cell was set into a stable rest state. Although most cases of the leech heartbeat HCO model restored functional pattern, 4% cases ended in a dysfunctional regime. One could observe that either a symmetrical or heterogeneous HCO can get trapped in a silent regime (Figure 8CD).
In our previous studies, the canonical models of a leech heart interneuron and of a HCO were instrumental in explaining experimental results and showing predictive power for new experiments [10], [26]. These models capture the electrical activity of an isolated interneuron and of a HCO under a variety of experimental conditions [10], [26], [40]–[47]. They were developed following the Hodgkin-Huxley formalism, and have incorporated the kinetics of currents measured in voltage-clamp experiments. These models provide a solid framework for studies of the variability of cellular properties among animals. As these models have been so valuable as a predictive tool in the past [10], [44], [47], we anticipate that the predictions made with the brute-force database of these models are credible. Our results suggest that it would be much easier to detect the multistability in the pharmacologically isolated neurons first, since the half-center motifs of the CPG circuit can restore functional bursting pattern in the face of potentially disruptive multistability. We predict that multistability of bursting and silent regimes can be demonstrated experimentally by increasing the leak conductance using a dynamic clamp technique [10]. To achieve this demonstration, we would have to solve a number of technical issues: the dynamic clamp requires intracellular recording with sharp microelectrodes. Such electrodes add significant leak. When applied to the leech heart interneurons, these electrodes transform the regime of neuronal activity from endogenously bursting into endogenously tonically spiking. In future we plan to identify modifications of ionic and leak currents (using Dynamic Clamp, e.g.), which would allow us to record endogenously bursting and silent regimes of the leech heart interneurons while recording intracellularly. Having this milestone achieved, we anticipate detecting multistability of the bursting and silent regimes within the border region between the transitions from silence into bursting and from bursting into silence. On the other hand, multistability could play unknown important roles in normal functioning of the leech heart CPG.
Multistability of single neurons could be a valuable requisite property for a variety of functions executed by neuronal networks. Multistable neurons are a substrate for memory units, toggle switches, and elements of multifunctional central pattern generators. A multifunctional central pattern generator is a neuronal network which can centrally generate more than one functional rhythmic motor pattern [48]–[56]. Studies of multifunctional CPGs in invertebrate preparations benefit from multiple advantages: the neuronal circuits can be described in terms of identified neurons and the connections between them. For example, in the medicinal leech there is a set of interneurons contributing to either a slow pattern, producing crawling, or a fast pattern, producing swimming. In the jellyfish Aglantha digitale multistability in the dynamics of single neurons is implicated in mechanisms underlying two different behaviors. This jellyfish shows two different swimming patterns: it swims slowly when feeding and rapidly when escaping [48]. These two different modes of swimming are driven by motoneurons and are explained by the fact that these motoneurons exhibit two regimes of spike propagation, a slow one based on T-type calcium current and a fast one driven by sodium current. The mechanisms of multistability are generally not well understood, however.
The mechanisms underlying multistability can be thoroughly studied by applying the theory of dynamical systems [5], [6]–[8], [11], [16], [17], [54], [57]–[63]. A key ingredient of a description of a mechanism is the identification of the unstable regime(s) which creates the boundary separating observable regimes. Among such unstable regimes leading to multistability, the most ubiquitous are saddle stationary states and saddle periodic orbits [60]. Using Morris- Lecar model as an example, Rinzel and Ermentrout showed two different mechanisms of bistability of spiking and silence [60]. In one case the stable manifold of a saddle stationary state separated a stable stationary state from limit cycle representing spiking activity, and in the other case the saddle periodic orbit separated these regimes [60]. In a mechanism described by Rinzel, an unstable periodic orbit representing unstable sub-threshold oscillations is responsible for the co-existence of periodic tonic spiking and silent regimes [5], [6]. The stable manifold of the periodic orbit separates silent and periodic spiking regimes [5]. Even a simple model of a single neuron can show a variety of different mechanisms underlying the coexistence of silence, subthreshold oscillations, spiking, and bursting in different combinations. In our previous work, we described six different types of multistability in a single simplified model [18]. Information on the bifurcations limiting the unstable regime is also important for understanding multistability in the neuronal dynamics and for description of the mechanisms of multistability [5], [6], [16]–[18], [60], [61]. Rinzel showed mechanisms of multistability which involve fold bifurcations for the stationary states and periodic orbits, homoclinic bifurcation, and torus bifurcation [61], [62]. The slow-fast decomposition techniques are instrumental in the analysis of the mechanisms supporting multistability [60]–[63]. They allow thorough description of multistability in terms of the averaged dynamics of the slow variables [2], [6], [11], [40], [41], [58]–[63]. The methods described here allow one to determine stationary states of the full model, stable and unstable, and to conduct large scale screening of a model database for new mechanisms of multistability, stationary states and other observed regimes. The brute-force database approach is a promising tool for screening cases of a model for novel mechanisms supporting multistability.
|
10.1371/journal.ppat.1004311 | Early Mucosal Sensing of SIV Infection by Paneth Cells Induces IL-1β Production and Initiates Gut Epithelial Disruption | HIV causes rapid CD4+ T cell depletion in the gut mucosa, resulting in immune deficiency and defects in the intestinal epithelial barrier. Breakdown in gut barrier integrity is linked to chronic inflammation and disease progression. However, the early effects of HIV on the gut epithelium, prior to the CD4+ T cell depletion, are not known. Further, the impact of early viral infection on mucosal responses to pathogenic and commensal microbes has not been investigated. We utilized the SIV model of AIDS to assess the earliest host-virus interactions and mechanisms of inflammation and dysfunction in the gut, prior to CD4+ T cell depletion. An intestinal loop model was used to interrogate the effects of SIV infection on gut mucosal immune sensing and response to pathogens and commensal bacteria in vivo. At 2.5 days post-SIV infection, low viral loads were detected in peripheral blood and gut mucosa without CD4+ T cell loss. However, immunohistological analysis revealed the disruption of the gut epithelium manifested by decreased expression and mislocalization of tight junction proteins. Correlating with epithelial disruption was a significant induction of IL-1β expression by Paneth cells, which were in close proximity to SIV-infected cells in the intestinal crypts. The IL-1β response preceded the induction of the antiviral interferon response. Despite the disruption of the gut epithelium, no aberrant responses to pathogenic or commensal bacteria were observed. In fact, inoculation of commensal Lactobacillus plantarum in intestinal loops led to rapid anti-inflammatory response and epithelial tight junction repair in SIV infected macaques. Thus, intestinal Paneth cells are the earliest responders to viral infection and induce gut inflammation through IL-1β signaling. Reversal of the IL-1β induced gut epithelial damage by Lactobacillus plantarum suggests synergistic host-commensal interactions during early viral infection and identify these mechanisms as potential targets for therapeutic intervention.
| The loss of intestinal CD4+ T cells in chronic HIV infection is associated with impaired immune responses to pathogens, aberrant immune activation, and defects in the gut epithelial barrier. While much is known about the pathogenesis of HIV in chronic disease, less is known about the defects that occur prior to gut CD4+ T cell depletion and whether these defects alter host interactions with pathogenic and commensal bacteria. Using a non-human primate model of HIV infection, we examined the immune and structural changes in the gastrointestinal tract 2.5 days following SIV infection. Paneth cells, in immediate proximity of SIV infected immune cells, generated a robust IL-1β response. This IL-1β response correlated with defects in epithelial tight junctions and preceded the IFN-α response, which is characteristic of innate antiviral immune responses. Despite this inflammatory environment, we did not observe defects in mucosal immune responses to pathogenic or commensal bacteria. In fact, commensal bacteria were able to dampen the IL-1β response and ameliorate tight junction defects. Our study highlights the importance of the gut epithelium in HIV infection, not just as a target of pathogenesis but also the initiator of immune responses to viral infection, which can be strongly influenced by commensal bacteria.
| Chronic inflammation and disease progression in HIV infection is attributed to dysfunction in the structure of the intestinal epithelial barrier as well as impairment of the mucosal immune response resulting in increased microbial translocation [1]–[3], dysbiosis of the gut microbiome [4]–[6], and enteric opportunistic infections [7]. Incomplete recovery of gut homeostasis, despite antiretroviral therapy, contributes to the persistence of immune activation in HIV infected patients [8]–[10]. Studies in HIV infected patients and SIV infected non-human primates have shown massive dissemination of viral infection in the gut mucosa during the primary acute stage of infection leading to severe and rapid CD4+ T cell depletion [11]–[14], which persists through all stages of infection [15], [16]. In contrast, CD4+ T cell loss is progressive in peripheral blood and lymph nodes. Loss of mucosal Th17 CD4+ T cell subset coincides with epithelial barrier disruption and is linked to increased microbial translocation and chronic immune activation [17], [18]. Although immune dysfunction following mucosal CD4+ T cell loss is well described, it is not known whether HIV can alter mucosal function and epithelial integrity prior to and independent of CD4+ T cell depletion in vivo. Further, our understanding of mucosal resident cells that are early responders to the virus and their inflammatory signaling networks is limited.
The intestinal epithelium is functionally diverse. In addition to the digestive and absorptive functions, it plays a critical role in microbial sensing and innate antimicrobial response [19]. Secretory lineages of the intestinal epithelium produce antimicrobial products such as mucins by Goblet cells and defensins and inflammatory cytokines by Paneth cells [20]. Expansion of Paneth cells during chronic SIV infection has highlighted its important role in imparting innate defense in gut mucosa during chronic SIV infection [21]. Although the Paneth cell response to microbial pathogens is well investigated, there is no information about their response to pathogens during early HIV and SIV infections and viral pathogenesis.
There is increasing evidence that viral infections can alter the host-commensal relationship [22]. HIV and SIV induced changes in the gut microenvironment may have a profound effect on the mucosal response to incoming enteric pathogens as well as local commensal bacteria. To assess the early changes in mucosal responses induced by SIV infection, use of an in vivo intestinal model is essential, as in vitro cell culture studies fail to replicate the complex cellular interactions and anaerobic microenvironment of the gut. We developed the simian ligated intestinal loop model, which most closely recapitulates the anaerobic gut microenvironment. By directly injecting bacteria into the intestinal lumen, this model facilitates the capture of the in vivo dynamics between microbes, the gut epithelium, and immune cell populations during the viral infection [17].
In the present study, we investigated the earliest effects of SIV, prior to acute mucosal CD4+ T cell depletion, on epithelial barrier integrity and mucosal immune response to pathogenic (Salmonella enterica serovar Typhimurium, S. Typhimurium) and non-pathogenic (Lactobacillus plantarum, L. plantarum) bacteria in vivo. Our findings showed that the gut epithelium was the initial target of viral pathogenesis, as evidenced by impaired expression and disorganization of epithelial tight junction proteins, which were correlated to increased expression of interleukin-1β (IL-1β). We identified Paneth cells as the dominant source of the early innate IL-1β immune response. At this time-point, no defects in mucosal immune response to either pathogenic or commensal bacteria were observed. In fact, mucosal exposure to L. plantarum rapidly dampened SIV-induced inflammation through the inhibition of the NF-κB pathway. Our study identified, for the first time, Paneth cells as an initial source of gut inflammation and IL-1β signaling during early viral infection. In addition, anti-inflammatory and epithelial repair effects of L. plantarum suggest the potential role of commensal bacteria in reversing the early effects of viral pathogenesis.
To identify the earliest targets of the pathogenic effects of SIV infection in the gut mucosa, prior to CD4+ T cell depletion, we examined rhesus macaques at 2.5 days following SIV infection (SIV+). Viral RNA was readily detected in plasma and intestinal tissue, indicating that productive viral infection was established in both mucosal and peripheral blood compartments (Figure 1A). Plasma viral loads ranged from 188–1106 RNA copies/ml (502.4±166.3 copies/ml) while viral loads in intestinal tissue ranged from 86–562 SIV copies/µg total RNA (249.7±86.44 SIV copies/µg total RNA). The localization and phenotype of SIV infected cells in intestinal tissues was determined by immunohistochemistry (IHC) (Figure 1B, Figure S1). A small number of SIV-positive cells were detected, mostly in clusters near the lower crypt regions of the intestinal mucosa, and were either CD3+ T cells or CD68+ macrophages. (Figure 1B, Figure S1). There was no detectable loss of CD4+ T cells, either in the peripheral blood (baseline uninfected: 1063±262.5 and SIV+: 952.5±322.4 cells/µl) (Figure 1C) or in the gut mucosa (percentage range 40.99–52.72%) (Figure 1D). Further, no significant changes were observed in CD4+ T cell activation, in either peripheral blood or gut mucosa, as determined by HLA-DR expression (Figure S2).
The rapid depletion of mucosal CD4+ T cells has been implicated in dysfunction of epithelial barriers and immune response during chronic HIV and SIV infections [16]. Despite the lack of detectable gut CD4+ T cell depletion at 2.5 days of SIV infection, we observed the onset of early defects in the gut epithelium by electron microscopy (EM). Epithelial tight junction structures were significantly shorter in SIV+ animals (253±10.73 nm) compared to uninfected controls (443.5±17.38 nm) (P<0.001, Mann-Whitney) (Figure 2A). Analysis of tight junction proteins by confocal microscopy confirmed that SIV infection also caused a significant reduction in the expression of tight junction proteins, ZO-1 and Claudin-1 (P = 0.027 and 0.015, respectively, Mann-Whitney) (Figure 2B–D). In addition, the distribution of ZO-1 was discontinuous in SIV+ animals; which may suggest an impairment of epithelial structure and organization since ZO-1 is an intracellular scaffolding protein integral to the organized assembly of epithelial tight junction complexes (Figure 2E). However, the reduction and restructuring of tight junction proteins during early SIV infection did not result in increased systemic microbial translocation, as determined by the levels of bacterial lipopolysaccharides (LPS) in the plasma (Figure S3).
To identify the earliest immune responses to viral infection in the gut mucosa, we performed transcriptome analysis of the intestinal tissues at 2.5 days following SIV infection using rhesus macaque specific DNA microarrays. There was no detectable increase in the expression of several known innate immune pathways or antiviral interferon (IFN) stimulated genes (ISG) in the intestinal tissues of SIV infected macaques compared to healthy controls (Figure 3A). In contrast, a striking induction of IL-1β expression and increased expression of IL-1β regulated genes was observed (Figure 3A). Evaluation of intestinal tissues by immunostaining confirmed a significant increase of IL-1β protein mean fluorescence intensity (MFI) in early SIV infection (P = 0.007) (Figure 3B).
To localize, and identify the major IL-1β expressing cells in response to SIV infection in the gut mucosa, immunohistochemical analysis was performed. The phenotype of IL-1β expressing cells was determined by the co-localization of IL-1β protein with several specific cellular markers. IL-1β was detected in the crypt epithelium as well as in lamina propria immune cells (Figure 3C). However, IL-1β expression in the crypt epithelium was approximately ten-fold higher than that observed in the lamina propria (Figure 3D). Macrophages are known to produce IL-1β following activation of the inflammasome [23]. We found that some of the IL-1β–expressing cells were positive for CD68, CD163 and CD206 expression, which served as the macrophage specific cell surface markers (Figure 3C, Figure S4). Paneth cells are differentiated, secretory cells that release defensins and antimicrobial enzymes into the intestinal lumen. Paneth cells in the intestinal tissues were identified based on their location at the base of the crypt epithelium, detectable secretory granules, absence of Ki67 cell proliferation marker (Figure 3E) and the presence of anti-microbial lysozyme protein in the granules (Figure 3F) by confocal microscopy. When we examined the localization of SIV infected cells by immunostaining for the SIV p27 antigen, we observed that the infected cells were localized in close proximity to Paneth cells in the intestinal crypts (Figure 3G). These findings suggested epithelial-immune cell interactions in the initial mucosal response to the virus.
IL-1β production in intestinal tissue during early SIV infection was negatively correlated with the expression of tight junction proteins ZO-1 (r2 = 0.874; P = 0.019) and Claudin-1 (r2 = 0.849; P = 0.026) (Figure 4A, B). These in vivo findings were validated by in vitro epithelial cell culture studies, where basolateral IL-1β treatment of Caco2 intestinal epithelial cells induced significant decreases in ZO-1 and Claudin-1 protein expression (Figure 4C), and increased permeability as measured by a decrease in trans-epithelial electrical resistance (TER) (Figure 4D). Addition of an IL-1β blocking antibody caused a significant rebound in TER (Figure 4D). To determine whether decreased TER reflected reduced barrier function, we measured 4 kDa-FITC dextran (FD4) flux, and found a significant increase in flux across the Caco2 monolayer following IL-1β treatment (Figure 4E). Together, these data provide compelling evidence that early IL-1β production following SIV infection plays a role in epithelial disruption.
We previously reported that the depletion of CD4+ Th17 cells in chronic SIV infection impairs gut mucosal immune response to pathogenic bacteria and leads to systemic microbial translocation [17]. Therefore, we sought to determine whether the onset of functional defects in the gut mucosal immune sensing and response to pathogenic (S. Typhimurium) or commensal (L. plantarum) bacteria occurred immediately upon viral exposure and prior to CD4+ T cell depletion. We utilized the ligated ileal loop model that allows for the real-time interrogation of mucosal immune responses to luminally-injected bacteria in an in vivo setting (Figure S5).
There was no systemic translocation of either S. Typhimurium or L. plantarum to peripheral sites from the lamina propria following intraluminal injection of bacteria into ileal loops of SIV infected animals (Figure S6A, B). Both live S. Typhimurium and L. plantarum could be detected in the lumen following incubation, however only pathogenic S. Typhimurium could be detected in the lamina propria (Figure S6 A–D). Further, to determine whether the presence of SIV infection in the gut mucosa was sufficient to induce aberrant mucosal immune response to S. Typhimurium, gene expression analysis was performed using DNA microarrays. A robust increase in mucosal gene expression associated with chemotaxis of neutrophils and monocytes, Th17 responses, and proinflammatory cytokines was detected in SIV-negative healthy controls in response to S. Typhimurium (Figure 5A). The transcriptional profiles in SIV+ macaques in response to S. Typhimurium were comparable to those present in SIV-negative controls with an exception for IL-6, whose expression was significantly elevated (P = 0.02) (Figure 5B). The percentages of Th17 and Th1 CD4+ T cell subsets in S. Typhimurium inoculated loops were not altered or depleted at 2.5 days post-SIV infection (Figure 5C). Thus, SIV infection did not dampen the ability of the gut immune system to mount a marked response against S. Typhimurium.
In contrast to the effects of S. Typhimurium, inoculation with L. plantarum in the intestinal loops of SIV-negative control animals had a minimal effect on the mucosal gene expression profiles (Figure 5D). However, intestinal loops from SIV+ animals had a significant change in the gene expression profiles in response to L. plantarum compared to control loops without L. plantarum. This included striking downregulation of genes involved in inflammation and cell trafficking of monocytes and neutrophils (CD86, TREM1 and CXCL8) and upregulation of genes associated with epithelial repair and tissue remodeling.
An exception to the general downregulation of chemokines in the intestinal loops following L. plantarum inoculation was the increased expression of CXCR4, CXCL12 and CCL20. The CXCR4-CXCL12 axis has been utilized by several pathogens, including HIV, for entry and invasion [24]. CCL20 is a chemokine involved in the recruitment of Th17 cells [25]. We found that there was a significant increase in IL-17 transcript levels (P = 0.03, Mann-Whitney) (Figure 5E) as well as a marked increase in the frequency of Th17 cells in intestinal loops of SIV+ animals compared to SIV-negative controls (P = 0.02, Mann-Whitney) (Figure 5F). In comparison, no changes were observed in inflammatory cytokine expression or Th1 cells between these two groups.
L. plantarum inoculation also resulted in the decreased expression of IL-1β and other genes involved in IL-1β production and signaling (Figure 6A). There was a similar reduction in IL-1β protein levels in both SIV+ and SIV-negative animals in response to L. plantarum (P = 0.087 and 0.061, respectively) (Figure 6B). The decrease in IL-1β protein expression in response to L. plantarum inoculation was inversely correlated to Claudin-1 mRNA expression, which was increased in SIV+ animals following L. plantarum inoculation (r2 = 0.781, P = 0.046) (Figure 6C). L. plantarum also significantly increased Claudin-1 protein expression in vivo in both SIV+ and SIV-negative animals (Figure 6D–F) compared to the controls (without L. plantarum) (Figure 1B–D) (P = 0.008 and 0.007, respectively). Our data suggest that L. plantarum has the potential to reverse IL-1β-associated epithelial barrier injury caused during early stages of SIV infection.
We sought to elucidate the mechanism by which IL-1β expression in the gut mucosa was reduced by L. plantarum. NF-κB is a transcription factor that regulates IL-1β expression and whose activation is characterized by its nuclear translocation [26]. NF-κB activation was detected by its nuclear localization using immunostaining. SIV infection alone caused an increase in the level of nuclear NF-κB translocation and but this increase was not significant (P = 0.193) (Figure 7A). Following L. plantarum inoculation, a trend of reduction in nuclear NF-κB protein localization was observed in both SIV+ animals and SIV-negative controls (P = 0.058 and 0.089, respectively) (Figure 7B, C). The levels of nuclear NF-κB positively correlated with the levels of IL-1β observed in all animals (r2 = 0.596, P = 0.008), regardless of the SIV infection status or presence of L. plantarum (Figure 7D). These observations suggest that L. plantarum is able to reduce IL-1β protein expression through the inhibition of the NF-κB nuclear translocation in the intestinal epithelium.
Our study, for the first time, reports that Paneth cells are early sensors of virally infected immune cells in the intestinal mucosa. Their inflammatory response is mediated through robust IL-1β signaling, with profound implications on early tissue damage. Thus, Paneth cells play a critical role in the induction of gut inflammation during the early stages of viral infection, prior to the depletion of CD4+ T cells. To our knowledge, this is the first description of IL-1β production by Paneth cells.
While the mechanism by which Paneth cells sense and respond to pathogenic bacteria is well characterized, our understanding of their response to HIV infection is limited [20]. We found that SIV infected cells were localized in close proximity of the crypt epithelium, potentially exposing Paneth cells to viral antigens or inflammatory cytokines released by the infected cells. In HIV infection, virus has been shown to induce NLRP3-inflammasome expression and IL-1β production in myeloid cells [27]. Though we cannot definitively attribute the induction of IL-1β to a specific stimulus, NLRP3 expression was increased in the gut mucosa suggesting potential involvement of an NLRP3-inflammasome mediated pathway in Paneth cells during SIV infection. Our findings highlight the need for future investigations to determine the mechanisms of Paneth cell sensing and response to viral infections and their role in the induction of host innate response to HIV.
We previously reported an increased expression of enteric defensins in Paneth cells during primary and chronic SIV infection that correlated with viral loads [21]. The loss of defensin accumulation in these cells correlated with disease progression and opportunistic infections. In the present study, we did not observe an increase in enteric defensin gene expression at 2.5 days of SIV infection. This suggests that the IL-1β response precedes the upregulation of defensin expression in Paneth cells. Similarly, we did not detect a significant increase in the expression of IFN-α or IFN stimulated genes (ISG). The type 1 IFN response is critical in the early containment of viral replication [28]. However, this involves the recruitment of plasmacytoid dendritic cells to the gut mucosa and may require higher levels of viral replication than occurs at 2.5 days following SIV infection [28]–[30]. Thus, IL-1β production by Paneth cells represents a local response to SIV infection at a time point when viral presence is low in the intestinal mucosa, and may critically impact innate immune cell subsets such as macrophages and innate lymphoid cells (ILC), which express IL-1β receptors [31], [32].
Inflammatory cytokines have been shown to disrupt epithelial barrier integrity [33]. Exposure to IL-1β increased permeability in intestinal epithelial cell cultures by decreasing epithelial tight junction protein expression [34]–[36]. Increased IL-1β expression at 2.5 days of SIV infection negatively correlations with expression of tight junction components in our study, suggesting that IL-1β initiates intestinal epithelial barrier defects. Other inflammatory cytokines, such as IFN-γ and TNF-α, have also been shown to cause disruption of epithelial cell tight junctions in vitro [37]. However, we did not detect an upregulation of IFN-γ or TNF-α expression by transcriptome analysis in vivo, suggesting that these cytokines might not contribute significantly towards intestinal epithelial changes during early infection. HIV envelope protein gp120 was shown to induce defects in epithelial tight junctions, only when added apically to epithelial cell cultures. No effects were observed when gp120 was added basolaterally [38]. This mechanism is unlikely to play a role in epithelial integrity defects observed in our study, given that the few SIV infected immune cells detected were localized to the basolateral side of the intestinal epithelium.
In chronic SIV disease, epithelial barrier disruption has been shown to lead to increased microbial translocation. However, the changes in the intestinal epithelial barrier that occur during early viral infection did not result in systemic dissemination of bacteria and microbial products. This discrepancy is likely due to the preservation of mucosal CD4+ T cells in early infection, as our previous study had shown that the depletion of Th17 cells, in chronic SIV infection, results in the increased dissemination of pathogenic S. Typhimurium [17].
The ability of the mucosal immune system to rapidly eradicate pathogens while maintaining tolerance to commensal bacteria is critical to the maintenance of intestinal homeostasis. The occurrence of aberrant host immune responses to commensal bacteria has been reported during chronic inflammatory conditions such as inflammatory bowel diseases (IBD) [39] and recently in acute Toxoplasma gondii infection [22]. Aberrant inflammatory response to commensal bacteria by peripheral monocytes of individuals with chronic HIV infection has been reported [40]. It is not known whether acute HIV infection might obfuscate the host's ability to distinguish between pathogen and commensals. Aberrant immune responses to commensal bacteria during chronic HIV infection may be attributed to increased microbial translocation [2], immune activation of antigen presenting cells [40], [41], and increased TLR2 and TLR4 expression [42], [43]. However, there have been no known studies that have interrogated gut mucosal immune responses to commensal bacteria in the context of early HIV infection. In our study, SIV infected animals had enhanced inflammatory responses to S. Typhimurium, compared to SIV-negative controls, but showed no significant changes in the response to L. plantarum. Thus, in early SIV infection, the host maintains its ability to distinguish pathogenic and commensal bacteria and mount the proper immune response.
We found that L. plantarum rapidly induced intestinal epithelial repair in SIV infected macaques through anti-inflammatory effects that were evident by decreased expression of IL-1β and inflammatory chemokines. Previous studies reported on the ability of Lactobacillus species to enhance epithelial barrier integrity via tight junction regulation [44]–[46]. Lactobacilli are known to regulate the NF-κB signaling cascade in both intestinal epithelial and antigen presenting cells [47], [48]. In the current study, significant correlations were found linking disruption of epithelial tight junctions, induction of IL-1β levels, NF-κB activation and the ability of L. plantarum to downregulate these pathologic processes. This raises a possibility of exploiting of L. plantarum to intervene the early mucosal-viral interactions that may influence gut inflammation. In addition to its anti-inflammatory effects, we observed enhanced recruitment of Th17 cells in response to L. plantarum, mostly likely due to the induction of CCL20 expression. This recruitment of Th17 cells may have a role in epithelial repair. Our findings suggest a supportive role of L. plantarum in overcoming SIV-induced gut inflammation and epithelial tight junction disruption. However, unintended consequences of an L. plantarum probiotic therapeutic adjuvant may include increased viral replication through recruitment of virus-susceptible Th17 cell targets and viral dissemination through the induction of the CXCR4-CXCL12 axis. Our findings raise an important consideration in the development of probiotic therapies for HIV infection and highlight the need for a better characterization of probiotic bacterial functions and effects [49], [50].
In summary, our study has identified the gut epithelium, specifically Paneth cells, as a site of sensing and response of viral infection and an inducer of gut inflammation through IL-1β signaling during early SIV infection. The ability of L. plantarum to modulate NF-κB activation and ameliorate epithelial defects makes it an attractive therapeutic adjuvant. These results highlight the importance of the trialogue between the epithelium, immune cells, and commensal organisms in the restoration and protection of the intestinal mucosa [51]. By further understanding the mechanisms that underlie the host/microbiota relationship in health and HIV disease, we can capitalize on their evolved synergy while identifying gaps in mucosal defenses that can be fortified through therapy.
Ten male rhesus macaques ranging from 3 to 6 years of age (tested negative for SIV, STLV, Salmonella) underwent ligated ileal loop surgery (Table S1). Five macaques were inoculated intravenously with 1000 TCID50 of SIVmac251 for 2.5 days, while five healthy, uninfected macaques served as negative controls. Animals were anesthetized and underwent ileal loop surgery as previously described [17]. Briefly, a laparotomy procedure was performed to expose the ileum before the ligation of 13 loops with an average of 5 cm in length, leaving 1-cm spacer loops in between. One ml of either stationary phase culture containing 1×109 colony-forming units (CFU) of wild type S. Typhimurium (IR715) or L. plantarum (WCFS1) was injected directly into the lumen of the ileal loops. Loops inoculated with sterile LB or MRS broth served as media controls. Each animal had three replicates of each inoculation and one loop that was not inoculated, and served as an injection control (Figure S5). All intestinal loops were collected at 5 hours (hr) following the bacterial administration. Six mm punch biopsies were collected from each intestinal loop as well as the jejunum, mesenteric lymph node, liver, and spleen for bacteriology as previously described [17]. Bacteriological data were obtained to confirm injected bacteria survival following 5 hours of incubation inside the intestinal lumen. All animals were housed at the California National Primate Research Center.
This study was carried out in strict accordance with the recommendations of the Public Health Services (PHS) Policy on Humane Care and Use of Laboratory Animals. All animals were housed at the California National Primate Research Center. All animal procedures were performed according to a protocol approved by the Institutional Animal Care and Use Committee of the University of California, Davis (protocol number: 17287). Appropriate sedatives, anesthetics and analgesics were used during handling and surgical manipulations to ensure minimal pain, suffering, and distress to animals. Furthermore, housing, feeding and environmental enrichment were in accord with recommendations of the Weatherall report. Animals were euthanized in accordance with the American Veterinary Medical Association (AVMA) Guidelines for the Euthanasia of Animals (Section 2.3)
SIV RNA loads in plasma and gut tissue samples were determined by real-time reverse transcription-PCR (RT-PCR) assay as previously described [52]. Briefly, viral RNA was isolated from 1 µg of tissue and reverse transcribed to cDNA using Supercript III. SIV gag sequences were detected using a previously published Taqman system using an Applied Biosystems ViiA 7 detection system, and data were analyzed with ViiA 7 RUO software (Applied Biosystem). The data was extrapolated against a standard curve and viral RNA copies/µg of total RNA or RNA copies/ml plasma were calculated and presented.
Six-mm biopsy punches were collected from the mesenteric lymph nodes and spleen. Biopsy punches were homogenized, serially diluted, and plated on LB + Carbenicillin (100 µg/ml) agar and MRS + Rifampicin (50 µg/ml) agar plates. To detect S. Typhimurium and L. plantarum in the lumen of ileal loops, 100 µl of luminal fluid was homogenized, serially diluted, and plated. Similarly, 1 mL of whole blood was homogenized and 100 µl was serially diluted and plated to determine the systemic dissemination of injected bacteria.
Plasma samples were diluted 1∶5 in endotoxin-free water and incubated for 15 minutes at 70°C to inactivate plasma proteins [2]. LPS was then measured using the Limulus Amebocyte Assay (Lonza) according to the manufacturer's protocol. Samples were run in triplicate and LPS levels were quantified using a standard curve after background subtraction.
Intestinal loop tissues were embedded in Araldite/Epon resin (Electron Microscopy Sciences) and 100 nm thin sections were produced using a Leica ultramicrotome. Sections were mounted on copper grids and then post-stained with 2% uranyl acetate and 1% lead citrate. Samples were imaged under a JEOL 1230 transmission electron microscope operated at 120 kV and the micrographs were digitally recorded on a TVIPS F214 CCD camera at magnification of 8000–10000×. The step size on the CCD is 14 um and the pixel size at specimen space was calculated for each micrograph according to its magnification and the post column modification in the microscope. The lengths of tight junction were measured with program GIMP, as number of pixels spanning the adhesive plasma membrane from the micrograph and then converted into nanometer by multiplying the corresponding pixel size.
Immunohistochemical analysis was performed using either frozen OCT embedded or 4% paraformaldehyde (PFA) fixed, paraffin embedded tissues. For Immunofluorescence: 5 µm sections were rehydrated and antigen retrieval (DAKO) was performed at 95°C for 30 min. Tissues were then blocked with 1% Fc blockers (Miltenyi Biotec) and 10% serum (Jackson ImmunoResearch Laboratories Inc.) for 30 min, incubated with primary antibody overnight at 4°C, followed by the secondary antibody for 1 hr at room temperature. For immunocytochemistry: 5 µm sections were fixed with cold acetone and blocked with DAKO dual endogenous enzyme block. Primary antibodies were incubated overnight at 4°C followed by development with 3,3′-diaminobenzidine (DAB). The primary antibodies were as follows: mouse monoclonal IgG1 anti-SIVmac251 Gag (clone: KK64) (NIH AIDS Reagents), rabbit polyclonal anti-human CD3 (DAKO), rabbit polyclonal anti-human ZO-1 and claudin-1 (Invitrogen), goat IgG anti-human IL-1β (R&D Systems), mouse monoclonal IgG1a anti-human CD68 (DAKO), mouse monoclonal IgG2a anti-human CD68 (Thermo Scientific), rabbit polyclonal anti-human lysozyme (DAKO), mouse monoclonal IgG1a anti-human Ki-67 (DAKO), polyclonal rabbit anti-human NF-κB p65 (Abcam), mouse anti-CD163 (clone: 10D6) (Leica Biosystems Newcastle), rabbit polyclonal anti-CD206 (Sigma-Aldrich), mouse monoclonal IgG1 anti-LTA (Santa Cruz Biotechnologies), and DifcoTM Salmonella O Antisera (BD Pharmingen). Alexa Flour 488 donkey anti-rabbit, Alex Fluor 488 goat anti-mouse IgG2a, Alexa Fluor 488 donkey anti-mouse IgG, Alexa Fluor 555 goat anti-mouse IgG1, Alexa Flour 555 donkey anti-rabbit and Alexa Fluor 555 donkey anti-goat secondary antibodies were used (Invitrogen). Isotype control was performed for IL-1β using a goat IgG UNLB (Southern Biotech). Nuclei were visualized using DAPI nucleic acid stain (Invitrogen). Images were collected using DeltaVision PersonalDV Deconvolution microscopy (Applied Precision), Leica DM IL LED microscope (Leica Microsystems) and LSM 700 microscope (Zeiss).
For the detection of epithelial tight junction proteins and NF-κB, gut tissues were imaged using Z-stack with 0.2 µm per section (25 sections total). These were performed in triplicate (3 slides with minimum of 30 µm distance separating tissue triplicates). An oil immersion, 60× objective (na = 1.42) was used with 2×2 binning during image acquisition. The sum of fluorescence intensity was calculated for the stack and mean fluorescence intensity (MFI) was determined. MFI of tight junction proteins within the epithelial regions of the tissue were quantified. For quantification of nuclear NF-κB, DAPI signal regions were selected and NF-κB signal within this region was analyzed. IL-1β localization was imaged using a 10×, 20× and 40× objectives. We utilized the 20× images to quantify the area of IL-1β within the crypts and lamina propria in the intestinal mucosa. Crypt epithelium was defined as epithelial cells most proximal to the basement membrane, as compared to protein in the lamina propria, which included regions of immune cells but not epithelial cells. Image J software (National Institute of Mental Health) was utilized for image processing and quantification.
For in vitro cell culture experiments, Caco-2 cells were treated with IL-1β for 24 hr and washed twice with PBS (Invitrogen) and fixed for 15 min with an acetone/methanol solution (1/1 v/v), permeabilized with a 1% Triton X-100 solution (Sigma-Aldrich) and blocked with 3% milk in PBS for 1 hr. Cells were then incubated with primary antibodies (Claudin-1, and ZO-1) overnight at 4°C followed by an incubation with a secondary antibody (1∶200) for 1 hr. Filters were mounted on slides with coverslip using Slow fade mounting media (Invitrogen). Slide were then analysed using a LSM 700 microscope (Zeiss) and fluorescence level was quantified using Image J software.
Total RNA was isolated utilizing the Qiagen RNeasy RNA isolation kit (Qiagen). Messenger RNA amplification, labeling, hybridization to rhesus macaque genome GeneChips, (Affymetrix) staining, and scanning were performed as described previously [52].
Assignment of genes to functional categories was performed through annotation of gene lists using the Affymetrix NetAFFX web interface, the DAVID (http://david.abcc.ncifcrf.gov/) annotation tool, and through literature-based classification by hand. Statistically over-represented (Fisher exact probability score <0.05) biological processes within sub-clusters were identified using Ingenuity Pathway Analysis (Ingenuity Systems Inc., Redwood City, CA).
Cryopreserved tissue samples were used for real-time PCR analysis. Primer-probe pairs tested, and validated to have an amplification efficiency of >95%, comparable to that of glyceraldehyde-3-phosphate dehydrogenase (GAPDH). Primers were either obtained from Applied Biosystems (Foster City, CA) or were designed, optimized, and validated for use by the Lucy Whittier Molecular Core Lab (University of California, Davis) (Table 1). Relative mRNA expression levels were calculated from normalized ΔCT (cycle threshold) values and are reported as the change. In this analysis, the CT value for the housekeeping gene (GAPDH) was subtracted from the CT value of the target gene for each sample for normalization. The target gene and GAPDH amplified with the same efficiency (data not shown). The ΔCT value for the tissue sample from the calibrator was then subtracted from the ΔCT value of the corresponding tissue sample from the experimental loop (ΔΔCT). The increase in mRNA levels in loop tissue samples of the experimental loops compared to tissue samples of baseline (calibrator) animals was then calculated as follows: increase = 2ΔΔCT; decrease = −(2(Abs (ΔΔCT))) (ViiA™ 7, Applied Biosystems).
Lamina propria lymphocytes (LPLs) were isolated from macaque tissue as described previously [53]. Following isolation, LPLs were incubated with or without 25 ng/ml PMA and 1 µg/ml ionomycin (Sigma-Aldrich) in the presence of Golgi Plug (BD Bioscience, San Jose, CA) for 6 hours. Cells were stained with Aqua LIVE/DEAD viability dye (Invitrogen) and subsequently stained for T cell phenotype markers CD3 (SP34-2, BD Bioscience), CD4 (OKT4, eBioscience), and CD8 (RPA-T8, Biolegend). Cells were then permeabilized with CytoFix/CytoPerm (BD Bioscience) and stained for IL-17 (eBio64CAP17, eBioscience), IFN-γ (4SB3, eBioscience), and TNF-α (MAb11, eBioscience). To assess T cell activation cells were ex vivo stained with HLA-DR (L243, Biolegend) in addition to the previously described markers: CD3, CD4, CD8.
Cells were analyzed on a LSRII flow cytometer (BD Bioscience). A minimum of one million events was collected per sample. Data analysis was performed using FlowJo version 8.8.6 (TreeStar).
Cellular proteins were extracted from 30–50 mg of ileal tissue using RIPA buffer (Sigma) with protease inhibitor (Roche) and homogenizing the tissue by bead agitation in a MagNA Lyser (Roche). Samples were then centrifuged and supernatant was utilized for further analysis. IL-1β protein level was measured by ELISA assay (IL-1β Quantikine, R&D Systems). Data was normalized to total tissue protein, assessed by the Bradford protein assay (Biorad).
Caco2 cells (ATCC®HTB-37) were grown in MEM media (Invitrogen) supplemented with 20% fetal bovine serum (Gemini Bioproducts), 1% Antibiotic-Antimycotic (Invitrogen). Caco2 cells only from passages 20 to 30 were used. Caco2 cells were cultured (5×105 cells/well) on permeable 0.4 µm polycarbonate filter membranes (Corning) until they reached confluence and a transepithelial electrical resistance (TER) higher than 1000. TER was measured using a Millicell-ERS voltohmeter (Millipore). Caco2 cells were then treated with IL-1β (Sigma-Aldrich) at the basolateral side of the membrane. TER measurements were performed just before and 24 hr after addition of IL-1β. FD4 (Sigma-Aldrich) flux across the Caco2 monolayer was assessed 24 h after IL-1β treatment. After withdrawing the media and washing the insert with HBSS, FD4 solution (500 µl; concentration 1 mg/ml) was added to the apical chamber and the fluorescent intensity of FD4 in the apical chamber was measured at 1 h by a fluorescent microplate reader (Chameleon V, Hidex).
For comparisons of tight junction length in SIV infection a 2-tailed, unpaired t-test with Welch's correction was performed. For data from IF, real-time PCR, and flow cytometry a two-tailed Mann Whitney test was performed. Pearson correlation was utilized to determine all coefficients of determination. Data pertaining to the changes observed due to bacterial inoculation, as compared to its media control within the same study animal, a paired two-tailed T-test was performed. P-values<0.05 were considered significant (GraphPad Software).
|
10.1371/journal.ppat.1002218 | The Fecal Viral Flora of Wild Rodents | The frequent interactions of rodents with humans make them a common source of zoonotic infections. To obtain an initial unbiased measure of the viral diversity in the enteric tract of wild rodents we sequenced partially purified, randomly amplified viral RNA and DNA in the feces of 105 wild rodents (mouse, vole, and rat) collected in California and Virginia. We identified in decreasing frequency sequences related to the mammalian viruses families Circoviridae, Picobirnaviridae, Picornaviridae, Astroviridae, Parvoviridae, Papillomaviridae, Adenoviridae, and Coronaviridae. Seventeen small circular DNA genomes containing one or two replicase genes distantly related to the Circoviridae representing several potentially new viral families were characterized. In the Picornaviridae family two new candidate genera as well as a close genetic relative of the human pathogen Aichi virus were characterized. Fragments of the first mouse sapelovirus and picobirnaviruses were identified and the first murine astrovirus genome was characterized. A mouse papillomavirus genome and fragments of a novel adenovirus and adenovirus-associated virus were also sequenced. The next largest fraction of the rodent fecal virome was related to insect viruses of the Densoviridae, Iridoviridae, Polydnaviridae, Dicistroviriade, Bromoviridae, and Virgaviridae families followed by plant virus-related sequences in the Nanoviridae, Geminiviridae, Phycodnaviridae, Secoviridae, Partitiviridae, Tymoviridae, Alphaflexiviridae, and Tombusviridae families reflecting the largely insect and plant rodent diet. Phylogenetic analyses of full and partial viral genomes therefore revealed many previously unreported viral species, genera, and families. The close genetic similarities noted between some rodent and human viruses might reflect past zoonoses. This study increases our understanding of the viral diversity in wild rodents and highlights the large number of still uncharacterized viruses in mammals.
| Rodents are the natural reservoir of numerous zoonotic viruses causing serious diseases in humans. We used an unbiased metagenomic approach to characterize the viral diversity in rodent feces. In addition to diet-derived insect and plant viruses mammalian viral sequences were abundant and diverse. Most notably, multiple new circular viral DNA families, two new picornaviridae genera, and the first murine astrovirus and picobirnaviruses were characterized. A mouse kobuvirus was a close relative to the Aichi virus human pathogen. This study significantly increases the known genetic diversity of eukaryotic viruses in rodents and provides an initial description of their enteric viromes.
| The order Rodentia is the single largest group of mammalian species accounting for 40% of all mammal species [1]. There are ca 2200 living rodent species, including mice, rats, voles, squirrels, prairie dogs, beavers, chipmunks, and guinea pigs. Many rodents have mixed diets but some eat mostly seeds or green vegetation. Rodents are known to vector more than 60 known human infectious diseases [2]. Some rodents live in close association with humans offering numerous opportunities for cross-species viral transmission through their urine, feces, or their arthropod ectoparasites such as ticks, mites, and fleas [2]–[8].
Rodents have been associated with numerous viruses including members of the Arenaviridae, Reoviridae, Togaviridae, Picornaviriade, and Flaviviridae families [2], [9]–[12]. The hantavirus pulmonary syndrome (HPS), an infection with an exceptionally high mortality first identified in the southwestern United States, was recognized as a zoonotic viral infection with Sin Nombre virus (SNV) in the Hantavirus genus in the Bunyaviridae family originating from deer mouse (Peromyscus maniculatus) [13]. Since then, HPS has been identified throughout the United States [14]–[17] with SNV responsible for most cases [18]–[23]. Deer mice captured in Montana showed an SNV antibody prevalence of approximately 11% [24]. Other members of the Hantavirus genus transmitted from rodents include Hantaan, Dobrava-Belgrade, Seoul, and Puumala viruses, causing hemorrahagic fever with renal syndrome (HFRS) worldwide [25]–[31]. HFRS was endemic in 28 of 31 provinces of China and is considered a major public health concern. Over 1,200 HFRS cases occurred in 2007 in China [32], [33]. It was reported that there were approximately 8,300 patients with HFRS in Inner Mongolia and 261 (3.14%) died during 1955–2006 [33]. Recently, Seoul virus was detected in 47 of 649 Norway rats (Rattus norvegicus) [34]. Several vole species (Microtus arvalis, Pitymys subterraneus, and M. subterraneus) have been linked with Tula virus also in the Hantavirus genus [35]–[39]. In Switzerland, acute infection with Tula virus was found in a 12-year-old boy after a rodent's bite [40]. Tick-borne encephalitis virus (TBEV) in the Flavivirus genus of the family Flaviviridae can cause fatal encephalitis in humans [41]–[43]. Several rodent species such as voles (Microtus agrestis and Myodes glareolus), field mice (Apodemus agrarius) are natural hosts of ticks that cause TBE [43], [44]. Lassa fever, an acute viral hemorrhagic fever first described in 1969 in Nigeria, is caused by Lassa virus, a member of the family Arenaviridae [45]. Its primary animal host is the Multimammate mouse (Mastomys [Praomys] natalensis) [3]. Lassa fever, endemic in West Africa, causes 30,000–500,000 cases and 5,000 deaths annually [46].
Because of their health and economic impact there is a growing awareness of emerging (and re-emerging) zoonotic infections [2], [7]. Increased interactions between rodents and humans occur when people build homes in wildlife habitat or conduct more recreational activities there [2], [47], [48]. Irruptions in density of rodent hosts, as occurred prior to the SNV infections in the southwestern US, also increase the risk of human viral exposure [49]. Viral surveys in wild and domesticated animals with extensive contacts with humans can be used to monitor for the presence of known zoonotic viruses or closely related viral species and to provide a baseline of the viruses present to help detect future changes associated with disease outbreaks. The identification of animal viruses closely related to human viruses also provides information regarding past successful zoonoses. To assist in these goals we performed an initial characterization of the fecal viromes of rodents from two locations in the US.
Viral particles in fecal samples were purified by filtration, digested with DNase and RNase enzymes to remove unprotected nucleic acids, amplified by random RT-PCR, and subjected to 454 pyrosequencing. A total of 1,441,930 sequence reads with an average of 177-bp were generated from the extracted nucleic acids in the present study and the sequences from each animal were assembled de novo into contigs of variable length. Both singlets and contigs longer than 100-bp were then classified using BLASTx and BLASTn as likely virus, phage, bacteria, or eukaryota based on the taxonomic origin in the annotation of the best-hit sequence (E score <10−5) with the GenBank non-redundant database. Fecal samples of mice, voles, and a woodrat revealed a large degree of microbial diversity. 26,846 sequence reads had best matches with viral protein, RNA or DNA sequences as shown in Figure 1. There were also ∼585,000 sequences for bacteria, 30,400 for eukaryota, and 154,000 for phage. A large proportion of the total reads (45%) did not have any significant hits to nucleotide or amino acid sequences in GenBank in agreement with viral metagenomic studies of feces from bats, turkeys, and humans [50]–[52].
The largest proportion of the rodent fecal virus-related sequences (52%) was related to mammalian viruses, with 91% of these being related to DNA viruses. Viral sequences related to single-stranded DNA viruses in the Circoviridae were abundant, comprising 90% of the mammalian DNA virus-like sequence reads. DNA viruses in the Parvovirinae subfamily, Papillomaviridae, and Adenoviridae families were also detected. RNA viral sequences were mostly related to the families Picobirnaviridae and Picornaviridae. A few RNA virus sequences (n = 23) related to the family Astroviridae were also identified. While some sequences showed >90% similarity at the amino acid level with known viruses, the majority exhibited <70% similarity. We further characterized some of these novel mammalian virus-like sequences by full or near full genome sequencing and compared them to their closest relatives by phylogenetic analyses. Partial viral genomes were similarly analyzed.
Papillomaviruses (PVs) are a highly diverse family of double-stranded circular DNA genomes ca 8-kb in size. PVs are known to infect a wide variety of mammals, as well as birds and reptiles. Some PV types cause benign or malignant epithelial tumors of the skin and mucous membranes in their natural hosts, while others are commonly present in the healthy skin of healthy humans, as well as a range of different animal species. PVs are highly species-specific and rarely transmitted between species. More than 100 human PV types have been detected, and the genomes of more than 80 have been completely sequenced [53], [54]. Only a few full genomes of PVs have been reported in non-human species.
We characterized the full-length genome of the deer mouse (Peromyscus maniculatus) PV type 1, hereafter referred to as PmPV1 (GenBank JF755418). The complete circular PmPV1 genome was 7,704-bp, with a GC content of 51%. Six distinct ORFs on the same coding strand were identified, including the early genes E6, E7, E1, and E2 and the late genes L2, and L1 (Figure 2A). Analysis of the deduced amino acid sequences revealed two characteristic zinc-binding domains (C-X2-C-X29-C-X2-C) in E6, separated by 36 amino acids. E7 also contained a zinc-binding domain and the conserved retinoblastoma tumor suppressor-binding motif (L-X-C-X-E). The C-terminal region of E1 protein had an ATP-dependent helicase motif (GPPDTGKS) and also contained a cyclin interaction RXL motif required for viral replication. The long control region (LCR) between the end of the L1 gene and the start of the E6 gene, was 469 bp. In the C-terminus of LCR, two consensus E2-binding sites (ACC-X6-GGT) were present. The TATA box (TATAAA) of the E6 promoter was located at position 7661 and a polyadenylation site (AATAAA; nt 7240) for processing of the L1 and L2 capsid mRNA transcripts were found at the N-terminus of the LCR. Taken together, many of the classic PV specific elements were identified in PmPV1.
The family Papillomaviridae currently contains at least 29 genera and mouse PVs have been found in the Pipapillomavirus and Iotapapillomavirus genera [53]. Phylogenetic analysis of the complete L1 protein was performed. PmPV1 shared the same root as the Multimammate mouse (Mastomys natalensis) PV type 1 (MnPV1) in the Iotapapillomavirus genus (Figure 2B). PmPV1 had the highest L1 similarity of 67% to the L1 of MnPV1 (Table S1). In addition, the closest amino acid similarities of other major PmPV1 proteins (E1, E2 and L2) were also to MnPV1. According to the International Committee on Taxonomy of Viruses (ICTV), different PV species share between 60% and 70% of nucleotide sequence similarity in the L1 ORF [53]. PmPV1 is therefore a proposed new PV species within the Iotapapillomavirus genus sharing 63% similarity with MnPV1.
Circular ssDNA viruses known to infect animals have the smallest viral genomes and are classified in the Circoviridae family and the unassigned genus Anellovirus. Circular ssDNA genome infecting plants belong to the Germiniviridae and Nanoviridae families. Despite very distinct host-specificities, these viruses share conserved motifs in their Replication initiator proteins (Rep) including a helicase domain [55]. Several circovirus-like genomes were also recently characterized from reclaimed water and marine environments [56], [57] and directly from a single cell of a marine protist [58]. Rep-like sequences were found in feces from 23% (12/52) of mice, 63% of voles (33/52), and 100% of cotton rat (1/1). Seventeen full circular DNA viral genomes were then sequenced using inverse PCR targeting the initial Rep sequence matches. The GenBank accession numbers are in Table S2. One genome was derived from feces of the Woodrat Neotoma cinerea, four from the mice Peromyscus maniculatus and Mus musculus and twelve from the voles Microtus pennsylvanicus. The smallest genome was 1,124-bp and the largest 3,781-bp.
These replicase-containing circular genomes could be placed into four different types of ORF organization following the classification of Rosario et al [56] (Figure 3). The type I genome had features similar to circoviruses and was characterized by a small circular DNA genome (approximately 2-kb), with the Rep protein and an unknown major ORF in opposite orientation. However, a majority of the type I genome had a stem-loop structure located in the 3′ downstream intergenic region distinct from the 5′ intergenic location seen in circoviruses, nanoviruses, and geminiviruses. The type II genome had the unique feature of encoding two separate Rep ORFs. The type III genomes contained two major ORFs in the same orientation in a manner similar to anelloviruses, a highly diverse but phylogenetically unrelated group of viruses with circular DNA genomes. The type IV genome had the largest genomes of nearly 4-kb, with up to eight ORFs. Due to their diverse genomic architectures, we preliminarily named these circular DNA genomes rodent stool-associated circular viruses (RodSCVs).
A DNA stem-loop containing a conserved nonamer is thought to have an important role in initiating rolling-cycle and a binding site for the Rep protein during circovirus replication [59]–[61]. The nonamer sequences in RodSCVs were slightly different from each other, but shared 5–7 nucleotide identity with those in circoviruses, nanoviruses, and geminiviruses [62]. Interestingly, RodSCV-M-13, M-45, and V-89 had the same nonamer sequence GGGTAATAC although they differed in genomic size and organization (Table S2).
The Rep proteins in RodSCVs possessed conserved motifs (DRYP and WWDGY) but not the DDFYGW motif typical of circoviruses. The Rep proteins in RodSCVs contained at least one well-known superfamily domain, either viral Rep, RNA helicase, or Gemini-AL1. In the type I genome, M-45 carried a homologue of the viral Rep superfamily; however, M-13, -44, V-72, -81, -84 and -97 contained both the viral Rep plus the RNA helicase superfamily domains within a single Rep ORF. By BLASTx, the Rep proteins of M-44 and 45 showed the most significant hits (E-value = 1×10−73 and 2×10−19, respectively) with circovirus-like genomes found in the environment [56] (Table S2). The Rep proteins of RodSCV-M-13, -44, V-72, -81, -84 and -97 showed the top BLAST hits (E-value = 4×10−35 and 2×10−82, respectively sharing 20–52% protein similarity) to the Rep protein of Giardia intestinalis, a major diarrhea causing parasite in humans. In the type II genome, M-53 had both the viral Rep and the RNA helicase superfamilies in two separate Rep ORFs. V-64, -86 and -91 did not contain any recognizable viral Rep superfamily protein domain, but their Rep-like proteins showed best BLASTx matches with the Rep of the plant geminiviruses. In the type III and IV genomes, RodSCV-R-15, V-76, -77 and -87 contained the Gemini-AL1 superfamily motif that is commonly found in geminiviruses (Table S2). None of the other ORFs in the all RodSCVs showed detectable similarities to known protein superfamilies or significant matches in GenBank.
In order to phylogenetically classify the RodSCVs their Rep proteins were aligned to those of circoviruses, cycloviruses, geminiviruses, nanoviruses, gyrovirus (Chicken anemia virus), environmental circovirus-like genomes, Bifidobacterium pseudocatenulatum plasmid p4M, and the protozoans Giardia intestinalis and Entamoeba histolytica. The RodSCV Rep were located on separate branches with some clustering with circovirus-like genomes derived from marine and reclaimed waters while other clustered with an integrated Rep in the genome of the parasitic protozoan Giardia intestinalis (Figure 4). Theses findings indicated that the RodSCVs likely represent several novel viral families.
The Aichi picornavirus was first identified in human cases of oyster-associated gastroenteritis in 1989 [63]–[65]. Aichi virus, a second species of human kobuvirus named salivirus/klassevirus also associated with diarrhea [66]–[70], bovine kobuvirus, porcine kobuvirus and a partial genome from a bat constitute the Kobuvirus genus in the family Picornaviridae (http://www.picornaviridae.com/kobuvirus/kobuvirus.htm). The reference Aichi virus genome is approximately 8-kb and similar to other picornaviruses contains a single large ORF that encodes a large polyprotein of 2,433 amino acids proteolytically cleaved into structural and non-structural proteins.
We found several sequence reads from two mouse (Peromyscus crinitus and P. maniculatus) fecal samples with similarities to Aichi virus. The sequence reads shared a high nucleotide similarity of 93% to each other, suggesting that the picornaviruses in these mice were from the same species. We successfully acquired the complete genome (8,171-bp, excluding the poly(A) tail) of the kobuvirus M-5 from Peromyscus crinitus mouse (GenBank JF755427), including 5′ UTR (610-bp), polyprotein ORF (2,439-aa) and 3′ UTR (244-bp). An optimal Kozak sequence, RNNAUGG (ATCATGG), was found at the beginning of the translated polyprotein. The leader (L) protein, located in the amino-terminal part of the polyprotein, showed the closest match (68%) to the L of human Aichi virus. The N-terminal zinc finger protein-binding motif C-X-H-X(6)-C-X(2)-C essential for the cytosol-dependent phosphorylation cascade found in L proteins of the Cardiovirus genus, such as Encephalomyocarditis virus and Theiler's murine encephalomyelitis virus could not be found in the mouse kobuvirus [71], [72]. In the 2C protein a highly conserved nucleotide-binding domain of the helicase (GPPGTGKS) was identified. In addition, the RNA-binding domain GLCG and the H-D-C catalytic triad were present at amino acid positions 42, 84 and 143 in the 3Cpro region as were the characteristic RdRp KDELR, YGDD and FLKR motifs in 3Dpol. Phylogenetic analysis of 3Dpol showed that the mouse kobuvirus was the closest relative to the human Aichi virus (Figure 5). This finding was also supported by pair-wise comparison in which mouse Kobuvirus M-5 had 81–84% similarities to human Aichi virus and 53–63% similarities to bovine kobuvirus at the amino acid level for the P1, P2 and P3 regions (Table S3). Genome analysis therefore indicated that the M-5 should be considered a new species and the first murine kobuvirus.
Molecular epidemiology of human Aichi virus worldwide has been performed using specific primers targeting the 3CD junction [73]–[75]. In order to understand the prevalence of Aichi-like viruses in rodents, we designed consensus primers based on 2C-3B regions conserved between rodent and human Aichi viruses and used RT-nPCR to screen other rodent fecal samples. Two additional mice (Peromyscus crinitus and P. maniculatus) were positive for Aichi virus. Their nucleotide sequences over that region (∼700-bp) were 94–95% similar to the mouse kobuvirus with which they also clustered phylogenetically (data not shown).
We detected a Sapelovirus-related sequence encoding 207 amino acids in house mouse (Mus musculus) feces M-58 (GenBank JF755421). This sequence covered about 25% of the capsid encoding P1 region and showed the best amino acid similarity (53%) to a porcine Sapelovirus. According to the ICTV, members of a picornavirus genus should share at least 40% amino acid similarity in the P1 region. Phylogenetically this sequence was also related to the Sapelovirus genus, indicating that it represents a fraction of the first reported murine Sapelovirus genome (Figure S1).
The family Picornaviridae currently consists of 12 genera, although recently characterized genomes are expected to nearly double that number (www.picornaviridae.com). In this study, we found a novel picornavirus in the feces of a canyon mouse (Peromyscus crinitus) we temporarily named Mosavirus for mouse stool associated picornavirus. We successfully acquired the nearly complete genome (6934-bp, excluding the poly(A) tail) of Mosavirus (GenBank JF973687) including a partial 5′ UTR (138-bp), complete polyprotein ORF (2235-aa) and complete 3′ UTR (88-bp). The hypothetical cleavage map of the Mosavirus polyprotein was derived from alignments with other known picornaviruses. Similar to Cardioviruses and Senecavirus, Mosavirus had a conserved cleaved site Q↓GN for a putative L protein preceding the capsid region. The P1 polypeptide (758-aa) contained the rhv-like superfamily and best BLASTx match to Theiler's encephalomyelitis virus in Cardiovirus genus, sharing 37% aa similarity. The P1 was hypothesized to be cleaved at VP4/VP2 (M↓D), VP2/VP3 (Q↓G) and VP3/VP1 (Q↓G). The VP4 was found to have the GXXX[T/S] myristylation site. The P2 polypeptide (501-aa) encoded non-structural proteins cleaved at 2A/2B (G↓P) and 2B/2C (Q↓G). BLAST showed that the P2 region had the highest aa similarity of 36% to the Saffold virus belonging to Cardiovirus genus. Similar to cardioviruses the 2A protein in Mosavirus had the conserved NPGP motif for 2A-mediated cleavage at the 2A/2B junction. The 2A protein was predicted to be 36 amino acids shorter than those of any known cardioviruses (∼150-aa). The 2C protein contained some characteristic features of picornaviruses such as the RNA-helicase superfamily, NTPase and helicase motifs. The P3 polypetide encoded proteins 3A, 3B (VPg, small genome-linked protein), 3Cpro (protease) and 3Dpol (RNA-dependent RNA polymerase). P3 with 822 amino acids in length was cleaved at 3A/3B (E↓G), 3B/3C (Q↓G) and 3C/3D (Q↓G). 3Cpro contained the peptidase C3 superfamily and conserved GXCG motifs. Some conserved YGDD, FLKR and GG[LMN]PSG motifs were identified in the 3D protein. The pair-wise amino acid sequence analysis demonstrated that the P3 polyprotein in Mosavirus shared very low aa similarity, less than 29%, to the genetically-closest picornaviruses (Table S4). Based on the 3Dpol phylogenetic tree (Figure 6) and genetic distance calculations, Mosavirus is not significantly linked to any recognized or proposed genera. According to the ICTV, the member of a picornavirus genus should share >40%, >40% and >50% amino acid similarity in P1, P2 and P3 regions respectively. The similarities over these three regions in Mosavirus were less than 40% at the amino acid level to those of any reported piconaviruses. Mosavirus is therefore proposed as a novel genus in the family Picornaviridae.
In the same house mouse feces where Mosavirus was found, we also found another picornavirus temporarily named Rosavirus for rodent stool associated picornavirus. A genome fragment of 3,956-bp was sequenced, including a partial 2C gene, the complete P3 region and a complete 3′ UTR (GenBank JF973686). The 2C segment contained the conserved NTPase motif GXXGXGKS (GGPGCGKS) and helicase motif DDLGQ typical of picornaviruses. The P3 region of Rosavirus was hypothesized to be cleaved at 3A/3B (E↓G), 3B/3C (Q↓I) and 3C/3D (Q↓G). The 3A (106-aa) and 3B (31-aa) proteins had typical lengths but did not show any detectable sequence similarity to other picornaviruses. 3B did have a conserved tyrosine at position 3 and another conserved glycine at position 5, consistent with its function as the genome-linked protein, VPg.
The 3C, with 206 amino acids, had the H-D-C catalytic triad at positions 47, 90, and 158, followed by 15 amino acids downstream of the GIH motif, similar to the substrate-binding site of chymotrypsin-like proteases. By BLAST, 3C contained the peptidase C3 superfamily and was genetically closest to turkey hepatitis virus, sharing 30% amino acid similarity. Rosavirus 3D contained conserved RdRp motifs KDELR, YGDD and FLKR. Interestingly, the 3D protein in Rosavirus has a mutated motif GAMPSG compared with conserved GG[LMN]PSG in other picornaviruses. The 3D protein in Rosavirus was most closely related to the 3D of avian turdivirus 2, sharing 44% amino acid similarity. In addition, Rosavirus had the longest reported 3′ UTR of 795-bp in length.
The P3 region in Rosavirus showed very low amino acid similarity (<31%) to the closest picornavirus (Table S4). The P3 regions of Rosavirus and Mosavirus shared only 23% similarity at the amino acid level. The 3Dpol-based phylogenetic grouping demonstrated that Rosavirus shared a monophyletic root with members of the genus Kobuvirus, turdiviruses and turkey hepatitis virus (Figure 6). A similar topology was found in a 3Cpro-based phylogenetic analysis (data not shown). Based on genetic distance criteria Rosavirus is therefore a candidate prototype for a novel genus in the family Picornaviridae.
Picobirnavirus (PBV), the only genus in the new Picobirnaviridae viral family, has a bi-segmented dsRNA genome. The large RNA segment encodes the capsid protein while the small segment encodes the viral RdRp. Originally found in the intestines of rat [76] PBV has since been found in numerous mammals [77]. In humans, ca 20% of fecal diarrheal samples in the Netherlands were positive for PBV [78]. This virus has also been reported as a causative agent of gastroenteritis in HIV-positive patients [79], [80]. Based on the small segment sequence, PBV has been classified into two genogroups, I and II [81]. However, there are few full-length sequences of this segment.
23% of feces from mice (12/52) and 19% from voles (10/52) contained PBV related sequences, the second highest prevalence and number of mammalian virus-like sequences after the Circoviridae-like reads (Figure 1). Previous reports indicated that the genogroup I was predominant in humans [78], [82]. Two fecal specimens had large number of PBV-like reads that assembled into nearly full-length RdRp for two strains, M-58 (house mouse Mus musculus) (412-aa at GenBank JF755419) and V-111 (vole Microtus pennsylvanicus) (414-aa at GenBank JF755420). Amino acid sequence alignment of nearly full-length RdRp allowed the construction of a phylogenetic tree using all other available sequences of similar length (Figure 7). M-58 and V-111 clustered with strains in genogroup I. M58 and V-111 were 63% similar at the amino acid level to each other and only 54% to 63% similar to human and bovine PBVs in genogroup I. These two picobirnaviruses therefore represent new PBV species, the first reported in mice and voles.
The family Astroviridae consists of two genera, namely, Mamastrovirus and Avastrovirus, that infect mammalian and avian species, respectively. Astrovirus (AstV), a member of the Mamastrovirus genus, is a small, non-enveloped, single-strand RNA virus that has been associated with human gastroenteritis and detected in association with other enteric pathogens [83]. The genome of astroviruses range from 6.1 to 7.3-kb and contains ORF1a, 1b, and 2, coding for serine protease, RdRp, and capsid protein, respectively. Astroviruses have been reported in fecal specimens from humans, bats, rats, pigs, cattle, and other animals [51], [84]-[88]. Astroviruses recently detected in human specimens were genetically related to animal astroviruses, indicating their possible zoonotic origins [89]–[91].
We sequenced the complete genome of an astrovirus from house mouse (Mus musculus) feces (M-52) (GenBank JF755422). The 6,519-bp genome length included a 14 bases 5′ UTR, ORF1a (848-aa), ORF1b (545-aa), ORF2 (707-aa) and 3′ UTR (332-nt, excluding the poly(A) tail). ORF1a contained a trypsin-like serine protease domain and ORF1b encoded RdRp following a -1 ribosomal frameshift induced by the presence of a heptameric “slippery sequence” AAAAAAC. The consensus promoter initiating ORF2 subgenomic RNA synthesis in the mouse astrovirus was identified as CUUUGGAGGGGUGGACCAAGAGGAGACAAUGGC (start codon in boldface). The 3′ UTR (332-nt) was longer than that of previously described astroviruses. This region contained the highly conserved 35-nt motif also found in other astroviruses from human, bat, porcine, and duck, in the avian picornavirus turdivirus as well as in a dog norovirus [92], [93]. Phylogenetic analysis based on the capsid protein showed that M-52 was linked to an astrovirus clade including bat, human, mink, and sheep astroviruses (Figure 8). BLASTx searches demonstrated that complete ORF1b and ORF2 of M-52 were most closely related to those of a bat astrovirus, sharing 65% and 38% amino acid similarities, respectively (Table S5). M-52 and rat astrovirus shared only 43% and 19% similarities in ORF1b and ORF2, respectively. M-52 is therefore the first characterized mouse astrovirus species.
Adeno-associated viruses (AAV) are small, ssDNA viruses with icosahedral capsid symmetry belonging to the Dependovirus genus of the family Parvoviridae. A characteristic of AAVs is their dependence on co-infection with adenovirus as a helper virus for their replication. AAV genomes are ca 4.7-kb in length and contain two major ORFs, and 145-bp inverted terminal repeats, which serve as the origins of replication [94]. AAV contains two major ORFs encoding the nonstructural Rep proteins and structural Cap proteins. Adenoviruses (AdV) are non-enveloped viruses composed of an icosahedral capsid and a double-stranded linear DNA genome ranging from 26 to 45-kb in length.
In our study, we found both AAV and AdV sequences in one deer mouse (Peromyscus maniculatus) fecal sample (M-6). The AAV sequence reads resulted in three separate contigs (GenBank JF755424-JF755426), covering 40% (304 amino acids) of the AAV VP7 protein showing the highest similarity to porcine AAV (61%) and AAV-5 (60%), followed by rat AAV (55%). It shared a lower similarity with mouse AAV-1 of only 43%. The phylogenic tree confirmed that AAV M-6 clustered with porcine AAV, AAV-5 and rat AAV (Figure S2A). AdV sequence reads, consisted of two separate contigs encoding 10% (154 amino acids) of the AdV hexon protein. M-6 (GenBank JF755423) exhibited the closet match of 80% to murine AdV2 with which it clustered phylogenetically (Figure 2B).
Thirty-nine percent of the rodent virome sequences were related to insect viruses (Figure 1). Detection of these viral sequences may be due to insect consumption. Insect RNA virus matches were more abundant than insect DNA viruses, making up 67% of insect viral sequences. Of note, 72% of RNA sequences were related to chronic bee paralysis virus (CBPV, an unassigned member of the Dicistroviridae family or of a distinct family in the picornavirales) known to infect honeybees [95]. Feces of infected honeybees are positive for CBPV (91). Other sequences also belonged to the family Dicistroviriade and had closest matches to newly identified viruses, including kelp fly virus and Nora virus [96], [97]. Insect DNA viruses from the viral families Iridoviridae and Polydnaviridae and the subfamily Densovirinae were found. The majority of insect virus-like sequences shared protein similarity of less <70% with annotated insect viral proteins.
About 3.4% of the rodent virome (911 sequence reads) were related to plant viruses. DNA viruses were predominant with 86% of the reads and the largest proportion was single-stranded DNA viruses from the family Nanoviridae (48%), followed by the families Geminiviridae (12%) and Phycodnaviridae (7.2%). Interestingly, 33% of the DNA viral sequences were related to fungi infecting Sclerotinia sclerotiorum hypovirulence-associated DNA virus 1 (SsHADV-1), the first reported mycovirus with a DNA genome [98]. RNA viruses accounted for only 14% of plant viral sequences and the majority was related to single-stranded RNA viruses in the family Secoviridae (64%), followed by the families Tymoviridae, Alphaflexiviridae and Tombusviridae. Double-stranded RNA viruses were responsible for 28% of the plant viral RNA sequences, belonging to the family Partitiviridae.
Animal and human viral discovery has long been focused on pathogenic infections and viruses that could be readily grown in cell cultures and cause visible cytopathic effects. Viral metagenomics is a recent approach to analyzing mixtures of viral nucleic acids enriched directly from a variety of sources including animals, plants, protozoa, bacteria, archaea, and diverse environments without a prerequisite in vitro or in vivo amplification [20], [50]–[52], [99]–[110]. Viral metagenomics has been successfully employed to identify both commensal viruses and viral pathogens and has the potential to detect all viruses recognizable through sequence similarity searches [111]. We describe here the fecal viral flora in several species of wild rodents. Sequences closely and distantly related to known viral sequences were identified as well as many sequences of unknown taxonomic origin [51], [99], [101], [108], [112]–[114]. A fraction of these currently unclassifiable sequences may be derived from still genetically uncharacterized viral families refractory to nucleotide or protein sequence similarity based searches. Multiple known DNA and RNA viral families infecting plants, insects, and mammals were also detected. Plant and insect viral sequences likely reflect the omnivorous diet of these rodents.
The number of Rep-containing circular DNA genomes recently found in various animal and environmental viral surveys has greatly expanded the genetic diversity in this group of viruses [56], [62], [115]–[119]. Some of the Rep proteins of the small circular DNA genomes characterized here clustered with viruses in marine environment or in protozoan genomes. Type II RodSCVs contained two separate Rep ORFs in their genomes, separately encoding the viral Rep and RNA helicase with each gene individually smaller than those in single Rep containing genomes. The many novel circular DNA genomes seen here indicate the likely existence of several previously unknown viral families. The closest relatives of some Rep sequences found in rodent feces are those integrated in the Giardia intestinalis genome, possibly reflecting their replication in protozoans in rodent guts.
We report on the first identification of a kobuvirus in mice (Peromyscus crinitus and P. maniculatus). Since its genetic characterization in 1998 in Japan [65], Aichi virus, the archetypical kobuvirus, has been associated with acute gastroenteritis worldwide [74], [120], [121]. Other kobuvirus species have been reported in pigs and cows as well as in humans and associated with diarrhea [68]–[70], [121]. Our phylogenetic analysis revealed that the mouse kobuvirus clustered with human Aichi virus. A recent study describes a similar close relationship between a canine kobuvirus and Aichi virus [122]. This close relationship between human, mouse, and canine kobuviruses suggests past zoonotic transmission of kobuviruses between these hosts or their recent ancestors followed by independent evolution leading to the close but distinct species seen today.
Sapelovirus is another genus in the family Picornaviridae and currently includes 3 species (www.picornaviridae.com). While phylogenetically related to the enteroviruses, these viruses have a different type of internal ribosome entry site and a leader protein (www.picornaviridae.com). A Sapelovirus-like genome fragment was identified in house mouse feces. Phylogenetically, this fragment clustered with the other Sapeloviruses. Mice therefore also appear to harbor Sapeloviruses, expanding the known host range of this group of viruses.
Two potential new picornaviridae genera we provisionally labeled Mosavirus and Rosavirus were genetically characterized. Only distantly related to the kobuvirus genus, the Rosavirus is also part of a larger clade containing several avian picornaviruses, including the recently described avian turdiviruses and turkey hepatitis virus [93], [123]. The recent characterization of numerous deep-branched members of the picornaviridae family (likely new genera) from mammals, birds, reptiles, and fish indicates this viral family is still greatly under sampled and likely to rapidly expand as more potential host species are analyzed (www.picornaviridae.com). A second potential new genus distantly related to the cardioviruses was also characterized. Until recently cardioviruses consisted of the Theiloviruses and Encephalomyocarditis viruses infecting largely rodents and the Saffold virus group infecting humans [124]–[126]. The Mosavirus also exhibited the large genetic distance relative to other picornaviruses required to tentatively qualify Mosavirus as a founding member of a new Picornaviridae genus.
Using consensus PCR primers PBV has been detected in the feces of mammals, birds, and snakes with or without diarrhea [127]–[132]. Infection was common in wild rodents with 22/105 animals shedding PBV-like nucleic acids, the second highest group of mammalian virus-like sequences after the highly diverse small circular DNA genomes (Figure 1). PBV therefore represent the largest quantity of viral nucleic acids released by these rodents into the environment. Two nearly complete PBV RdRp sequences clustered together with a larger clade of human PBV cluster I sequences possibly reflecting past rodent-human transmission. Complete PBV RdRp sequences from more animal species will be required to further test this hypothesis.
The number of mammalian astroviruses species has recently undergone a rapid expansion [85], [88]–[91], [133], [134]. The first astrovirus genome from a mouse clustered with a clade containing mink, sheep, and recently identified human astroviruses [89], [91]. The detection of multiple unrelated astrovirus species within some host species (human, pigs, and sea lions) may reflect frequent cross-species transmission rather than diversification from a common source.
AAV infects humans and other primate species [16], [135], [136]. Although over 80% of humans are seropositive, AAV has not been associated with any diseases in humans [137]. Members of the family Adenoviridae infect various species of vertebrates, including humans [51], [138]–[140]. The majority of AdV infection cause upper respiratory diseases; however, AdV also causes other symptoms, such as acute gastroenteritis [141]. The small AAV and AdV fragments identified here are too limited for definitive classification but provisionally appear to represent a fourth species of murine adenovirus [142].
Rodents are known reservoirs of numerous viruses capable of causing human diseases [2]. The characterization of the viromes of animal species with frequent contacts with humans can provide baseline viral content to more quickly identify the possible sources of future zoonotic infections and therefore assist in their control. Changes in the baseline virome of various animal may also assist in identifying changes associated with future population crashes. The addition of annotated viral genomes to public databases will also facilitate their inclusion on microarrays as well as assist in their precise identification after high-throughput sequencing [143], [144]. Monitoring for rodent viruses in humans will require extensive geographical sampling in persons with high exposure such as hunters, hikers, and farmers, and in regions of high biodiversity. The large increase in rodent viral species and higher level taxa revealed here by a limited species, geographic, and individual sampling reflects the large extent of mammalian virus diversity awaiting characterization.
Fecal specimens were collected from 52 mice, 52 voles and one Woodrat and stored at −80°C. Rodents were humanely captured and released according to the international guidelines of the American Society of Mammalogists (www.mammalsociety.org). Four mouse species Peromyscus crinitus (Canyon mouse), P. maniculatus (Deer mouse), P. truei (Pinyon mouse), and P. boylii (Brush mouse) were trapped in three different California counties, Siskiyou, El Dorado and Lassen, in May and June 2010 (Table 1). Two species of voles were analyzed: one Microtus longicaudus (Long-tailed vole) was caught in El Dorado Co., CA in June 2010 and 51 Microtus pennsylvanicus (Meadow vole) plus 20 Mus musculus mice were caught in an old field habitat in southeastern Virginia during April-November 2008. A Neotoma cinerea (Woodrat) was trapped in Siskiyou Co., CA in May 2010. The feces collected from the traps were used in this study. Sample collection was exempted from review by the IACUC committees of the CDPH and of the Old Dominion University.
Fecal samples were resuspended in Hanks buffered saline solution (Gibco BRL) and vortexed. The samples were clarified by 15,000 x g centrifugation for 10 min. A total of 200 µl of supernatants was filtered through a 0.45-µm filter (Millipore) to remove bacterium-sized particles. The filtrate was then treated with a cocktail of DNases (Turbo DNase from Ambion, Baseline-ZERO from Epicentre, and Benzonase from Novagen) and with RNase (Fermentas) to digest unprotected nucleic acids [52]. Nucleic acids protected from nuclease digestion within viral capsids were then extracted using QIAamp spin-columns according to the manufacturer's instructions (Qiagen).
cDNA synthesis was performed as described previously [52]. Briefly, RNA only and DNA plus RNA virus sequence-independent amplifications were separately performed and then combined before sequencing. For RNA virus-only amplification, 10 µl of the extracted nucleic acid was treated with DNase (Ambion) and was used as a template to synthesize cDNA, using SuperScript III reverse transcriptase (Invitrogen) and a primer containing an arbitrary set sequence followed by a randomized eight nucleotides at the primer's 3′ end. For the RNA plus DNA virus amplification, the DNase step prior to RT was excluded. Following reverse transcription followed by heat denaturation and re-annealing of the primer a single round of primer extension DNA synthesis was performed using Klenow fragment polymerase (New England Biolabs). PCR amplification was then performed using primers consisting of only the set portion of the random primer. To increase sampling of the viral nucleic acids, the random PCR amplifications were performed in duplicate resulting in four PCR products (2 from viral RNA-only and 2 from viral RNA plus DNA). The four PCR products were pooled and purified using the QIAquick Purification Kit (Qiagen). The purified DNA level was determined by Nanodrop (Thermo Scientific). Equal amounts of amplified DNA from up to 40 different fecal samples (using different primers subsequently recognizable by their set sequences) were combined into larger pools to generate 3 libraries. A total of 120 µg of DNA from each library was run on a 2% agarose gel, yielding a DNA smear. DNA ranging in apparent size from 500 to 1,000-bp was cut from the gel and purified using the QIAquick Gel Extraction Kit (Qiagen). The extremities of the PCR products DNA were then polished using T4 polynucleotide kinase. The Roche/454 adaptors were then ligated, and small DNA fragments removed, according to the manufacturer's protocol (GS FLX Titanium General Library Preparation Kit, Roche).
The set sequences on the different random PCR primers were used to assign sequence reads to the corresponding fecal samples. Sequence reads were trimmed of their set primer sequences and adjacent eight nucleotides corresponding to the randomized part of the primers. Trimmed sequences from each bin were then de novo assembled into contigs using Sequencher (Gene Codes), with a criterion of at least 95% identity over 35-bp to merge two fragments. The assembled sequence contigs and singlet sequences greater than 100-bp were compared to the GenBank nonredundant nucleotide and protein databases using BLASTn and BLASTx, respectively. Based on BLAST output, sequences were classified as viruses, phage, bacteria, and eukaryota based on the taxonomic origin of the best-hit (lowest E score) sequence match. An E value of 10−5 was the cutoff value for significant hits. Sequences whose best alignment E value was >10−5 were deemed unclassifiable.
Complete circular DNA viral genomes were amplified using inverse PCR (iPCR) with specific primers designed from 454 derived short-sequence fragments. iPCR amplicons were then directly sequenced by primer walking. PCR reactions contained 2.5 U of LA Taq polymerase (Takara) in 2.5 µl of 10X LA PCR buffer (Takara), 4 µl of 2.5 mM dNTP (Takara), 2.5 µl of forward and reverse primers (10 pmol/µl), and 2.5 µl of nucleic acids (for first round) or 1 µl of the first-round PCR product (for second PCR round) as a template in a 25 µl total volume. PCR was performed at 94°C for 1 min, followed by 30 cycles of 98°C for 10 s, 68°C for 4–10 min depending on the sizes of the expected amplicon, and a final extension at 72°C for 10 in, and then held at 4°C.
All primers are described in Table S6. Sequence reads showing significant BLASTn or BLASTx hits to Aichi virus were linked together using RT-PCR. The 5′ and 3′ rapid amplification of cDNA end (5′ and 3′ RACE) was used to acquire the 5′ and 3′ extremities of the Aichi-like virus genome [89], [145]–[147]. For the complete genome of astrovirus, pairs of specific reverse (Ast-R1 and Ast-R2) and forward primers (Ast-F1 and Ast-F2) designed from an initial astrovirus-like-sequence of 414-bp were used in 5′ and 3′ RACE [89], [145]–[147] to amplify ∼1.5-kb and 5-kb PCR products, respectively. In a mouse (Peromyscus crinitus) fecal specimen two small picornavirus-like-genome fragments (300-bp and 251-bp, respectively) were detected. PCR failed to link these two fragments together, suggesting that they belonged to two different viruses. For the complete genome of the Mosavirus, specific reverse (Mosa-R1 and Mosa-R2) and forward primers (Mosa-F1 and Mosa-F2) were used in 5′ and 3′ RACE [89], [145]-[147] to amplify ∼1.5-kb and 6-kb PCR products, respectively. For the Rosavirus, a pair of specific forward primers (Rosa-F1 and Rosa-F2) were used in 3′ RACE [89], [146], [147] to amplify ∼4-kb amplicon. 5′ RACE was not successful. To investigate the prevalence of Aichi-like virus, consensus primers were used for PCR screening designed on a nucleotide alignment of the 2C-3B region of all human Aichi virus genotypes available in GenBank and the mouse Aichi-like virus strain characterized here. For the RT reaction, 10 µl of extracted RNA was added to 10 µl of RT mixture containing 4 µl of 5X First-Strand buffer (Invitrogen), 1 µl of 10 mM dNTP (Fermentas), 1 µl of random primer, 1 µl of SuperScript III Reverse Transriptase (Invitrogen), 1 µl of RNase inhibitor (Fermentas), and 1 µl of DEPC-treated water. The RT reaction mixture was incubated at 25°C for 5 min, 50°C for 60 min, 70°C for 15 min to inactivate the enzyme, and then held at 4°C. For the first round PCR, 2.5 µl of cDNA template was mixed with 2.5 µl of 10X ThermoPol Reaction buffer (New England Biolabs), 0.5 µl of 10 mM dNTP (Fermentas), 2.5 µl of each primer (10 pmol/µl) (Ai-Deg-F1 and Ai-Deg-R1), targeting the Achi-like virus, 0.4 µl of Taq DNA Polymerase (New England Biolabs). DEPC-treated water was added up to a 25 µl total volume. The PCR condition was as follows: denaturation at 95°C for 5 min, 35 cycles of 95°C for 30 s, 63°C for 30 s and 72°C for 1 min, a final extension at 72°C for 10 min, and then held at 4°C. The second round of amplification was performed using the same conditions except that the annealing temperature was 60°C, and inner primers Ai-(Deg-F2 and Ai-Deg-R2). The second round PCR amplification resulted in the amplicon size of 735-bp.
Reference viral sequences from different viral families were obtained from GenBank. Sequence analysis was performed using CLUSTAL X with the default settings. Sequences were trimmed to match the genomic regions of the viral sequences obtained in the study. A phylogenetic tree with 100 bootstrap resamples of the alignment data sets was generated using the neighbor-joining method [148]. The genetic distance was calculated using Kimura's two-parameter method (PHYLIP) [149]. Sequence identity matrix was measured using BioEdit. GenBank accession numbers of the viral sequences used in the phylogenetic analyses were shown in the trees. Putative ORFs in the genome were predicted by NCBI ORF finder. The circular genome architectures were predicted using Vector NTI 11.5 Advance (Invitrogen) with the following conditions: minimum ORF size of 100 codons, start codons ATG and GTG, stop codons TAA, TGA and TAG. To identify stem-loop structures, nucleotide sequences were analyzed with Mfold. Full and partial genome sequences are at GenBank accession numbers JF755401-JF755427, and JF973686-JF973687. The 454 pyrosequencing data is in the short read archive at SRA030869.
|
10.1371/journal.pntd.0003098 | Non-Participation during Azithromycin Mass Treatment for Trachoma in The Gambia: Heterogeneity and Risk Factors | There is concern that untreated individuals in mass drug administration (MDA) programs for neglected tropical diseases can reduce the impact of elimination efforts by maintaining a source of transmission and re-infection.
Treatment receipt was recorded against the community census during three MDAs with azithromycin for trachoma in The Gambia, a hypo-endemic setting. Predictors of non-participation were investigated in 1–9 year olds using random effects logistic regression of cross-sectional data for each MDA. Two types of non-participators were identified: present during MDA but not treated (PNT) and eligible for treatment but absent during MDA (EBA). PNT and EBA children were compared to treated children separately. Multivariable models were developed using baseline data and validated using year one and two data, with a priori adjustment for previous treatment status. Analyses included approximately 10000 children at baseline and 5000 children subsequently. There was strong evidence of spatial heterogeneity, and persistent non-participation within households and individuals. By year two, non-participation increased significantly to 10.4% overall from 6.2% at baseline, with more, smaller geographical clusters of non-participating households. Multivariable models suggested household level predictors of non-participation (increased time to water and household head non-participation for both PNT and EBA; increased household size for PNT status only; non-inclusion in a previous trachoma examination survey and younger age for EBA only). Enhanced coverage efforts did not decrease non-participation. Few infected children were detected at year three and only one infected child was EBA previously. Infected children were in communities close to untreated endemic areas with higher rates of EBA non-participation during MDA.
In hypo-endemic settings, with good coverage and no association between non-participation and infection, efforts to improve participation during MDA may not be required. Further research could investigate spatial hotspots of infection and non-participation in other low and medium prevalence settings before allocating resources to increase participation.
| As the target year for Global Elimination of Trachoma (GET2020) approaches, the scale up of mass drug administration (MDA) with azithromycin will lead to more endemic areas becoming low prevalence settings. In such areas, identification of those at highest risk of Chlamydia trachomatis infection and at highest risk of non-participation during MDA could inform control planning, especially if correlation is present. We investigated non-participation in children aged 1–9 years during three annual MDAs in The Gambia, a low prevalence setting. We found evidence that non-participation is associated with household membership and decision-making, as seen in medium and high prevalence settings in East Africa. In addition, we demonstrate geographical heterogeneity (spatial clustering) of non-participation, persistent non-participation behaviour over time and different non-participator types. Between the first and third MDA, non-participation increased significantly overall from 6.2% to 10.4%, whilst spatial clusters became smaller with non-participation more focused in single households or small groups of households. There was no evidence of association between infection and non-participation. In low prevalence settings with no evidence to suggest non-participation as a risk factor for infection, resources to improve participation may not be required. Spatial hotspot analysis could address this research question in areas with more infection.
| Trachoma is a leading cause of preventable blindness in endemic areas [1]. Control is through the SAFE strategy [2], of which a key component is mass drug administration (MDA) with the antibiotic azithromycin. Entire communities are targeted during MDA in order to reach both pre-school and school aged children who form the reservoir of infection for Chlamydia trachomatis, the causative bacterial agent for trachoma [3].
There is renewed commitment from the World Health Organization (WHO), donors of funding for disease control and research and also pharmaceutical companies to support efforts to eliminate Neglected Tropical Diseases (NTDs), including trachoma, by 2020 [4]–[6]. The success of MDA for NTDs is thought to depend heavily on adequate population coverage in affected areas and participation amongst those offered treatment [7], [8]. With increasing provision of MDA for trachoma, prevalence is expected to fall so that endemic areas will, over time, become low prevalence settings on a trajectory towards the endgame of elimination [4]. In such settings, MDA participation amongst those at highest risk of infection is important. If spatial clusters, or hotspots, of non-participation occur during MDA and correlate with hotspots of infection, it is possible that reservoirs of infection could remain to facilitate continued transmission [9]. This would in turn increase the time needed to reach elimination goals. Identification of factors associated with persistent non-participation in low prevalence settings could therefore provide important clues about how to minimise non-participation. Determining whether infected individuals are amongst non-participators in previous annual MDAs may also provide information regarding the importance of non-participation in low prevalence areas and the potential need for resources to improve participation.
C. trachomatis infection, follicular trachoma (TF) and non-participation with azithromycin MDA have all been found to cluster within communities and also within households [10]–[16]. Limited data on non-participation in trachoma control suggest that non-participation is associated mainly with household level decision-making factors, related to knowledge and awareness of trachoma control and also mode of delivery (for example, perception of community drug distributors). A case-control study in Tanzania found household level risk factors such as guardians of children reporting better health in themselves, increased burden due to poor family health, more children per household and younger guardians [3]. At community level, enhanced effort to increase coverage during implementation of MDA was successful in achieving higher participation rates. Studies in Nigeria and South Sudan identified prior household head knowledge of trachoma control and prior notification of MDA as factors associated with better participation but no association with age or gender [17], [18]. In a cluster randomised trial (CRT) in Ethiopia, women and younger children were more likely to be non-participators [15].
For CRTs evaluating the impact of MDA intervention, non-participation can be problematic as it can reduce power to detect intention-to-treat effects [19] and lead to bias in results if there is systematic or heterogeneous non-participation due to reasons also associated with the outcome [20], [21].
In the Partnership for Rapid Elimination of Trachoma (PRET) CRT in The Gambia [22], [23] which represents a hypo-endemic setting (prevalence of TF of 10–20%[7], [24]), MDA took place over a three year period to evaluate the effectiveness of different frequency and coverage MDA delivery strategies on C. trachomatis infection and TF in children aged 0–5 years. The aims of this study are to quantify non-participation amongst children aged 1–9 years during PRET, to identify factors associated with non-participation of different types at child, household and community level, to investigate the presence of heterogeneity of non-participation at household and, or community level and determine if any observed household or community heterogeneity is spatially clustered.
Approval was obtained from the London School of Hygiene & Tropical Medicine Ethics Committee, and The Gambia Government/Medical Research Council Unit, The Gambia Joint Ethics Committee. Written informed consent was obtained from a parent or guardian prior to examination for all children.
In PRET, 48 communities (enumerations areas, or EAs) were randomised in a 2×2 factorial design to MDA delivery strategies [11], [22], [23]. A frequency strategy allocation resulted in either three annual MDAs of all community members in 24 communities or MDA at baseline only in the remaining 24 communities. A coverage strategy allocation (24 communities per arm) was either standard (one day visit to each community by the treatment team of National Eye Health Program (NEHP) in The Gambia) or enhanced (two visits to each community to achieve higher coverage). At the end of the trial, the overall prevalence of TF was around 3% and of C. trachomatis, less than 1%.
All community members in treated EAs were eligible to receive azithromycin, with the exception of pregnant women and children under six months old who were offered tetracycline ointment if needed. The study took place in two adjacent districts on the northern Bank of the River Gambia and two adjacent districts on the southern Bank (Figure 1) identified for azithromycin MDA. Twelve EAs per district were randomly selected so that only one EA within settlements of more than one EA was chosen. A restricted randomisation of EAs within districts to trial arms was performed by the trial statistician, such that all EA within larger settlements of multiple EA received the same allocation to avoid contamination.
Every six months, between baseline and 36 months inclusive, a complete census was taken. Following this, a random sample of children aged 0–5 years was taken from each community in order to measure trachoma outcomes (the primary outcomes of PRET; presence of TF and C. trachomatis infection). Full details of survey methods, sampling strategies and measurement of trachoma outcomes are published elsewhere [11], [22]. MDA took place within approximately one month of the examination rounds. Treatment receipt for each individual was recorded against the census.
A central treatment station was set up in each community during MDAs. Adults aged 14 years or above received 1 g of azithromycin and height was used as a surrogate for weight for children's dosing on the basis of 20 mg/kg [25]. Treatment was administered and directly observed by NEHP treatment teams and the number of tablets or ml of suspension recorded within pre-printed fields included in census forms.
NEHP staff attended the initial training workshop for the PRET trial. Prior to each MDA, treatment team leaders received training about recording treatment status on census forms from the trial coordinator and about dosing from NEHP. Team leaders trained their team. Data review and feedback took place throughout MDAs. Communities were sensitised to MDA by the trial field team before fieldwork started. During the census prior to treatment, the study was again explained to households, and the expected dates for examination and treatment teams' visits were explained. Supervisory field visits were conducted by the NEHP to ensure appropriate distribution. Treatment team members were given per diems to cover food and accommodation for days spent in the field, as a single payment at the end of the fieldwork based on the expected number of days needed.
For each MDA, treatment receipt and eligibility were categorised according to one of the following categories:
Two binary outcomes were analysed for each MDA; 1) PNT versus treated and 2) EBA versus treated.
EA level variables included coverage allocation, North or South river bank and district, EA type (single settlement, multi-settlement, or segment of a settlement) and population size (small: <600, medium: 600–800, large: >800 individuals). For households, variables included size (small: <11, medium: 11–16, large: >16 individuals), latrine access, time to primary water source, recall of community health education, years of education of household head, a diagnosis of TF for a child aged 0–5 years in the household during the survey immediately prior to the MDA and treatment status of the household head. Child level variables were gender, age, participation in a previous ocular examination survey and treatment status at previous MDAs. Latitude and longitude coordinates were measured for each household.
Data were analysed using Stata, version 13 Special Edition [26] and SaTScan [27] and mapped using Quantum GIS [28]. All EAs were treated at baseline and 24 EAs at year one and year two. All available data for children aged 1–9 at the time of each MDA in treated EAs were used to analyse non-participation in this sub-study of the PRET trial. Children with unknown (missing) outcome data were excluded.
The number (%) of children treated, PNT or EBA was summarised overall and by characteristics of interest for each MDA, treating each as a cross-sectional survey (Table S1). Using random effects logistic regression, multivariable models were developed for both outcomes using the baseline data. EA level random intercepts were included in all models and household level random intercepts for EBA versus treated comparisons. PNT children were too few to include a household level random effect. Factors associated with the outcome by a likelihood ratio test (LRT) p-value of <0.1 in univariable analyses were included in a step-wise model building approach to obtain a final multivariable model. Coverage delivery allocation was included in all multivariable models a priori since by design the enhanced allocation was intended to increase participation. The same multivariable models were fitted to the year one and two MDA data for validation. Treatment status at previous treatment rounds was added to each of these final models a priori. Tests for interaction with coverage allocation were pre-specified if an association between coverage allocation and the outcome was found. Intracluster correlation coefficients (ICCs) with corresponding 95% confidence intervals were obtained from final multivariable models.
Considering the study areas north and south of the River Gambia separately, spatial point patterns were investigated using Kulldorf's scan statistic [29] for each MDA round (baseline, year one and year two), in order to test whether PNT and EBA cases were randomly distributed over space compared to treated children and to identify the location of any significant spatial clusters. Within SatScan software, a circular window is moved systematically throughout the geographic space to identify clusters by centring the window on each household location with a window size of 0% to 50% of the study population, to allow detection of small and large clusters. Clusters containing more than 50% of the population are ignored. A LRT test for a Poisson based model was conducted for each location and size of scanning window to test the hypothesis of an increased rate of non-participator type compared with the distribution outside the window. P-values corresponding to the most likely and secondary clusters are determined using Monte Carlo replications of the dataset. Spatial clusters of PNT and EBA children were added to maps showing the location of children and their treatment status. The locations of infected children at year three are shown on the map for the year two MDA for visual inspection.
Treatment status was unknown for 403 (3.6%), 88 (1.6%) and 187 (3.0%) eligible children at baseline, year one and year two, respectively. Participation was high overall during each MDA. The overall prevalence of non-participation at baseline was 6.2% (604/9777) with 1.0% (99/9777) of children being PNT and 5.2% (505/9777) of children EBA (Table S1). The distribution of treatment status was similar at year one. Over the three MDAs, the percentage of EBA children appeared to increase and the percentage of PNT children to decrease. By year two, overall non-participation increased to 10.4% (paired t-test of EA summary data p<0.01) due to the increase in EBA children. Reductions in PNT non-participation were not significant.
Of 1626 households eligible for treatment in 24 annually treated communities, one household (0.1%) had PNT children in all three MDAs and 34 (2.1%) had EBA children in all three MDAs. Persistent EBA households were generally larger and within EAs comprised of multiple settlements. The persistent PNT household was further from water, without latrine access and with a household head with no recall of health education or education.
Univariable analyses of baseline data are presented in Table 1. The final multivariable model for being PNT versus treated at baseline included coverage allocation, time to water, household size, household head treatment status and district (Table 2). Children residing in a medium or large household compared to small (p<0.001) and within 15 minutes of primary water source (p<0.001) were less likely to be PNT. A child was more likely to be PNT if the household head was untreated (p<0.001). An association with district was also found (p = 0.002), due to a difference between districts south of the River Gambia. No effect of coverage allocation was found (p = 0.842). A TF diagnosis in the household during the baseline examination round, approximately one month prior to treatment, was associated with lower odds of being PNT in univariable analyses (Table 1) but not after adjustment for other factors in the final model.
The same final model was fitted to the year one and year two data, adding previous treatment status. For these follow-up MDAs, the fixed term for district was removed due to zero PNT cases north of the river. Treatment status one year previously was an important predictor of non-participation at both years one and two, with children who were PNT at the previous round being more likely to be PNT again the following year (baseline treatment status at year one MDA: p = 0.034, year one treatment status at year two MDA: p = 0.032, Table 2). Treatment status at baseline was not associated with being PNT at year two (p = 0.656).
The final multivariable model for being EBA versus treated at baseline (Table 3) suggested being EBA was more likely for children who were not included in the baseline examination round (p<0.001), aged 3–5 or 1–2 years compared to 6–9 years (p<0.001), whose household head was also EBA compared to treated, who resided in households further from water (p = 0.018) and possibly for those whose household head could not recall community health education (p = 0.060). Coverage allocation was not associated with being EBA (p = 0.166). Children who were EBA at each previous round were more likely to be EBA at later time points (Table 3). Results also suggest that children who were ineligible at both previous treatment rounds were more likely to be EBA at year two.
In the EBA versus treated comparisons ICCs suggested substantially more variation was present between households within EAs, than between EAs (Table 3). ICCs from PNT models at EA level were closer to the ICCs estimated at household level for EBA children, possibly because between-household variation could not be determined due to the very low prevalence of PNT non-participation.
GPS coordinates were missing for 11 out of 1626 households, excluding 23 children from spatial analyses.
Spatial clusters of PNT and EBA children were detected at baseline in study areas on each side of the river (Table 4). No PNT children were reported in year one or year two in the northern river bank districts. Spatial clusters of PNT and EBA children reduced in size in each subsequent MDA and by year two, clusters included either single households or a small group of adjacent households (Figures 2 and 3).
Cases of C. trachomatis infection in annually treated communities at year three (n = 14) were found within three kilometres of Senegal in all but one child. In Senegalese districts adjacent to The Gambia, MDA had not yet taken place. Infections were detected amongst children who were ineligible or treated during the three prior MDAs, apart from one child residing on the north side of the river who was persistently EBA during the MDAs. Two cases were located in an EA with households within a year two EBA cluster on the south side of the river (Figures 2 and 3). In the two EAs with households in this spatial cluster, approximately 15% of 1–9 year olds were EBA during the year two MDA.
In this large study of non-participation during azithromycin MDA from a low prevalence trachoma setting, we demonstrate further evidence of heterogeneity of non-participation in children aged 1–9 years, particularly at household level, in line with studies in higher prevalence settings. We also observed persistent non-participation over time in annual MDAs, as seen elsewhere in a CRT setting [3]. Geographical clustering of non-participation represents a new finding and we found two different types of non-participators. We found circumstantial rather than statistical evidence of an association between infection and non-participation during a previous MDA, however, the overall prevalence of infection and TF in 0–5 year olds at the end of PRET was below a level requiring any SAFE interventions. Detection of infection in communities close to untreated areas [22], relatively high EBA rates in those communities during the previous MDA and literature from The Gambia and elsewhere linking travel with re-infection [30], [31] together, suggest the observed infections could have resulted from exposure to untreated persons. Travel plans could have been set prior to notification of MDA timing and therefore could have been unrelated to intentional non-participation, although intentional decision making to miss treatment is a possibility.
Household level variables were associated with greater likelihood of being PNT and EBA. Household head non-participation and their type of non-participation predicted PNT and EBA status in children, implying household decision making with respect to MDA participation behaviour. The finding that children in households further from their primary water source were more likely to be PNT or EBA is probably indicative of some other unmeasured risk factor, for example, marginalisation within the community due to either household head or community leader choice, or a mixture of the two. Non-participation during MDA subsequent to participation in a previous MDA has been found to be associated with possible markers of marginalisation in another CRT [32]. Active trachoma has been found to be associated with lower socio-economic status (SES) and isolation of households from the community [33] so access to, or participation in, trachoma control activities could also be affected by these unmeasured factors. Smaller household size was important for predicting PNT status but not EBA, compared to treated children, which could represent some effect of lower SES. Recent migration into the community could also mean less access to community decision making and activities. Participation in a previous TF examination survey could be indicative of increased awareness and acceptance of control activities in annually treated communities, however, a proxy effect cannot be concluded in case of potential bias introduced by households more willing to take part in all control and assessment activities. Results from the Gambian setting suggest that enhanced efforts to increase coverage of mass treatment programs, by means of an extra treatment team visit to the community do not improve participation, in contrast to the PRET trial conducted in Tanzania [3].
Studies of MDA participation in Africa for onchocerciasis and lymphatic filariasis, other NTDs for which control is through mass community-wide treatment, have also linked non-participation to household level decision making factors, for example, a perception of low disease risk or lack of family or household support [34]–[36]. The Gambia has relatively high childhood immunisation coverage [37], elimination of trachoma by 2020 is attainable [24] and non-participation was higher in the districts south of the river where the prevalence of TF was consistently lower during PRET [22]. It is perhaps plausible therefore that a household level decision based on a perceived lack of need for treatment could apply in this low prevalence setting, although we do not have data from each community to assess this. Reasons for being EBA in this setting could be logistic and independent of participation choices, for example, population movement and travel where children are sent away for weaning which is common practice in The Gambia, or farming related activities. PNT and EBA comparisons to treated children were performed separately as it was hypothesised that there may be differences in reasons for non-participation that may or may not be related to refusal of treatment or a perceived lack of need for treatment. The data do suggest some differences between PNT and EBA children but further information is unavailable to determine if and why there was an active decision to refuse treatment.
Due to the very low prevalence of TF and infection in both MDA frequency arms (annual and baseline only MDA) throughout the original trial, it is unlikely that the heterogeneous non-participation observed here had an additional negative effect on power to detect differences between arms in intention-to-treat analyses in the PRET trial. It is also unlikely that heterogeneous non-participation introduced bias in comparative analyses given the low prevalence of TF and infection. We found a geographical effect on non-participation and on trachoma outcomes [22]. Infections did occur in one part of the study area with notable EBA non-participation at the previous MDA, however, even if all PNT and EBA children at the year two MDA had been found to have infection and TF, the overall prevalence of each outcome at year three would have been less than 5% and thus still below MDA continuation thresholds for TF. Therefore, for the Gambian national trachoma control program, efforts and resources to address non-participation are not required.
For national control programs in low and medium prevalence settings, heterogeneous non-participation linked to increased risk of infection could present challenges for elimination efforts. Links between infection and non-participation in prior MDA rounds could undermine MDA where corresponding prevalence levels for TF meet criteria for continued MDA at the time of impact assessment. Identification of hotspots of infection and non-participation, along with modifiable risk factors for non-participation could take place during impact assessment following repeated MDA. The results could then aid control program managers working towards elimination goals in low and medium prevalence settings, by enabling them to target delivery resources for continued MDA and to improve coverage in areas with a greater threat of continued transmission.
|
10.1371/journal.ppat.1004071 | Histone Deacetylase Inhibitor Romidepsin Induces HIV Expression in CD4 T Cells from Patients on Suppressive Antiretroviral Therapy at Concentrations Achieved by Clinical Dosing | Persistent latent reservoir of replication-competent proviruses in memory CD4 T cells is a major obstacle to curing HIV infection. Pharmacological activation of HIV expression in latently infected cells is being explored as one of the strategies to deplete the latent HIV reservoir. In this study, we characterized the ability of romidepsin (RMD), a histone deacetylase inhibitor approved for the treatment of T-cell lymphomas, to activate the expression of latent HIV. In an in vitro T-cell model of HIV latency, RMD was the most potent inducer of HIV (EC50 = 4.5 nM) compared with vorinostat (VOR; EC50 = 3,950 nM) and other histone deacetylase (HDAC) inhibitors in clinical development including panobinostat (PNB; EC50 = 10 nM). The HIV induction potencies of RMD, VOR, and PNB paralleled their inhibitory activities against multiple human HDAC isoenzymes. In both resting and memory CD4 T cells isolated from HIV-infected patients on suppressive combination antiretroviral therapy (cART), a 4-hour exposure to 40 nM RMD induced a mean 6-fold increase in intracellular HIV RNA levels, whereas a 24-hour treatment with 1 µM VOR resulted in 2- to 3-fold increases. RMD-induced intracellular HIV RNA expression persisted for 48 hours and correlated with sustained inhibition of cell-associated HDAC activity. By comparison, the induction of HIV RNA by VOR and PNB was transient and diminished after 24 hours. RMD also increased levels of extracellular HIV RNA and virions from both memory and resting CD4 T-cell cultures. The activation of HIV expression was observed at RMD concentrations below the drug plasma levels achieved by doses used in patients treated for T-cell lymphomas. In conclusion, RMD induces HIV expression ex vivo at concentrations that can be achieved clinically, indicating that the drug may reactivate latent HIV in patients on suppressive cART.
| Combination antiretroviral therapy has greatly improved the clinical outcome of HIV infection treatment. However, latent viral reservoirs established primarily in memory CD4 T cells persist even after long periods of suppressive antiretroviral therapy, which hinders the ability to achieve a prolonged drug-free remission or a cure of the HIV infection. Activation of HIV expression from latent reservoirs is a part of proposed strategies that may potentially lead to virus elimination and ultimately cure of the infection. In this study, we show that romidepsin, a histone deacetylase inhibitor approved for the treatment of T-cell lymphomas, is a potent activator of HIV expression in an in vitro model of viral latency as well as ex vivo in resting and memory CD4 T cells isolated from HIV-infected patients with suppressed viremia. Importantly, the ex vivo activation of latent HIV occurred at romidepsin concentrations lower than those achieved in drug-treated lymphoma patients. In addition, romidepsin exhibited a more potent effect than other drugs in the same class that have already been shown to activate HIV expression in vivo. Together, these results support the clinical assessment of romidepsin in HIV-infected patients on suppressive antiretroviral therapy.
| Combination antiretroviral therapy (cART) has dramatically improved the life expectancy and health of patients infected with HIV. In the setting of controlled clinical trials with optimal cART, up to 90% of treatment-naïve patients can achieve undetectable virus in plasma and normalization of CD4 T-cell levels [1], [2]. However, when cART is interrupted in patients who initiated therapy during the chronic phase of infection, virus replication resumes in virtually all patients [3]–[5], indicating that current cART is not sufficient to cure HIV infection. The failure of cART to cure HIV infection is due, in part, to the ability of HIV to establish latency in a subset of infected CD4 T cells [6]. The state of latency is characterized by the presence of integrated but transcriptionally silent proviral HIV DNA, which makes the infected cells invisible to the immune system and resistant to both innate antiviral defenses and antiretroviral therapy [6], [7].
Although latent proviral DNA has been detected in multiple different immune cell subsets permissive to HIV infection, long-lived resting memory CD4 T cells are believed to represent the predominant reservoir of proviruses that can be activated to produce infectious virions [8], [9]. Initial quantification of latent HIV proviruses in peripheral blood lymphocytes from patients on cART revealed approximately 200 copies per 106 resting CD4 T cells; however, in general, less than 1% of these proviruses was shown to produce infectious HIV after T-cell mitogenic stimulation with substantial inter-patient variation observed in the fraction of total proviruses that could be activated [10]. The pool of latently infected memory CD4 T cells is believed to be maintained throughout a patient's life by homeostatic proliferation of memory T cells and/or intermittent antigen-driven clonal expansion [11]. Alternatively, low levels of HIV replication confined to lymphatic tissues and undetectable in the periphery may also contribute to the maintenance of the latent virus reservoir [11], [12]. The decay rate of latent virus reservoirs in peripheral blood lymphocytes has been estimated to have a half-life of >3 years, indicating that even life-long cART is unlikely to cure HIV infection [7].
Chronic HIV infection, even when suppressed by cART, poses long-term health risks that include accelerated cardiovascular disease, liver and renal disease, non-AIDS-associated cancers, neurocognitive impairment, and accelerated senescence of immune responses [13]–[15]. Thus, there is a clear unmet medical need for novel therapeutic interventions that could lead either to host-mediated control of HIV in the absence of cART or complete clearance of viral reservoirs. Such virus eradication interventions will need to be well tolerated with minimal side effects. Specifically, therapeutic interventions need to be found that do not cause global T cell activation or a chronic state of inflammation. These interventions should also exhibit minimal drug-drug interactions with medications frequently administered to HIV-infected patients.
Identifying a safe modality to activate latent HIV in memory CD4 T cells is an important goal and potentially represents the first key step towards a cure for HIV. Establishment and maintenance of HIV latency is a complex process that appears to involve multiple mechanisms restricting productive viral transcription. These mechanisms include promoter occlusion via steric hindrance, insufficient levels of cellular transcription factors, modification of the HIV 5′ long-terminal repeat (LTR) by methylation and alteration of the chromatin environment in the vicinity of the LTR by histone deacetylation and other epigenetic modifications [16], [17]. Histone deacetylases (HDACs) have been implicated in maintaining HIV in a latent state. In this process, HDACs are recruited to the LTR by various transcriptional regulators and deacetylate lysine residues on histones, inducing chromatin condensation, thereby repressing proviral transcription [18], [19]. Consistent with this mechanism, HDAC inhibitors (HDACi) have been reported to activate latent HIV in cell lines and primary cells, including CD4 T cells from HIV-infected patients on cART [20], [21]. Vorinostat (suberoylanilide hydroxamic acid; VOR) is an HDACi approved for clinical use to treat cutaneous T-cell lymphomas and has also been shown to activate HIV transcription in various latency models [22]–[24]. Administration of a single dose of VOR to eight HIV-infected patients on cART increased HIV RNA levels in resting CD4 T-cells by a mean of 4.8-fold [25]. Although it is unclear whether a single dose of VOR diminished the activatable latent HIV reservoir, these results represent an important milestone by demonstrating that HIV expression can be increased pharmacologically in HIV-infected patients on cART.
In a recent study exploring a panel of HDACi for HIV activation using an in vitro latency assay, panobinostat (LBH589; PNB) displayed superior potency to multiple other HDACi tested including givinostat, belinostat, and VOR [26]. Notably, a potent HDACi romidepsin (RMD; Istodax) was not tested in this study. RMD is a cyclic peptide naturally synthesized by Chromobacterium violaceum [27] that has received regulatory approval for the treatment of patients with peripheral T-cell lymphomas (PTCL) or cutaneous T-cell lymphoma (CTCL) [28].
The present study explores the ability of RMD, in comparison with VOR and other HDACi currently in clinical development, to reverse HIV latency in vitro in primary T cells infected with a reporter virus as well as ex vivo in resting and memory CD4 T cells from HIV-infected patients on suppressive cART.
To assess clinically tested HDACi for their ability to activate HIV from latency, we first employed a previously described in vitro HIV latency model [29], [30] with several modifications to increase the sensitivity of detection. The assay involves freshly isolated naïve CD4 T cells from healthy donors polarized to a Th0 phenotype, which mimics memory CD4 T cells. These Th0 cells are then infected with HIV expressing a luciferase reporter gene and cultured for an additional 7 to 10 days until a latent infection is established.
The HDACi that we tested in this in vitro HIV latency assay included RMD, VOR, PNB, givinostat, mocetinostat, and pracinostat (SB939). All compounds showed dose-dependent activity in the assay, but displayed varying levels of potency in activating HIV expression (Fig. 1A). Based on experiments conducted with cells prepared from three independent donors, RMD was the most potent HDACi with a mean EC50 value of 4.5 nM (Table 1). Global T-cell activators such as anti-CD3/CD28 antibodies and PMA+ ionomycin consistently showed 2- to 4-fold higher maximum induction of the luciferase signal relative to RMD (data not shown). When cytotoxicity was determined under the identical conditions, RMD displayed a CC50 (50% cell viability reduction) value of 100 nM, resulting in an approximately 20-fold selectivity window (Table 1). PNB was the second most potent compound tested with an EC50 value of 10 nM and a relatively high selectivity window of >250-fold. VOR was substantially less potent in this assay with EC50 and CC50 values of 4 µM and >25 µM, respectively.
To confirm that the HDACi were activating HIV expression in this latency model, we used flow cytometry analysis to quantify intracellular p24 antigen levels following treatment of latently infected CD4 T-cell cultures with RMD, VOR, and the positive control anti-CD3/CD28 antibodies (Fig. 1B). Treatment with 5 and 80 nM RMD, representing the minimal and maximal concentrations leading to the activation of HIV expression in this model, resulted in 3.3% and 5.5% of cells expressing p24 antigen, respectively. In comparison, treatment with 3.0 µM VOR induced p24 antigen expression in approximately 4.4% of cells. As expected, anti-CD3/CD28 antibodies induced p24 antigen expression in 2- to 3-fold higher fractions of CD4 T cells than RMD (Fig. 1B). While the p24 expression in anti-CD3/CD28-treated cells reached a plateau at 48 hours, the fraction of p24 positive cells in RMD-treated cultures continued to increase for 72 hours (Fig. 1C). These data indicate that RMD induces the expression of HIV proteins following the activation of latent provirus in vitro.
Since the specific mechanism of HIV latency reversal by HDACi remains to be fully understood, we investigated whether the relative potency of selected compounds in the HIV latency assay correlates with their ability to directly inhibit the activity of individual HDAC isoenzymes. RMD, VOR, and PNB were tested against 11 individual HDAC isoenzymes (HDAC-1 to -11) from four distinct classes: 1, 2a, 2b, and 4. Overall, RMD was the most potent inhibitor, especially against class 1 HDACs (HDAC-1 to -3) as well as HDAC-10 and HDAC-11 isoenzymes (Table 2). PNB showed lower potency, particularly against class 1 and 4 enzymes with up to 40-fold higher IC50 values relative to RMD. VOR was a substantially weaker inhibitor of the majority of HDAC enzymes than RMD with the exception of HDAC-6. The greatest difference in potency between RMD and VOR was observed with the class 1 HDAC enzymes that were 60- to 2,500-fold more susceptible to RMD than VOR (Table 2). The greater activity of RMD against a broad range of HDAC enzymes correlated with its higher potency to activate HIV expression relative to VOR (∼1,000-fold difference; compare Tables 1 and 2).
Previous studies revealed that VOR activates HIV transcription ex vivo in resting CD4 T cells isolated from HIV-infected patients on cART [21]. We used a similar approach to compare RMD to VOR in total memory as well as resting CD4 T cells isolated from patients chronically infected with HIV who were treated with cART and maintained their plasma HIV RNA at <50 copies/ml for at least 12 months (See Supplementary Table S1 for patient descriptions). In the initial set of experiments, isolated CD4 cells were treated with 40 nM RMD for 4 hours or 1 µM VOR for up to 24 hours to mimic the clinical exposure profiles of these drugs. To examine the kinetics of HIV induction, we analyzed cell-associated viral RNA levels at 6, 12, 24, and 48 hours after treatment initiation. Consistent with the data of Archin et al. [21], a 2- to 4-fold increase in HIV RNA levels was observed both in memory and resting CD4 T cells after 6 hours of VOR treatment compared with control vehicle (DMSO)-treated cells (Fig. 2). However, HIV RNA in VOR-treated cells decreased by 48 hours to levels detected in vehicle-treated cells. In contrast, cell-associated HIV RNA continued to increase in both memory and resting CD4 T cells isolated from the same donors following their exposure to RMD (Fig. 2). Levels of intracellular HIV RNA in both cell types were 5- to 6-fold higher compared with vehicle-treated controls and peaked between 24 and 48 hours after the addition of RMD. These findings suggest that the activation of HIV transcription with RMD is more durable than with VOR.
We next examined whether RMD and VOR induce the release of HIV particles from patient-derived memory and resting CD4 T cells. We used HIV RNA in cell culture supernatants as a surrogate marker for virion release from treated cells. We assumed that extracellular virions would accumulate over the course of several days; therefore, the duration of culture was extended to 6 days to maximize assay sensitivity. Memory CD4 T cells from multiple patients were treated with 5 and 20 nM RMD for 4 hours, or 0.5 and 1 µM VOR for 24 hours. Under these conditions, RMD, but not VOR, increased extracellular HIV RNA in culture supernatants of memory CD4 cells isolated from multiple HIV-infected virally suppressed patients (Fig. 3A). The release of viral RNA from cells treated with RMD is unlikely to be due to cytotoxicity of the compound because the tested concentrations did not substantially affect cell viability (Supplementary Fig. S1). In addition, highly cytotoxic compounds such as blasticidine caused >80% cell death during the same incubation period, but did not lead to release of detectable levels of HIV RNA into cell culture supernatants (data not shown).
To determine whether the observed lack of HIV RNA release from cells treated with VOR was due to an insufficient duration of drug treatment, memory and resting CD4 T cells isolated from HIV-infected patients on cART were treated with 1 µM VOR for 6 days. Similar to the cultures treated with VOR for 24 hours, there was no significant increase in HIV RNA in culture supernatants following a 6-day VOR treatment of memory CD4 T cells isolated from two HIV-infected patients on suppressive cART. In contrast, continuous 6-day treatment with RMD resulted in the detection of extracellular HIV RNA in memory CD4 T-cell cultures from the majority of tested donors (Fig. 3B). Continuous treatment with RMD was conducted at a lower drug concentration (5 nM) to avoid cytotoxicity. Similarly, 2.5 nM RMD induced HIV RNA release in resting CD4 T-cell cultures from 6 of 8 tested donors. In comparison, treatment of resting CD4 T cells with 1 µM VOR resulted in a measurable increase of extracellular HIV RNA in 3 of 7 patient-derived cultures tested, but 0.5 µM VOR did not increase extracellular HIV RNA significantly above the levels of untreated controls (Fig. 3C). Importantly, HIV RNA released into cell culture supernatants of the resting CD4 T cells treated with RMD can be pelleted by high-speed centrifugation, supporting the conclusion that the extracellular HIV RNA is associated with virion particles (Supplementary Table S2).
Although the relative in vitro HIV activation potency of RMD, VOR, and PNB correlated with the inhibitory activity of these compounds against multiple HDAC isoforms, we next set out to determine whether inhibition of cellular HDAC activity in latently infected cells also correlated with HIV activation. To address this question, resting CD4 T cells isolated from three cART-suppressed HIV-infected patients were treated with RMD, VOR, or PNB, and activation of intracellular HIV RNA expression was determined in parallel with cell-associated HDAC enzymatic activity. We found that the activation of HIV expression correlated with HDAC inhibition across multiple time points ranging from 6 to 48 hours (Fig. 4). In RMD-treated cells, both HIV activation and cell-associated HDAC inhibition persisted throughout the 48-hour incubation period. In contrast, the inhibition of HDAC activity by VOR and PNB was transient and decreased after 24 hours of incubation, which paralleled the levels of cell-associated HIV RNA over time (Fig. 4). Concentrations of VOR and PNB used in these experiments corresponded to 100% systemic drug exposure observed in treated cancer patients, while the concentration of RMD was 40% of the clinical drug exposure. These results reveal that sustained inhibition of cell-associated HDAC activity by RMD correlates with persistent activation of HIV expression.
Since RMD-related toxicities have been observed at doses used for the clinical treatment of hematologic malignancies [31], [32], we tested whether RMD can activate latent HIV in patient-derived CD4 T cells at concentrations below its clinical exposure achieved with the dose of 14 mg/m2 used for the treatment of lymphomas [33]. Resting CD4 T cells from three HIV-infected patients were treated ex vivo with 3.5 to 40 nM RMD for 4 hours to mimic 40% or less of the systemic drug levels achieved during the i.v. administration of the clinical dose based on the relative free drug fraction in human plasma and cell culture media (Supplementary Table S3). While minimal induction of HIV transcription in resting CD4 T cells was observed with 3.5 nM RMD, treatment with 15 or 40 nM RMD induced 4- to 6-fold activation of HIV RNA expression (Fig. 5). This contrasted with results obtained following treatment with 1 µM VOR, which induced rapid, but transient and less-pronounced activation of latent HIV expression (Fig. 2 and Fig. 4A). Based on the measured free drug concentration in cell culture media and human serum, 1 µM VOR is above the clinical exposure achieved in patients following the oral administration of the clinical dose of 400 mg (Supplementary Table S3). Thus, RMD is capable of durable induction of HIV transcription in cells from HIV-infected patients on suppressive cART at concentrations substantially below those achieved in cancer patients following the administration of the clinically approved dose of the drug.
HIV latency reversing agents that would be suitable for the therapeutic use in vivo should not induce non-specific immune activation. To determine if RMD may activate immune cells, PBMCs isolated from HIV-infected subjects on suppressive cART were treated with a 4-hour pulse of 15 or 40 nM RMD. The activation status of selected immune cell subsets, including CD4+ and CD8+ T cells, as well as CD19+ B cells was assessed by flow cytometry analysis of cellular activation markers relative to treatment with either 1 µM VOR or vehicle control. While RMD treatment induced dose-dependent expression of CD69 in 10% to 50% of T and B cells, it did not lead to any changes in the expression of other prominent cell activation markers such as CD25 or HLA-DR in any of the cell subsets (Fig. 6). In contrast to effects observed in the context of complete PBMC cultures, RMD treatment induced minimal increases in the fraction of CD69-positive cells in resting CD4 T cells isolated from cART-suppressed HIV-infected patients (Supplementary Fig. S2). In addition, no significant induction of IFN-α, IFN-γ, TNF- α, TGF-β, IL-2, IL-7 or other cytokines were detected in PBMC cultures from the same patients following the treatment with RMD (data not shown). Together, these data indicate that RMD effectively activates the expression of latent HIV without inducing the global activation of T or B cells.
Recent results showed inter-patient variation in ex vivo HIV activation by VOR [25]. Thus, it was important to determine whether HIV reactivation by RMD was consistent in samples collected longitudinally from the same patient over time. We measured HIV RNA levels following RMD treatment of resting CD4 T cells from three sequential blood samples collected several weeks apart from the same two HIV-infected patients with suppressed plasma viral load. One of the tested subjects showed robust and reproducible dose-dependent HIV RNA increase in culture supernatants after 7 days of RMD treatment in all three longitudinal samples (Fig. 7; high-responding patient). Although the other subject exhibited weaker HIV activation following treatment with RMD, a concentration-dependent effect on viral expression was observed in 2 of 3 longitudinal samples (Fig. 7; low-responding patient). These results show that RMD is capable of eliciting reproducible reactivation of latent HIV ex vivo in samples collected longitudinally from the same patient.
We used a single-genome sequencing (SGS) [34]–[36] approach to analyze sequences of HIV gag-pol RNA in cell culture supernatants of resting CD4 T cells from the high-responding patient (depicted in Fig. 7) following the latency reversal by RMD and compared those with sequences of HIV RNA induced by the stimulation of CD3/CD28. In parallel, we assessed the sequence diversity of HIV proviruses integrated in genomic DNA of resting CD4 T cells from the same culture. The SGS analysis confirmed the presence of multiple HIV RNA sequences in the supernatants of resting CD4 T-cell cultures following a 7-day treatment with RMD or anti-CD3/CD28 antibodies. Importantly, several proviral DNA sequences were identified that matched some of the RMD-induced HIV RNA sequences in culture supernatants (Fig. 8). However, many proviruses did not have matching RNA sequences induced in the cell culture supernatants. This is consistent with previous observations that a large proportion of integrated proviruses are refractory to the activation by latency reversal agents [37]. Notably, a few instances were found in which identical HIV RNA sequences were induced by both RMD and anti-CD3/CD28 antibodies, but many HIV RNA sequences were also detected that are unique for each treatment. This is likely, at least in part, due to different mechanisms by which the two stimuli activate latent HIV. Although more extensive sequence analyses of samples from a larger set of HIV-infected patients on suppressive cART are needed to characterize proviruses that can be specifically activated by RMD, these initial results further confirm that RMD treatment activates a subset of latent HIV proviruses in resting CD4 T cells.
In this report, we show that the HDACi RMD acts as a potent activator of latent HIV with effects observed in resting and memory CD4 T cells either infected in vitro or isolated from HIV-infected patients on cART with suppressed viremia. RMD was a more potent inducer of HIV expression compared to VOR and PNB, two HDACi that are being evaluated clinically in HIV-infected patients on cART. The greater potency of RMD compared with VOR to induce HIV expression was observed in multiple patient samples and at multiple time points in both total memory and resting CD4 T cells. RMD also exhibited a greater magnitude and longer persistence of HIV RNA expression relative to VOR, and was capable of inducing the release of HIV virions from infected cells. The lasting effect of RMD compared to other HDACi may be related to its unique intracellular pharmacology and interaction with HDAC enzymes. RMD acts as an intracellular prodrug that undergoes reduction of its intramolecular disulfide bond upon entering cells [38]. The released free sulfhydryl groups tightly interact with the Zn2+ ion in the active site of various target HDAC isoforms, a mechanism of inhibition that does not apply to VOR or PNB.
In addition, RMD activated latent HIV ex vivo in cells from HIV-infected patients on cART at concentrations that are lower than those achieved in cancer patients receiving the clinically approved dose of 14 mg/m2. In patients with either hematologic malignancies or solid tumors, RMD showed an acceptable short-term safety profile and was not associated with severe toxicities [33]. The most common adverse reactions related to once-weekly dosing of RMD were thrombocytopenia, neutropenia, lymphopenia, nausea, and fatigue [33].
In our study, we observed reproducible ex vivo activation of HIV by RMD at concentrations corresponding to systemic clinical exposures expected after dosing at 2–5 mg/m2. Given these results and the established clinical safety profile of RMD, clinical testing is warranted to assess whether RMD can activate latent HIV and potentially reduce the size of the latent reservoir in HIV-infected patients on suppressive cART.
Recently, results from two clinical studies of single and multiple doses of VOR administered to HIV-infected patients on cART have been reported. The first study demonstrated a mean 4.8-fold increase in cell-associated HIV RNA levels in resting CD4 T cells following the administration of a single 400 mg dose of VOR to 8 subjects [25]. Another study examined once-daily dosing of VOR for 14 consecutive days and reported an approximate 3-fold mean increase in cell-associated HIV RNA levels in a cohort of 20 patients [39]. However, increases in the level of plasma viremia were not observed in either study. This finding could be consistent with our observation that VOR does not cause release of virions as measured by extracellular HIV RNA. In addition to VOR, multiple doses of PNB have been shown to induce intracellular HIV transcription and low-level viremia in peripheral blood of HIV-positive patients on suppressive cART [40].
Although monitoring the induction of HIV RNA expression in resting CD4 T cells from virally suppressed subjects treated with VOR was shown to be feasible, determining whether such treatments with HDACi can affect the size of the latent virus reservoir remains a challenge. There was no apparent effect on total proviral DNA in peripheral CD4 T cells in the multiple-dose (14-day) trial of VOR [39]. This outcome may be explained in part due to the presence of a high background of defective proviruses that cannot be activated and thus would not undergo depletion from viral cytopathic effect or immune-based clearance. In addition, Ho et al. recently showed that a large fraction of intact proviruses are refractory to latency reversal [37]. Therefore, in future clinical trials of HIV latency-reversing agents, it will be critical to use methods that will allow more direct assessment of the changes in the size of the inducible HIV reservoir following its pharmacological activation.
Induction of HIV expression from latent reservoirs is being pursued as a component of the reservoir eradication strategy that may ultimately lead to a prolonged drug-free remission or even cure [11], [12]. It should be noted that RMD activates latent HIV to a lesser degree than the mitogenic T-cell activators PMA+ionomycin or anti-CD3/CD28 antibodies; therefore, additional significant efforts would likely have to be devoted to improving the HIV activation effects of RMD. As already suggested, this may potentially be achieved by combining HDACi with other latency-reversing agents that work through complementary mechanisms. For example, synergistic effects on HIV expression have been documented in vitro between VOR and the PKC activator prostratin [41]. Other classes of agents shown to activate latent HIV in vitro and/or ex vivo include histone methyltransferase inhibitors [42], [43], as well as other small molecules such as disulfiram [44], [45]. Currently, however, no specific combination of agents suitable for testing in patients has emerged. Therefore, in future studies, the latency reversing agents should be considered for testing in combination with RMD to identify potential synergies for HIV activation. Concomitantly, any identified combinations of synergistic agents would have to undergo rigorous safety evaluations as the combined safety profiles of compounds independently modulating gene expression cannot be easily deduced.
At this early discovery stage of novel agents that activate latent HIV, it is also important to realize that the fate of cells expressing pharmacologically reactivated proviruses is not fully understood. While acute HIV infection of CD4 T cells is associated with cell death via both direct and bystander mechanisms, it is unknown whether reactivation of latent HIV with RMD or other latency-reversing agents would lead to virus-induced cell death. Recent data generated in an in vitro model of HIV latency has shown that HIV activation with VOR does not lead to cell death [46]. However, this result could be at least in part attributed to the nature of the latency model that relied on the ectopic expression of the anti-apoptotic factor Bcl2 to prolong the viability of the latently infected cells and/or to other unrecognized differences between in vitro models and the in vivo state. Addressing the fate of infected CD4 T cells after virus reactivation with RMD is a high priority of ongoing studies.
If reactivation of latent virus does not lead to cell death, additional interventions would be needed to eliminate cells expressing viral proteins. Strategies may include agents that selectively enhance apoptosis in infected cells following the latency reversal [47]. Other approaches under consideration include antibody-based therapies to recruit immune effector cells such as natural killer cells via antibody-dependent cell-mediated cytotoxicity [48], a strategy that has been used clinically for the treatment of various malignancies [49]. One potential advantage of the antibody-mediated clearance of virally infected cells is the high degree of selectivity of monoclonal antibodies for cells expressing the HIV envelope antigen. However, it is unclear whether the pharmacological activation of latent HIV will induce sufficient levels of gp120/gp41 expression on the surface of the treated cells to trigger an antibody-mediated effector mechanism resulting in cell lysis. Although we and others have detected intracellular HIV p24 capsid protein following treatment with HDACi using in vitro HIV latency models, little is known about levels of viral protein expression that can be induced by HDACi in latently infected resting CD4 T cells from cART-treated patients. One of the major challenges in addressing this question is the very low frequency of T cells harboring latent HIV provirus in virally suppressed patients [37]. New approaches to selectively enrich the fraction of latently infected cells isolated from patients are being explored [50] and may ultimately enable more in-depth characterization of RMD and other latency-reversing agents.
It should be noted that the treatment with HDACi could modify immune functions and responses (reviewed in [51], [52]). Various HDACi have been shown to inhibit innate immunity responses and cytokine production [53], [54] as well as immune cell trafficking and functions [55]–[57]. While the clinical doses of RMD approved for oncology applications can diminish responses of immune effector cells in lymphoma patients [58], our data support the exploration of lower doses of RMD for the activation of latent viral reservoirs in HIV-infected patients. Such modified dosing regimens could reduce or eliminate the perturbations of immune functions potentially associated with RMD treatment. In addition, treatment with agents that stimulate innate immunity such as TLR7/8 agonists could mitigate the suppressive effects of RMD on the activity of immune effector cells [58]. If RMD use eventually requires combination with immunostimulatory strategies to enhance antiviral immune effector functions and the clearance of latent HIV reservoirs, optimized temporal separation in the administration of RMD and immune stimulators could further minimize the potential of RMD-mediated interference with immune responses.
In summary, this report shows that RMD is a more potent and robust inducer of HIV expression in latently infected cells compared with other HDACi in clinical testing. This profile warrants the assessment of RMD in virally suppressed HIV-infected patients.
HIV-infected patients were enrolled into the study at the Quest Clinical Research (QCR; San Francisco, CA) and the University of Pittsburgh Medical Center (UPMC; Pittsburgh, PA). The QCR and the UPMC part of the study were approved by the Western Institutional Review Board and the University of Pittsburgh Institutional Review Board, respectively. In both cases, written, informed consent was obtained from the patients prior to any study procedures.
Romidepsin was obtained from a US-based pharmacy as a marketed commercial formulation of the drug. Vorinostat, panobinostat, givinostat, mocetinostat and SB939 were obtained from Selleck Chemicals (Houston, TX). Phorbol 12-myristate 13-acetate (PMA) and ionomycin were obtained from Sigma Aldrich (St. Louis, MO). All compounds were dissolved in DMSO. Effects of HDAC inhibitors in all assays and models used in this study have been compared to vehicle (i.e. DMSO) treated controls.
Plasmid pKS13 is a NL4-3-based vector in which the vpr and env genes were inactivated by inserting a ‘T’ base at the AflII site and two bases (TT) at the NdeI site, respectively, causing frame shifts in both open reading frames. A codon-optimized firefly luciferase gene was introduced in place of the nef gene by replacing the BamHI and NcoI fragment, yielding the plasmid pKS13. NL4-3-Luc virus was generated by co-transfection of HEK-293T cells with pKS13 and a plasmid containing the HIV-1 env gene using Lipofectamine 2000 (Life Technologies, Grand Island, NY).
Total peripheral blood mononuclear cells (PBMCs) were obtained from healthy HIV-negative donors by leukapheresis (AllCells, Inc, Emeryville, CA). Naive CD4+ T cells were purified by negative selection using EasySep magnetic beads (StemCells, Inc, Vancouver, Canada) and cultured in RPMI with 10% fetal bovine serum (FBS), penicillin/streptomycin, 1% nonessential amino acids (Life Technologies, Carlsbad, CA), 1% sodium pyruvate (Life Technologies) and 495 nM beta-mercaptoethanol (Sigma Aldrich, St. Louis, MO) in a 37°C, 5% CO2 incubator. Purified naive CD4+ T cells were activated by incubation with anti-CD3/CD28 magnetic Dynabeads (1 bead: 2 cells ratio, Life Technologies), 1 µg/ml anti-IL-4 (R&D Systems, Minneapolis, MN), 2 µg/ml anti-IL-12p70 antibodies (R&D Systems), and 10 ng/ml TGF-β (R&D Systems) for 3 days [29], [59]. Following the removal of anti-CD3/CD28 beads and antibodies, cells were maintained in 30 U/ml IL-2 (Life Technologies) for 2 days. Cells were then infected with NL4.3-Luc in the presence of 50 µg/ml DEAE for 3 hours. Cells were maintained in the continued presence of 30 U/ml IL-2 throughout the infection and subsequent rest period with culture medium with fresh IL-2 replaced every 2–3 days. Seven days post-infection, 20 µl of latently infected cells were dispensed into 384 well plates using a MicroFlo dispenser (Biotek Insturments, VT) at 10,000 cells/well containing 100 nl of compound solutions delivered by the Echo acoustic-based liquid dispenser (Labcyte, Sunnyvale, CA). After a 48-hour incubation, 16 µl/well BriteGlo (Promega, Madison, WI) was added and luminescence measured using the Envision plate reader (Perkin Elmer, Waltham, MA). Compound-associated cytotoxicity was determined in latently infected cells in parallel with the virus activation assay. Cells were incubated with compounds for 48 hours at 37°C, and cell viability was determined using Cell Titer Glo reagent (Promega).
Seven days post-infection, cells infected with NL4-3-Luc virus were stimulated for 48 hours with test compounds. Cells were stained with fixable viability dye v450 (eBioscience, San Diego, CA) in phosphate buffered saline (PBS) +2% FBS for 30 minutes at 4°C. Cells were fixed with IC Fixation Buffer (eBioscience) for 20 minutes at room temperature in the dark followed by treatment with a permeabilization buffer (eBioscience) for 30 minutes at 4°C. Intracellular p24 was stained with anti-p24 antibody clone KC57-RD1 (Beckman Coulter, Fullerton, CA), then washed and resuspended in PBS +2% FBS. The cells were analyzed on a LSR Fortessa (BD Biosciences, San Jose, CA) and collected data were processed using a FlowJo Analysis Software (TreeStar, Ashland, OR).
Determination of HDACi potency against individual HDAC enzymes was performed at Reaction Biology Corporation (Malvern, PA). In brief, human HDAC-1 to -11 enzymes were incubated with serial dilutions of HDACi and fluorogenic peptide substrates (50 µM) in a 96-well format assay. Concentrations of HDACi reducing the control activity of tested enzymes by 50% (IC50) were determined by a five-parameter curve fit of collected fluorescent signals. Fluorogenic peptide derived from p53 protein (amino acid residues 379–382; RHKKAc) was used as a substrate for HDAC enzymes 1, 2, 3, 6, 10, and 11. A different fluorogenic peptide from p53 (amino acid residues 379–382, RHKACKAC) was used as a substrate for HDAC-8. Fluorogenic peptide Boc-Lys(trifluoroacetyl)-AMC was used as a substrate for HDAC enzymes 4, 5, 7, and 9.
HIV-infected patients participating in the study were selected based on sustained plasma viral load suppression (<50 copies/ml for >12 months), CD4 counts (>350 cells/µL), and absence of co-infection with hepatitis B or C virus (Supplementary Table S1). Clinical laboratory results were reconfirmed 2 weeks before leukapheresis or blood draw. Leukapheresis was conducted for 3–4 hours, and samples were processed within 2 hours after collection. The leukapheresis product was diluted 1∶1 with PBS and layered over Ficoll for isolation of PBMCs. PBMCs were treated with red blood cell lysis buffer (eBioscience) and rested overnight (10 million cells/ml) in tissue culture medium (RPMI 1640 supplemented with 10% FBS and PenStrep) before the isolation of memory CD4 T cells according to the manufacturer's recommendation (EasySep Human Memory CD4 T cell Enrichment Kit). Resting CD4 T cells were isolated by first enriching total CD4 T cells (StemCell Technologies) and subsequently depleting HLA-DR-, CD25-, and CD69-positive cells via negative selection (Miltenyi Biotec, Auburn, CA). Flow cytometry was used to assess the purity of both T-cell subsets (>98%). Resting CD4 T cells for the analysis of continuous RMD exposure and longitudinal responses to RMD in the same donors were purified from fresh blood using the same protocol.
To assess HIV activation by HDACi, up to 5 million CD4 T cells were plated in 24-well plates in 2.5 ml of media, supplemented with antiretrovirals (ARVs) (100 nM elvitegravir and 100–300 nM efavirenz) for the entire duration of culture incubation. HDACi were added at specified concentrations on day 0 and cells or culture supernatants were harvested for the analysis of HIV RNA at the indicated time points (6–48 hours or 6–7 days, respectively). For HDACi pulse treatment, cultures were washed twice using ARV-containing media at 4 or 24 hours after addition of HDACi and incubated with ARVs until harvest. To measure HIV RNA levels, 1 ml of culture supernatant was analyzed by a robotic COBAS AmpliPrep/TaqMan system (Roche Diagnostics, Indianapolis, IN), which extracts total nucleic acid and quantifies HIV RNA in copies per milliliter using the HIV-1 Test, v2.0 kit (Roche Diagnostics). For the measurement of cell-associated HIV RNA levels, cells were washed with PBS, counted, and lysed (Qiagen RLT buffer, 400 µl for every 2 million cells; Qiagen, Venlo, Netherlands). Lysates were filtered through a Qiashredder (Qiagen). Total RNA was prepared from the lysates (400 µl) using a robotic system (QIAsymphony, Qiagen) that incorporates a DNase I digestion step to eliminate cellular DNA. The resulting total RNA was eluted with 200 µl of buffer, diluted to 1 ml with nuclease-free water, and analyzed by COBAS using the HIV-1 Test v2.0 kit. Preparations of both cell-associated RNA and supernatant total nucleic acids were tested for potential contamination with HIV DNA and/or host DNA by performing the PCR amplification in the presence and absence of reverse transcriptase and by a detection of GAPDH-encoding host DNA sequence. These methods confirmed that there was no contaminating HIV DNA in either the intracellular RNA or supernatant total nucleic acid preparations (Supplementary Figs. S3 and S4).
For cell activation analysis, either complete PBMC cultures or resting CD4+ T cells isolated from HIV-infected cART-suppressed patients were incubated with the tested agents under specified conditions and were stained with relevant antibodies. All antibodies were purchased from BD Biosciences and included Alexa Fluor 700-labeled antibody to CD4, APC-H7-labeled antibody to CD8, PE-Cy7 -labeled antibody to CD19, PE-labeled antibody to CD69, APC-labeled antibody to CD25, and V450-labeled antibody to HLA-DR. Live cells were gated by forward and side scatter and exclusion of the dead cells by Live/Dead Fixable Aqua Stain (Invitrogen). Marker staining was assessed by flow cytometry analysis on a LSR Fortessa with data processing using FlowJo software.
Resting CD4 T cells obtained from HIV-infected patients on suppressive cART were incubated with serial dilutions of HDACi in 96-well plates at 10,000 cells/well for indicated time and then lysed by repeated freezing and thawing. Total cellular HDAC activity was measured using the HDAC-Glo I/II assay kit (Promega) according to the manufacturer's protocol.
Single-genome sequencing of a portion of HIV-1 gag-pro-pol amplified from cell culture supernatants or cellular DNA extracts was performed as previously described [34]–[36]. Sequences were aligned using ClustalW and neighbor-joining phylogenetic analysis was performed using MEGA5. Trees were rooted on the subtype B consensus sequence (www.hiv.lanl.gov).
HIV RNA-induction experiments were conducted using 5 to 6 replicates for every experimental condition, including vehicle-treated controls. Fold induction in HIV RNA levels was calculated as a ratio of mean signals from compound-treated and vehicle (DMSO)-treated control samples from the same donor under identical conditions. Student's t-test (one-tailed distribution, two-sample equal variance) was used to assess statistical significance of the difference between the detected HIV RNA copies in vehicle control- and HDACi-treated samples or between RMD- and VOR-treated samples. Time course experiments testing the latency reversal agents had matching vehicle-treated controls for each time point. For experiments using blood-derived resting CD4 T cells, culture supernatants from 3 replicate wells were combined and analyzed using the COBAS system. Primary results in absolute HIV RNA copy numbers from all experiments presented in this study are provided in Supplementary material (Supplementary Table S4).
|
10.1371/journal.pcbi.1005641 | Classification and analysis of a large collection of in vivo bioassay descriptions | Testing potential drug treatments in animal disease models is a decisive step of all preclinical drug discovery programs. Yet, despite the importance of such experiments for translational medicine, there have been relatively few efforts to comprehensively and consistently analyze the data produced by in vivo bioassays. This is partly due to their complexity and lack of accepted reporting standards—publicly available animal screening data are only accessible in unstructured free-text format, which hinders computational analysis. In this study, we use text mining to extract information from the descriptions of over 100,000 drug screening-related assays in rats and mice. We retrieve our dataset from ChEMBL—an open-source literature-based database focused on preclinical drug discovery. We show that in vivo assay descriptions can be effectively mined for relevant information, including experimental factors that might influence the outcome and reproducibility of animal research: genetic strains, experimental treatments, and phenotypic readouts used in the experiments. We further systematize extracted information using unsupervised language model (Word2Vec), which learns semantic similarities between terms and phrases, allowing identification of related animal models and classification of entire assay descriptions. In addition, we show that random forest models trained on features generated by Word2Vec can predict the class of drugs tested in different in vivo assays with high accuracy. Finally, we combine information mined from text with curated annotations stored in ChEMBL to investigate the patterns of usage of different animal models across a range of experiments, drug classes, and disease areas.
| Before exposing human populations to potential drug treatments, novel compounds are tested in living non-human animals—arguably the most physiologically relevant model system known to drug discovery. Yet, high failure rates for new therapies in the clinic demonstrate a growing need for better understanding of the relevance and role of animal model research. Here, we systematically analyze a large collection of in vivo assay descriptions—summaries of drug screening experiments on rats and mice derived from scientific literature of more than four decades. We use text mining techniques to identify the mentions of genetic and experimental disease models, and relate them to therapeutic drugs and disease indications, gaining insights into trends in animal model use in preclinical drug discovery. Our results show that text mining and machine learning have a potential to significantly contribute to the ongoing debate on the interpretation and reproducibility of animal model research through enabling access, integration, and large-scale analysis of in vivo drug screening data.
| Testing potential therapeutic compounds in animal disease and safety models is a crucial part of preclinical drug discovery [1]. Although many in vitro methods have been developed to rapidly screen candidate molecules, no such simple assay system can recapitulate the complexities and dynamics of a living organism [2]. By contrast, an in vivo assay, depending on the animal species, allows a potentially far more realistic and predictive measure of a compound’s effect, and can capture the complexity of target engagement, metabolism, and pharmacokinetics required in the final therapeutic drug. Testing novel therapeutics in vivo is therefore most likely to accurately predict patient responses and successfully translate from bench to bedside [3]. In fact, a proof of efficacy and safety in animals is usually an essential requirement by regulatory agencies before progressing a compound into human studies [1, 4].
Drug efficacy tests are carried in animal models that mimic some aspects of human pathology. Based on how the disease state is created, animal models can be generally classified into three main groups [5]:
Regardless of the type of an animal model used, the main purpose of in vivo drug screening is to offer useful insights into human biology and to predict human responses to novel treatments [9]. High attrition rates in the clinic, however, show that animal studies do not always reliably inform clinical research for both drug efficacy and safety [10]. Many scientists have drawn attention to the need for more systematic, rigorous, and objective analysis of animal research data before designing studies in patients [11, 12]. In particular, the probability of successful model species to human translation should be assessed based on careful meta-analyses and systematic reviews that integrate results across all relevant animal studies [9]. These should involve experiments performed on different strains, species, and experimental models [9], since they might recapitulate distinct aspects of human disease and offer variable accuracy of predictions. Several such systematic reviews performed retrospectively exposed various challenges of successful translation including publication bias, design flaws, insufficient reporting of experimental details, and lack of coordination between scientists involved in animal research and those designing clinical trials [13–15].
Currently, systematic reviews are performed manually and involve analysis of large quantities of published articles and internal proprietary reports. Attempts to automate some aspects of this time-consuming process have mainly focused on systematic identification and ranking of potentially relevant articles with robust search filters [16, 17] and machine learning methods [18]. More recently, Flórez-Vargas et al. used text mining to analyze the full text of 15,000 research papers describing mouse studies across a diverse range of therapeutic areas [19]. In addition to exposing insufficient reporting of gender and age of laboratory mice, the study found evidence of sex bias across specific fields of research. These results demonstrate the ability of text mining to offer insights into large-scale emergent trends and weaknesses of animal research by systematically analyzing large unstructured datasets [19, 20]. One discussed limitation of the analysis was due to the fact that details of animal experiments are typically reported in the full text of articles (as opposed to abstracts), which can only be obtained for open access publications [19], ~24% of biomedical research literature [21].
Our own study aims to contribute to the ongoing debate on the need for the integration of animal research data, not least in enabling discovery and reuse of previous research. As a basis of our analysis of in vivo drug testing information, we use ChEMBL [22]–a drug discovery focused bioactivity database widely known for its large curated and consistently indexed in vitro bioassay datasets. Animal model data in ChEMBL include descriptions and results of more than 100,000 drug screening experiments in rats and mice—the most widely used animal model species. These have been manually extracted by database curators from the full text of scientific articles.
In contrast to the molecular target annotated in vitro content of ChEMBL, its in vivo screening data are currently understudied and arguably under-curated. This is very likely due to their relative complexity compared to in vitro bioassays, and their inherent abstracted, unstructured format. The in vivo information is encoded in textual assay descriptions, written by database curators and intended for expert human users, not for computational analysis. The descriptions take the form of compact summary accounts such as: “Inhibition of carrageenan-induced paw oedema in Sprague-Dawley rat at 5.16 mg/kg, sc after 3 hrs”. In less than twenty words, the example above summarizes important details of the screening system (strain of the animal, experimental stimulus, phenotypic readout) as well as compound administration details (dosage, timing, and administration route). Hence—despite their concise format, assay descriptions can be information rich and we consider they have future potential in translational drug discovery research. However, the variety of possible expressions used to describe the same assay makes comparison and clustering of in vivo screening data from resources such as ChEMBL extremely challenging.
In this paper, we present the first, to our knowledge, computational analysis of the entire ChEMBL in vivo assay description dataset in rat and mouse. We use natural language processing (NLP) methods to parse the descriptions of in vivo assays and then mine them for information connecting animal models to human disease, genetic strains, experimental treatments, and phenotypes. To this end, we apply an approach that leverages existing community-maintained and stable vocabularies alongside manually crafted extraction rules. To automatically organize the extracted information, we construct a “semantic space” of assay descriptions using a neural network language model, and build several random forest (RF) classifiers. Finally, we show that combining information mined from text with structured curated data offers new and useful insights into the in vivo dataset in ChEMBL as well as trends in the use of animal models in drug discovery research in general. We restrict our current analysis to animal models used in the evaluation of the efficacy of drugs, not animal model usage in ADME or toxicology studies—although a similar analysis strategy can be applied to these.
ChEMBL is a large open access database covering bioactivity information for about 1.6M compounds tested in 1.2M distinct bioassays. Publications on analysis and use of ChEMBL indicate that most users focus on protein-binding/biochemical data from in vitro experiments. However, as shown on Fig 1, more than half of the bioassays in ChEMBL corresponds to “higher-level” functional screening involving cell lines, tissues, and whole organisms. The last category includes laboratory rodents. Rats and mice represent the most commonly studied model organisms in ChEMBL, reflecting their central, historical and current, importance for preclinical drug discovery. Jointly, these two species were used as targets in 100,250 functional assays (77.7% of all animal-based experiments and 84.1% of experiments in mammals); see Fig 2.
As the input for our analysis, we selected all rat and mouse assay data that were extracted from scientific publications (as opposed to those that came from direct depositions or were loaded from PubChem Bioassay). In summary, the assay data we used come from the full text of 10,851 primary research articles published between 1976 and 2015 in 17 high-impact drug discovery and pharmacology journals. Each assay is summarized by a short description (mean of 20.7, and median of 20 words, see Fig 3) and a set of structured additional annotation fields including species name, molecular structures of compounds tested in the assay, and information about the associated publication including title, year, and journal. Although the total counts of distinct assays performed in rats and mice are similar (49,313 and 50,937 respectively), time frequency analysis shows that mouse is becoming increasingly used in recent years. Altogether, the assays involve 100,432 distinct compounds from all stages of drug discovery, including 1,215 molecules that have at least reached clinical development (based on max phase field in ChEMBL). 19,975 (20% of the total) assays involve approved drugs covering various drug classes and therapeutic areas; see Fig 4 showing the coverage of drug classes defined by Anatomical Therapeutic Chemical (ATC) classification [23].
Following preprocessing of the raw assay descriptions (described in Methods section), the first step in our analysis was to mine the descriptions for phrases representing animal disease models and genetic strains of mice and rats. Currently, there exists no dedicated software for this task although many systems have been developed for the recognition of other biomedical concepts, such as genes, cell lines, and diseases [24–26]. Similarly, there are no labelled datasets that could be used to train supervised machine learning models for animal model identification. Therefore, instead of a supervised approach, we explored dictionary and rule-based methods that make use of structured terminologies, syntactic information, and custom lexical patterns.
To identify genetic strains in text, we built two dictionaries based on specific community nomenclature guidelines [27, 28] and official strain listings maintained by public mouse and rat genome databases where new strains are registered [29, 30]. Each dictionary lists basic strains (1,307 mouse and 648 rat strains), together with the strain type (e.g. inbred or hybrid), available synonyms, and substrains.
For the task of identification of induced (experimental) disease models, where no such controlled vocabularies exist, we applied a method that identifies relevant expressions using a set of manually defined rules. These extraction patterns combine keyword matching with information about the structure of a sentence to capture animal models, often represented by multi-word phrases such as “maximal electroshock induced” or “high-fat diet-fed”. See Methods section for details on grammatical analysis of assay descriptions and noun phrase extraction.
Our next goal was to automatically organize and cluster concepts extracted in the previous step. This involved finding related entities—such as animal models of the same disease—identified based on linguistic patterns and contexts in assay descriptions. The underlying assumption is that two terms are likely to be related if they often occur in similar contexts, i.e. surrounded by the same or analogous phrases [31].
To find semantic similarities between words/phrases, we used the assay descriptions to train Word2Vec [32]–an unsupervised neural network (NN) model that converts text into a set of numerical vectors. These word vectors, also called word embeddings, correspond to points in a high-dimensional semantic space where distance correlates with differences in meaning. In other words, vectors representing semantically related words lie close to each other in the constructed semantic space, while unrelated words are far apart [33].
Next, we considered the problem of assay classification. Specifically, we tested whether features learnt by Word2Vec could be used to group together related in vivo assays and to predict the type of an assay based on its textual description. To train supervised classification models, we used the Word2Vec embeddings labeled with the curated information associated with compounds tested in individual assays. Specifically, many assays in ChEMBL involve well-known reference molecules (pharmacological standards/positive controls) used to calibrate and validate the resulting measurements on novel molecules. Commonly, such molecules are comparator approved drugs, whose biological activities and therapeutic effects are already well studied. These characteristics are summarized by Anatomical Therapeutic Chemical (ATC) classification codes—the most widely recognized drug classification system administered by the World Health Organization [23]. Here, we used ATC codes of involved drugs to link assays to various disease/therapeutic area indications.
To better understand how animal models were used in drug discovery, we have combined the results of our text mining analysis with the manually abstracted content of ChEMBL. Specifically, we used the curated information about compounds evaluated in each in vivo assay to link animal disease models with approved drugs that had been tested in them. Based on the relationships generated this way, we built and visualized a network representation of the ChEMBL in vivo data.
As shown on Fig 9, the drug-model network shows strong local clustering with several densely-connected modules that correspond to distinct disease classes including inflammation, cancer, epilepsy, and hypertension. The clusters bring together related drugs and animal models of related diseases. For instance, in the highlighted cluster, antidiabetic drugs (such as metformin or rosiglitazone) and lipid-modifying agents (e.g. gemfibrozil, fenofibrate) are connected through various animal models used in diabetes and obesity research. The disease models grouped in the cluster belong to different classes. Some of the models include diabetic animals whose condition was artificially induced in the laboratory: either through adjusted diet (e.g. high-fat diet and glucose load models) or through administration of toxic compounds such as streptozotocin (STZ) or alloxan—chemicals that destroy pancreatic cells thus dramatically reducing insulin production [7]. In addition to the experimental models, several spontaneous (genetic) models can be found in the cluster. For instance, db/db mice and Zucker Diabetic Fatty (ZDF) rats develop symptoms similar to human diabetes due to a mutation in the Lepr gene, which encodes the receptor for a “satiety hormone”, leptin [7]. Finally, one of the smaller nodes in the network corresponds to a transgenic model: a genetically engineered mouse model expressing high levels of human apolipoprotein A-I (APOA1).
In addition, we combined the extracted animal model information with curated species annotations from ChEMBL to investigate differential usage of mouse and rats in drug discovery. First, we compared organism annotation across all assays involving different experimental models; representative results are shown on Fig 10A. Next, we divided assays into classes corresponding to different disease areas (based on the ATC codes of involved approved drugs) and found species distribution for the most frequent indications (Fig 10B). The figures show that whilst some screening experiments are routinely performed on animals of different species, in some cases the rat might be preferred to mouse, or vice versa. As further discussed in the next section, the variation may be attributed to various factors such as anatomical and behavioral differences between the two rodents.
The ChEMBL in vivo data, which capture in an unbiased way reported bioactivities from animal-based drug screening experiments, are arguably the database’s least appreciated resource. However, there are several reasons why the dataset could be valuable for data-driven translational drug discovery research. Firstly, the ChEMBL in vivo data are unique: there are almost no other large publicly available resources for efficacy screening in animal models. Secondly, they have been derived from scientific articles published over the last forty years and, hence, they reflect long-term, community-wide trends in preclinical drug discovery. Finally, the information from these publications has already been extracted, filtered and condensed by specialists and trained curators. In some sense, it is more efficient to analyze the concise curated summaries resulting from these efforts than to attempt to mine the original publications directly.
To our knowledge, this is the first systematic analysis of the in vivo assay description data in ChEMBL and the first attempt to index and identify spontaneous and induced animal models from such a resource.
Although most information on the in vivo assays in ChEMBL is in the form of unstructured assay descriptions, we show that it can be efficiently extracted and systematized using modern text mining techniques. In this work, we first identify mentions of animal models and phenotypes, and automatically organize them using a neural network language model. Then, we use the learned neural embeddings (word vectors) to train a random forest classifier predicting ATC codes of drugs tested in different assays. In the following, we further discuss our approach as well as its potential applications and limitations. In addition, we examine how the information extracted from assay descriptions reflects trends and practices in preclinical drug discovery.
For ChEMBL users, it is difficult to identify which in vivo assays correspond to a disease of interest. Some assays involve approved drugs with known indications, however these correspond to just 20% of the dataset. For the remaining 80% assays involving only novel, unannotated structures, there are not many options beyond a simple keyword search, which does not benefit from synonym mapping and in addition requires knowledge of animal models and phenotypes.
Here, we have shown that natural language processing (NLP) and neural language models can be used to automatically classify animal-based assays in ChEMBL based on the information encoded in the free text of assay descriptions. The approach could be further extended by incorporating models based on chemical similarity of associated molecules to known drugs or clinical candidates. Meaningful classification of animal-based assays would provide better access to the currently less widely analyzed in vivo screening data in ChEMBL by helping users find a subset of data that are most likely to be relevant.
Apart from application in the use of the ChEMBL resource, our work has applications in translational bioinformatics and data-driven drug discovery [50]. We demonstrated the first attempt to extract the mentions of animal models from text and produced a network of relationships between animal models and approved drugs. Given the functional relevance of in vivo screening and substantial differences between different model systems, we believe that bulk analysis of animal-based drug screening data could improve our understanding of biological activities of small molecules as well as the challenges of successful translation. In this regard, text mining techniques will play an important role in extracting and integrating animal model information since most relevant studies are disseminated in the unstructured format in scientific publications and patents.
Some of the limitations of this work arise from the shortcomings of the NLP workflow. Except for genetic strains, we do not normalize synonyms, although Word2Vec helped identify many related phrases for commonly used expressions (e.g. the metrazole—pentyneletrazole example discussed above). Furthermore, our dictionary-based NER methods favor precision, but may overlook terms that are not covered by the underlying vocabulary. On the other hand, the rule-based method used to identify mentions of experimental animal models include extraction patterns that are highly dataset-dependent and would have to be evaluated and potentially re-optimized before application to documents outside of the ChEMBL corpus.
Other limitations are due to the features of the underlying dataset, including its relatively limited size—larger corpus would certainly improve the performance of computational models we used. In addition, some therapeutic areas are underrepresented resulting in unbalanced training sets. Furthermore, descriptions frequently lack essential information about animal models, and some assays are currently misclassified.
In addition, some of our basic assumptions may be incorrect. For instance, our animal model—approved drug relationships are based on a co-occurrence assumption which might not hold in all cases since approved drugs might not always be active in disease models in which they were tested. Hence, we might have erroneously annotated some assays with incorrect ATC codes.
Testing drug candidate molecules in proven, or putative, animal models of mechanistic efficacy is an important part of all drug discovery programmes. Yet, there are almost no publicly available resources storing the results of historical in vivo compound screening. A notable exception is ChEMBL—a bioactivity database widely known, primarily for its ligand-protein binding data.
In this work, we demonstrate that in vivo assay data in ChEMBL are, despite their largely unstructured format, a valuable resource for direct use in data-driven drug discovery and optimization. We show that the descriptions of screening assays can be effectively and efficiently mined and classified using a combination of modern text processing techniques and neural language models, and that the extracted information, particularly when combined with the structured database content, provides fundamental insights into the inter-relationships of experimental models, drugs, and disease phenotypes. Finally, there is currently an active debate in the literature over the reproducibility of in vivo bioassay results and publication bias in reports of animal studies in general [51], approaches developed in our work have potential to further inform this debate.
The goal of this study was to mine the descriptions of animal-based bioassays for information about animal disease models and their role in drug discovery. To this end, we built a system that uses NLP techniques and machine learning models to extract relevant information from the ChEMBL in vivo bioassay description dataset and automatically organize extracted concepts as well as whole assay descriptions based on their semantic similarity.
We begin with preprocessing and grammatical analysis of the assay descriptions extracted from ChEMBL (the overview of this step is illustrated by Fig 11). We use the GENIA tagger to tokenize the sentences and to annotate the words with part-of-speech (POS) tags and other linguistic features. Next, we use custom grammatical patterns to chunk the descriptions such that noun phrases of optimal length and content are retrieved. To identify mentions of animal models and their phenotypes in text, we use a combination of dictionary and rule-based named entity recognition (NER) methods that consider lexical and syntactic patterns. Next, we build an unsupervised neural network model to convert text of the descriptions into numerical vectors, which we then use to cluster extracted concepts and to train random forest classifiers. Finally, we combine information extracted from text with curated fields in ChEMBL to associate animal models with approved drugs tested in them.
The first input for our system is a text corpus of experimental descriptions of in vivo assays involving animal models of human disease and physiology. To generate the corpus, we identified a subset of relevant experiments in ChEMBL (release 21 [52]) by selecting all functional whole organism-based assays in mouse and rat. There are 100,250 such assays in the database, each summarized by a concise description and a set of curated annotations. The latter include: species name, structures and ATC classification codes of compounds tested in the assay, and details of the publication in which the assay was reported.
To easily query and manipulate the dataset, we indexed the textual descriptions of selected assays and associated structured data with Elasticsearch—a distributed full-text search engine built on Apache Lucene [53]. All SQL queries used to retrieve the data and to calculate their statistics can be found in S2 Text; the data are available in S1 and S2 Datasets.
In the text of assay descriptions, complex multi-word noun phrases (NPs) are often used to represent meaningful concepts such as experimental stimuli (e.g. “high fat diet”), experimental readouts (“systolic blood pressure”, “insulin level”) or assay names (“tail flick assay”, “tail suspension test”). In this work, we use such phrases in the rule-based named entity recognition and as input for the language model. In order to extract NPs from text, we applied shallow parsing analysis (chunking [54]) with custom grammatical patterns. First, we assign each word in a preprocessed description its corresponding part-of-speech (POS) category (e.g. noun, adjective, verb). Next, we search for chunks of text corresponding to individual noun phrases using custom grammatical and lexical rules.
Prior to the syntactic analysis, we modified all assay descriptions as follows: we first normalized the species names, e.g. by substituting each mention of “Rattus norvegicus” with “rat”; next, we expanded commonly used acronyms of administration routes (e.g. sc—subcutaneous(ly)) and administration regimes (e.g. qd—daily) using a custom dictionary covering 20 common abbreviations (S3 Dataset).
We adopted two distinct procedures for the detection of relevant concepts in the text of assay descriptions. 1) To identify the names of genetic strains and phenotypes, we used a dictionary-based approach in which documents are matched against comprehensive lists of terms from existing ontologies and custom-built dictionaries. 2) For the detection of induced (experimental) disease models, we adopted a rule-based approach combining syntactic information and custom lexical extraction patterns. We describe the basic aspects of our NER workflow in the section below, while additional details and explanations can be found in S1 Text.
In the next step, we converted words and phrases from assay descriptions into numerical vectors that can be used to find semantic similarities between concepts and to train classification models.
To calculate distributed vector representations for the assay description dataset, we used an unsupervised neural network model, Word2Vec [33], implemented in the Python gensim module [76]. As input for the model, we used preprocessed assay descriptions tokenized into words and noun phrases (see: “Chunking and noun phrase extraction” section). For better results, we normalized the names of genetic strains mentioned in the descriptions, and removed numbers and standard English stop words (“the”, “in”, etc.) from the training sentences. The parameters of the model were set to standard values as follows: window (the maximum distance between the current and predicted word) = 5; minimum count (minimum word frequency) = 30; number of features (the dimensionality of resulting word embeddings) = 250. The output of the model was a set of 250-dimensional numerical vectors, each corresponding to a single word or phrase from the input corpus.
To find semantic similarity values for pairs of terms, we calculated the cosine distances between their corresponding vector embeddings [33]. We used analogous method to find similarities between entire assay descriptions. In this case, we first converted assay descriptions into numerical format by averaging their corresponding word vectors and normalizing the resulting mean representation vectors to unit norm [77, 78].
We visualized pairwise semantic similarities for a set of 35 most frequent animal models (with an exception of general purpose strains) and 35 phenotypic terms using a hierarchically clustered heatmap implemented in Python seaborn module for statistical data visualization [79].
To qualitatively analyze the distribution of functional assays, we projected the neural embeddings from 250-dimensional space into 2D using t-Distributed Stochastic Neighbor Embedding (t-SNE) [37]. To reduce the noise and speed up the computationally expensive t-SNE calculations, we first reduced the dimensionality of the vectors to 20 dimensions [37] using principal component analysis (PCA) [80].
Next, we considered the problem of assay classification. To train supervised models, we used the fact that many assays include reference molecules—mostly approved drugs with known indications. Specifically, we used the Anatomical Therapeutic Chemical (ATC) classification codes assigned to those drugs to annotate the assays and divide them into classes [23]. In most cases, we used the second level of the ATC hierarchy corresponding to the main therapeutic group of a drug, e.g. drugs used in diabetes (“A10”) or in epilepsy (“N03”). We considered different classification problems and built four distinct random forest models that predict assay class membership based solely on the textual information from the descriptions.
The first classifier predicts whether an assay involves any cidal/cytotoxic drugs. These are therapeutics whose main mechanism often involves causing cell death or inhibiting the growth of microbes): antineoplastic drugs (ATC code “L01”), antibacterials (“J01”), antivirals (“J05”), antiprotozoals (“P01”), etc. (see S1 Text).
The second model predicts whether an assay involves any drugs acting on the nervous system (ATC code “N”).
The third model is a multiclass classifier that assigns an assay with one of the five most common ATC code (level 2) combinations—a proxy for the most common disease areas in ChEMBL. In order of frequency these are: antiepileptic drugs (ATC code “N03”), psycholeptic drugs (“N05”), antineoplastic drugs (“L01”), drugs used in diabetes (“A10”), and anti-inflammatory drugs (combination of 4 ATC codes: “C01”, “M01”, “M02”, “S01”). The last combination of ATC codes is assigned to a nonsteroidal anti-inflammatory drug, Indomethacin, which is very commonly used as reference standard in the models of inflammation and pain.
The fourth classifier predicts specific drug classes for assays involving drugs acting on nervous system. The six class labels are: antiepileptics (“N03”), psycholeptics (“N05”), analgesics (“N02”), psychoanaleptics (“N06”), antiparkinsonians (“N04”), and anaesthetics (“N01”).
Dividing assays into classes is not a straightforward task since some assays might involve multiple drugs and, in addition, one drug might be assigned multiple ATC codes. To deal with this complexity, we selected a subset of assays that can be unambiguously assigned to one of the classes in every classification problem. For the details on class assignment and number of assays allocated to different classes, see S1 Text.
We applied two different methods for splitting the data into ten subsets used for training and testing in the 10-fold cross-validation procedure. In the first method, we split all assays randomly into equally sized subsets; in the second method, we partitioned the assays by randomly splitting the documents (scientific publications) from which the assay data were curated. This is important since the descriptions of assays reported in the same document are often very similar; commonly, they correspond to the same experiments that differ solely in dose or timing details. The second splitting method assured that such assays would never be used for both model training and testing, which might result in classifiers with overly optimistic performance and poor generalization to new data. See [38] for in-depth discussion on splitting assay data for QSAR model building.
We trained each random forest (RF) model with a set of vectors calculated for assay descriptions as average of their corresponding word embeddings (see previous section). Each RF classifier consists of an ensemble of decision tree estimators trained on an equally sized dataset sample drawn with replacement. The final classification is given by averaging probabilistic predictions of individual decision trees. We set the number of tree estimators to 200; we used adjusted class weights to reduce the impact of dataset imbalance (class_weight parameter set to “balanced”). Other parameters remained set to default values of the Python scikit-learn implementation [81]. Following 10-fold cross-validation procedure, we calculated standard performance measures (precision, recall, accuracy, F1-score), confusion matrix, and out-of-bag estimate (OOB) [82] for each model. Additional details on performance and comparison with other text-based methods (paragraph2vec and bag-of-words with TF-IDF weighting) are reported in S1 Text.
To generate a network of drugs and animal models, we combined the information extracted from assay descriptions with curated compound data from ChEMBL. For a subset of assays involving approved drugs, we created a dataset of animal disease models, including experimental and transgenic models as well as genetic strains. From the latter, we excluded general purpose models (such as Wistar rat or Swiss mouse), which are indiscriminately used for screening compounds for diverse indications (leaving only strains that spontaneously develop disease-related phenotypes). In addition, we manually consolidated synonymous names of experimental disease models; for instance “maximal electroshock”-“maximum electric shock”, “pentylenetetrazole (PTZ)”-“pentylenetetrazole”, and “hot plate”-“hotplate” are pairs of equivalent terms which were merged together. To create the network, we linked animal models to approved drugs such that a drug is connected to an animal model if it was tested in at least five assays involving this model. This resulted in a bipartite graph with 554 nodes and 710 edges, which we visualized using the Gephi network visualization software [83].
|
10.1371/journal.pntd.0005669 | Identification of antigenic Sarcoptes scabiei proteins for use in a diagnostic test and of non-antigenic proteins that may be immunomodulatory | Scabies, caused by the mite, Sarcoptes scabiei, infects millions of humans, and many wild and domestic mammals. Scabies mites burrow in the lower stratum corneum of the epidermis of the skin and are the source of substances that are antigenic or modulate aspects of the protective response of the host. Ordinary scabies is a difficult disease to diagnose.
The goal of this project was to identify S. scabiei proteins that may be candidate antigens for use in a diagnostic test or may be used by the mite to modulate the host’s protective response.
An aqueous extract of S. scabiei was separated by 2-dimensional electrophoresis and proteins were identified by mass spectrometry. A parallel immunoblot was probed with serum from patients with ordinary scabies to identify IgM and/or IgG-binding antigens. The genes coding for 23 selected proteins were cloned into E. coli and the expressed recombinant proteins were screened with serum from patients with confirmed ordinary scabies.
We identified 50 different proteins produced by S. scabiei, 34 of which were not previously identified, and determined that 66% were recognized by patient IgM and/or IgG. Fourteen proteins were screened for use in a diagnostic test but none possessed enough sensitivity and specificity to be useful. Six of the 9 proteins selected for the possibility that they may be immunomodulatory were not recognized by antibodies in patient serum.
Thirty-three proteins that bound IgM and/or IgG from the serum of patients with ordinary scabies were identified. None of the 14 tested were useful for inclusion in a diagnostic test. The identities of 16 proteins that are not recognized as antigens by infected patients were also determined. These could be among the molecules that are responsible for this mite’s ability to modulate its host’s innate and adaptive immune responses.
| Scabies, caused by the mite, Sarcoptes scabiei, infects millions of humans, and many wild and domestic mammals. Scabies mites burrow in the lower stratum corneum of the epidermis of the skin and are the source of substances that are antigenic or modulate aspects of the protective response of the host. Ordinary scabies is a difficult disease to diagnose. We identified 50 different proteins produced by S. scabiei, 33 of which bound IgM and/or IgG from the serum of patients with ordinary scabies. A set of 23 recombinant proteins were produced and screened for use in a diagnostic test but none possessed enough sensitivity and specificity to warrant further consideration although some could be among the molecules that are responsible for this mite’s ability to modulate its host’s innate and adaptive immune responses.
| Scabies is a worldwide disease that affects millions of humans, other species of primates, and many wild and domestic mammals. It is caused by the itch mite, Sarcoptes scabiei, that burrows in the lower stratum corneum of the epidermis of the skin. Scabies mites are the source of substances that modulate certain aspects of the inflammatory innate and adaptive immune response of the host allowing it to evade detection by the host until it is able to establish a thriving population [1–12]. Ordinary scabies is a difficult disease to diagnose and there are no diagnostic blood tests with adequate sensitivity and specificity available to identify patients early in the course of an infection [13].
The goal of this project was to identify S. scabiei proteins that (1) may be candidate antigens for use in a diagnostic test or (2) may be among those used by the mite to modulate the host’s protective responses.
Serum from patients with confirmed ordinary scabies was collected under Human Subjects Protocol (HSP) #0205 as approved by the Wright State University Institutional Review Board (IRB). All patients were adults and all provided written informed consent. Negative control sera were previously provided to us without personal identifiers under protocol SC #2714 approved as EXEMPT under CFR 46.101(b)(4) by the Wright State University IRB.
An aqueous extract of Sarcoptes scabiei var. canis was prepared by homogenizing mites in endotoxin-free water as previously described [14]. Following two 24-hr extractions, the supernatants were collected by centrifugation, sterile-filtered (0.22 μm) into sterile vials and stored at 4°C. The protein content of this and all other samples was determined using the method of Bradford with bovine serum albumin (BSA) as standard [15].
Unless otherwise noted, the materials used for protein separation and analysis were obtained from Bio-Rad Laboratories, Inc., Hercules, CA.
Proteins in the S. scabiei extract (40 mL containing 175 mg protein) were concentrated using preparative isoelectric focusing (IEF) as previously described [16] using a Bio-Rad Rotofor apparatus with ampholytes of pH 3–10 (BioLyte 3/10, 2% wt/vol final) and 5% glycerol. Focusing at 5°C for 5 hr at 12 W yielded 20 fractions with pH 1.6–13. Fractions 4–15 (pH 4–8 containing ~ 120 mg protein) were recombined and subjected to a second IEF separation. Fraction 14 had the highest protein concentration (2.2 mg/mL) and a pH of 5.0 and was selected for further study.
Two-dimensional (2D) gel electrophoresis was performed as previously described [14]. An aliquot of Fraction 14 was prepared using the ReadyPrep 2-D Cleanup Kit and the resulting protein sample was extracted into ReadyPrep Rehydration/Sample Buffer. Two identical samples, each containing ~200 μg of protein, were loaded onto 11 cm ReadyStrip pH 5–8 IPG strips using overnight passive rehydration. Second dimension separation was carried out using Criterion TGX Any kD precast gels as before. At the conclusion of the electrophoretic separation, one gel was stained with GelCode Blue Stain Reagent (Thermo Scientific, Rockford, IL). The other gel was prepared for electrophoretic transfer.
Following 2D separation, the proteins on the second gel were transferred to an “Immun-Blot PVDF Membrane for Protein Blotting” using condition as previously described [17]. PBST, composed of Dulbecco’s Phosphate Buffered Saline + 1% Tween 20, was used as wash. BPBST (PBST+ 1% BSA + 1% normal goat serum) was used to block the membranes and for antibody dilutions except as noted.
A pool of serum from patients with confirmed ordinary scabies was prepared by combining equal volumes of 5 individual serum samples [18]. The serum pool was diluted 1/60 and used to probe the blot for 2 hrs. For IgM binding, the blot was probed for 1 hr in biotinylated-Goat anti-Human IgM at 1/5000 and 1hr in streptavidin-Alkaline Phosphatase at 1/5000 (both from Southern Biotechnology Associates, Birmingham, AL). Tris-buffered saline (TBS) replaced PBS in wash and diluent prior to the Alkaline Phosphatase step. The blot was developed using AP Blue Membrane Substrate (Sigma-Aldrich, St. Louis, MO) yielding blue spots where IgM bound. The blot was imaged and subsequently re-probed for IgG binding using biotinylated-Goat anti-Human IgG at 1/5000 and streptavidin-Horseradish Peroxidase at 1/5000 (Southern Biotechnology Associates). IgG binding proteins were stained reddish-brown using the substrate of Young [19]. Proteins that bound both IgM and IgG appeared purplish on the finished blot.
Both the stained gel and probed immunoblot were imaged and the images were overlaid with a 1,000-cell grid (25 row x 40 cells/row) as described before [14]. This allowed each stained protein spot on the gel and on the corresponding blot to be assigned a unique “spot number” identifier.
Ninety-seven blue-stained spots were excised from the gel using a 1-mm spot picker, collected into labeled LoBind tubes (Eppendorf, Westbury, NY) and frozen. Samples were shipped to Applied Biomics (Hayward, CA) for trypsin digestion and sequencing by mass spectrometry. Proteins were identified by MASCOT (Matrix Science, London, UK) search of the National Center for Biotechnology Information non-redundant database (NCBInr) with taxonomy restricted to “Sarcoptes scabiei”. This database contains the complete genome and predicted proteome for S. scabiei var. canis [20].
Gene sequences for selected proteins were synthesized by GenScript (Piscataway, NJ) with the open reading frame being codon-optimized for expression in E. coli. Additional modifications to the open reading frame were made to eliminate any internal BamHI, HindIII, and KpnI restriction sites. The termini of each gene contained in-frame 5' BamHI and 3' HindIII restriction sites for cloning into the pET-45b(+) expression vector. Expression vectors were transformed into E. coli Rosetta(DE3) competent cells (EMD Millipore, Billerica, MA). Transformants were selected on ampicillin-containing solid media plates, and 3-mL overnight liquid cultures were generated from five separate single colonies. The overnight cultures were incubated in liquid LB media that included ampicillin. All five cultures were then combined the following morning and subcultured into 500 mL of LB media without ampicillin for 3 hrs, followed by induction of protein expression by the addition of 1 mM (final concentration) IPTG for three hours. All liquid cultures were maintained in a MaxQ 4000 orbital shaking incubator (Thermo, Waltham, MA) shaking at 250 rpm and held at 32°C. Cells were harvested by centrifugation at 5000 x g for 20 min. Cell pellets were stored at -80°C until protein purification. Frozen cell pellets were resuspended in 10 mL of ice cold 1x Tris Buffered Saline (25 mM Tris, 150 mM NaCl, pH 7.2) containing Pierce Protease Inhibitor without EDTA (Thermo, Waltham, MA). Resuspended cells were disrupted by sonication on ice using 10 pulses of 30 sec on, 30 sec off with a 4710 series ultrasonic homogenizer (Cole Parmer, Vernon Hills, IL) set at 40% amplitude. Cellular debris was pelleted by centrifugation, and the supernatant was filtered through a 0.4 μm syringe filter. His-tagged proteins were then purified by column purification on Pierce His Pur Cobalt chromatography columns (Thermo) according to the manufacturer's recommendations and using a final elution volume of 3 mL. Purified proteins were quantified and analyzed as follows.
All recombinant proteins were subjected to an initial immunoblot screening. Aliquots of purified proteins (3–10 μg) were loaded onto the single prep-well of Mini-Protean TGX Any kD Gels and electrophoresis was carried out at 200 V as recommended by the manufacturer (BioRad) and as described previously [21, 22]. Separated proteins were then transferred to PVDF membranes that were blocked as described above.
Ten pools of serum from patients with confirmed scabies infestations were prepared based on prior assessment [13, 18]. A pool of serum from healthy control subjects was included as a negative control. All proteins were tested with these sera. Eight proteins of interest were also screened with serum from 30 individual patients with ordinary scabies and 10 uninfested controls [13, 18].
Blots were loaded into a mini-slot blot apparatus (Mini-Protean Multiscreen, BioRad) [21, 22] and probed for 2 hrs with the sera as described above. After removal from the slot blot apparatus, blots were sequentially developed for IgM and IgG binding as described above.
The purity of all individual proteins was also determined using electrophoresis on Mini-Protean TGX Any kD Gels run as above and stained with GelCode Blue.
Our previous analysis revealed that most of the soluble proteins present in an aqueous extract of scabies mites had isoelectric points (pIs) in the range of pH 5–8 [14]. In the present analysis, we used preparative IEF to concentrate proteins with pIs in this vicinity and then used IPG strips of pH 5–8 for final separation.
Ninety-seven protein-containing spots were excised from the GelCode Coomassie blue stained gel and were submitted for sequence analysis (Fig 1). All 97 spots were identified as containing one or more proteins of S. scabiei var. canis (Table 1). There were a total of 50 different S. scabiei proteins identified and 34 of these had not been previously reported (Fig 1, Table 1). The proteins from an identical gel were transferred to a PVDF membrane that was probed using a pool of sera from 5 patients with confirmed ordinary scabies infections that had previously been determined to have high levels of circulating antibodies that recognized antigens in S. scabiei extracts [13, 18]. Of the 97 protein-containing spots, one bound only IgM, 32 bound only IgG and 29 bound both IgM and IgG (Fig 2, Table 1). No antibody bound to 33 of the spots.
We previously postulated that a diagnostic test for scabies would require identifying a set of antigens that selectively bind antibody (especially IgM) from the serum of patients suspected of being infected with scabies mites [13]. Based on the antibody-binding profiles of the proteins identified on the 2D gel and blot, we selected 14 proteins for further study as diagnostic antigen candidates (Table 2). We also selected 9 additional proteins from the > 150 previously-identified proteins that could be among the molecules that are responsible for this mite’s ability to modulate its host’s immune responses (Table 2) [12, 14]. The genes coding for these 23 proteins were deduced from the S. scabiei var. canis genome [20], chemically synthesized, and cloned into E. coli. The recombinant proteins were expressed and partially purified before being subjected to immunoblot screening. GelCode Coomassie blue stained gels showed that the purity of the recombinant proteins ranged from 10% to 95%.
For the first round of immunoscreening, ten pools of serum from patients with confirmed scabies infections were prepared based on prior screening [13, 18]. Another pool of sera from uninfected subjects was also prepared to serve as a negative control. Each of the 23 proteins was then screened with these 11 serum pools by slot blot. Eight of the 14 proteins that were selected as diagnostic antigen candidates were recognized by antibodies in ≥ 50% of the test serum pools (Table 2). None of the candidate immunomodulatory proteins bound antibodies in more than 20% of the test sera.
The 8 most promising diagnostic antigen candidates were subjected to a second round of screening using the serum of 30 individual US patients with ordinary scabies and 10 uninfected controls [13, 18]. One protein (KPM11752), a hypothetical protein that appears unique to scabies mites, was recognized by antibodies present in the serum of all scabies patients and all control subjects (Table 3). Three other proteins (KPM03215, KPM07763 and KPM10468) bound antibodies present in the serum of 40–67% of the scabies patients but they were also recognized by 10–40% of the control sera. Two of these are homologs of the Group 13 and 25 dust mite allergens. The remaining candidate proteins were recognized by ≤ 30% of the serum from the scabies patients.
This research builds on previous proteomic work by identifying 50 different proteins produced by S. scabiei, 34 of which were not identified previously [14]. We determined that 66% of the protein-containing spots were recognized by IgM and/or IgG that is circulating in the serum of patients with ordinary scabies at the time of initial diagnosis and selected 14 of these for screening as candidates for use in a diagnostic test for scabies. Additionally, we identified 33 protein-containing spots, representing 16 different proteins, that were isolated from a Coomassie blue stained gel that did not bind patient antibody. Included among this set of proteins may be molecules that participate in the parasite’s immune evasion mechanisms and are responsible for modulating the host’s immune responses [1–12].
Unfortunately none of the 14 proteins selected as potential diagnostic antigens shows enough promise to warrant further study. Only two proteins (KPM07763 and KPM11752) had sensitivities of ≥ 67% but neither offered a specificity of > 40% (Table 3). An additional 19 different antibody-binding proteins were identified on the 2D immunoblot and these are also potential candidates for use as diagnostic antigens (Table 1, Fig 2). It is possible that screening of these proteins could yield candidates promising better sensitivity and specificity than those reported here.
Perhaps more interesting are the data for the 9 proteins selected for the possibility that they may be immunomodulatory. All were identified in a previous study [14] or were predicted from the genome [20] and none were detected on the 2D immunoblot (Fig 2). Among these were calmodulin-, calreticulin- and cystatin-like proteins, all of which have been shown to be produced by other parasites and to possess immunomodulatory properties [23, 24]. For a protein to be effective in assisting the mite to evade the host’s immune response it would likely also be able to escape detection by the adaptive immune system and would not elicit an antibody response. As would be expected, 3 of the 9 proteins tested were recognized by antibodies in the serum of ≤ 20% of the scabietic patients while the other 6 did not bind any antibody. A logical next step would be to test these proteins for their immunomodulatory properties, although this was beyond the scope of the present study.
Thirty-three proteins that bound IgM and/or IgG from the serum of patients with ordinary scabies were identified. Although none of the 14 tested are useful for inclusion in a diagnostic test, the identity of 19 other candidates is provided. The identity of 16 proteins that are not recognized as antigens by infected patients was also determined. These could be among the molecules that are responsible for this mite’s ability to modulate its host’s innate and adaptive immune responses.
|
10.1371/journal.pgen.1008107 | Risk of spontaneous preterm birth and fetal growth associates with fetal SLIT2 | Spontaneous preterm birth (SPTB) is the leading cause of neonatal death and morbidity worldwide. Both maternal and fetal genetic factors likely contribute to SPTB. We performed a genome-wide association study (GWAS) on a population of Finnish origin that included 247 infants with SPTB (gestational age [GA] < 36 weeks) and 419 term controls (GA 38–41 weeks). The strongest signal came within the gene encoding slit guidance ligand 2 (SLIT2; rs116461311, minor allele frequency 0.05, p = 1.6×10−6). Pathway analysis revealed the top-ranking pathway was axon guidance, which includes SLIT2. In 172 very preterm-born infants (GA <32 weeks), rs116461311 was clearly overrepresented (odds ratio 4.06, p = 1.55×10−7). SLIT2 variants were associated with SPTB in another European population that comprised 260 very preterm infants and 9,630 controls. To gain functional insight, we used immunohistochemistry to visualize SLIT2 and its receptor ROBO1 in placentas from spontaneous preterm and term births. Both SLIT2 and ROBO1 were located in villous and decidual trophoblasts of embryonic origin. Based on qRT-PCR, the mRNA levels of SLIT2 and ROBO1 were higher in the basal plate of SPTB placentas compared to those from term or elective preterm deliveries. In addition, in spontaneous term and preterm births, placental SLIT2 expression was correlated with variations in fetal growth. Knockdown of ROBO1 in trophoblast-derived HTR8/SVneo cells by siRNA indicated that it regulate expression of several pregnancy-specific beta-1-glycoprotein (PSG) genes and genes involved in inflammation. Our results show that the fetal SLIT2 variant and both SLIT2 and ROBO1 expression in placenta and trophoblast cells may be correlated with susceptibility to SPTB. SLIT2-ROBO1 signaling was linked with regulation of genes involved in inflammation, PSG genes, decidualization and fetal growth. We propose that this receptor-ligand couple is a component of the signaling network that promotes SPTB.
| Worldwide, more than 10% of babies are born prematurely without effective means of prevention. Premature birth is associated with mortality and lifelong comorbidities. Aggregation of spontaneous preterm birth in certain families suggests that there are underlying genetic factors that predispose to preterm birth. Both maternal and fetal genomes likely affect susceptibility. We set out to identify fetal genetic variants that predispose infants to premature birth in a population of Finnish origin. Our results from a genome-wide association study indicate that a variant of slit guidance ligand 2 (SLIT2) is associated with the risk of spontaneous preterm birth. Furthermore, SLIT2 and its receptor roundabout guidance receptor 1 (ROBO1) are expressed in placental cells, and their mRNA levels are higher in placentas from spontaneous preterm deliveries compared to term controls. Based on gene knockdown experiments in cultured placental tissue–derived cells, ROBO1 regulates expression of pregnancy-specific beta-1-glycoprotein (PSG) genes and genes involved in inflammation. Thus, our results indicate that the fetal SLIT2 variant and expression of both SLIT2 and ROBO1 in placental cells are correlated with susceptibility to spontaneous preterm birth. We propose that this receptor–ligand pair is a component of the signaling network that promotes spontaneous preterm birth.
| Preterm live births that take place before 37 completed weeks of gestation and even as early as 22–24 weeks are a global problem. Up to 11.1% (15 million babies) of all births worldwide occur prematurely, and approximately 45–50% of them are idiopathic or spontaneous [1–3]. Complications caused by preterm birth are the most common cause of neonatal deaths and the largest direct cause of deaths of children <5 years of age [1,3]. The research focusing on spontaneous preterm birth (SPTB) has been complicated by etiological, pathophysiological, and genetic heterogeneities. Multiple events are associated with SPTB, either independently or in concert [4]. These include intrauterine inflammation, called chorioamnionitis, preterm premature rupture of fetal membranes (PPROM) and abnormal fetal growth relative to uterine size [5,6]. It is important to find new biomarkers for early detection of SPTB. Currently, our understanding of the early molecular pathways leading to SPTB is incomplete and there are no effective means to prevent SPTB. Knowledge of how maternal and fetal genomes contribute to the risk of SPTB could provide more personalized tools to prevent it [7].
Epidemiological studies have shown that both fetal and maternal genes affect fetal growth, birth weight, birth length, head circumference, and gestational age (GA) [8–11]. A recent study indicated that variants of the fetal and maternal genome independently affect normal variations in birth weight [12]. In addition, maternal and fetal genomes are also considered to affect the susceptibility to preterm birth and duration of pregnancy in general [13–15]. The intrauterine environment influences fetal growth, and adverse intrauterine events affect pregnancy length not only in elective preterm pregnancies but also in SPTB [11]. Genetic analysis of 244,000 Swedish births resulting in twins, full siblings, and half-siblings revealed that 13% and 21% of the variation in birth timing is explained by fetal and maternal genetic factors, respectively [9]. Overall, preterm birth is a phenotype with contributions from both, maternal and fetal genomes that may have separate contributions and together with environmental factors, interactively determine the outcome [7,13,16].
A recent study that included a population of > 40,000 women and replication cohorts of > 8000 women identified several common variants in EBF1, EEFSEC, and AGTR2 that showed associations with preterm birth at a genome-wide significance level [13]. In addition, other genome-wide association studies (GWAS) and SPTB genetic studies focused on mother [17–19] or infant [19–21] genomic signals have discovered genetic loci associated with preterm birth and gestational length. Studies that focused on fetal genomes have not revealed replicable associations between fetal genetic factors and SPTB.
Many pathways and cellular processes are reported to be associated with SPTB, including response to infection, regulation of inflammation, stress, and other immunologically mediated processes [3]. According to our current understanding, inflammatory pathways also have roles in the initiation of spontaneous term birth, as normal labor starts when there is a shift in signaling between anti-inflammatory and proinflammatory pathways in the myometrium. This shift appears to involve many chemokines such as interleukin 8 (IL8), cytokines such as IL1 and IL6, and contraction-associated proteins such as oxytocin receptor (OXTR), connexin 43 (CX43), and prostaglandin receptors [22]. Therefore, it is likely that changes in inflammation-associated pathways also contribute to preterm birth. Evidence from candidate gene studies supports the role of inflammation-related factors in SPTB. For example, polymorphisms of the genes encoding TLR4, TNF, IL1B, interferon gamma (IFNγ), IL6, and matrix metalloproteinases may be associated with increased risk of SPTB [5].
The aim of the present study was to use a GWAS to investigate fetal genetic variants that may predispose infants to SPTB in a homogeneous population of Finnish origin. A variant of slit guidance ligand 2 (SLIT2) had the most suggestive association with SPTB in the GWAS and in a genetic pathway analysis. Therefore, we characterized SPTB-associated expression of SLIT2 and its receptor ROBO1 in the placenta and subsequently conducted experiments with relevant placenta-associated cells.
In order to find fetal genetic factors associated with predisposition to SPTB, we analyzed polymorphisms encompassing the entire genome for associations with SPTB. After quality control, 247 infants born spontaneously preterm and 419 infants born at term remained for inclusion in the GWAS. We performed the analysis for both GA (quantitative trait) and SPTB (dichotomous setting). However, due to sample collection bias resulting in skewed GA distribution, we acknowledge that the results of the quantitative trait analysis should be interpreted with caution. Fig 1 summarizes the study workflow.
We detected several suggestive associations (p < 10−5) in the GWAS (Fig 2, Table 1). The two most promising regions were within the genes encoding SLIT2 (rs116461311, p = 1.6 × 10−6) and succinyl-CoA:glutarate-CoA transferase (SUGCT; rs57670997, p = 1.8 × 10−6). We also detected suggestive associations for GA as a quantitative trait (S2 Table, S1 Fig). SNP rs116461311 within the SLIT2 gene showed the most significant signal for GA (p = 3.1 × 10−7, S1 Fig). In addition to SLIT2, four regions showed suggestive signals both in the primary setting and in the GWAS of GA; these signals were within SUGCT, an intergenic region in chromosome 6 (nearest loci LOC105377949 and LOC107986634), and within the genes encoding anaplastic lymphoma receptor tyrosine kinase (ALK) and DLC1 Rho GTPase activating protein (DLC1).
We further analyzed very preterm and moderate-to-late preterm SPTB infants separately against term-born controls (S3 and S4 Tables). Three regions showed suggestive signals (p < 10−5) both in the primary setting and in the analysis of very preterm birth (S3 Table). These signals were within SLIT2, in an intergenic region on chromosome 2 (nearest genes THUMPD2 and SLC8A1), and within the region encompassing the EXOSC1, ZDHHC16, MMS19, and UBTD1 genes, which encode exosome component 1; zinc finger DHHC-type containing 16; MS19 homolog, cytosolic iron-sulfur assembly component; and ubiquitin domain containing 1, respectively. The minor allele of SLIT2, SNP rs116461311, was overrepresented (OR 4.06, p = 1.55 × 10−7) in very preterm-born infants (GA < 32 weeks) compared to term-born infants. In moderate-to-late preterm infants (GA 32–36 weeks), two regions showed suggestive signals (S4 Table) that were also evident in the primary analysis: an intergenic region on chromosome 3 (nearest loci LOC105377173 and ROBO1) and within ADAMTS14, which encodes ADAM metallopeptidase with thrombospondin type 1 motif 14.
We also studied associations with SPTB within the contexts of PPROM and no PPROM. There were separate suggestive associations with SPTB-PPROM and with SPTB without PPROM (S5 and S6 Tables). For the SLIT2 region, the effects were similar for infants born after PPROM (rs116461311, OR = 3.5, p = 3.0 × 10−5) and for those born after spontaneous onset of labor with intact fetal membranes (rs116461311, OR = 3.6, p = 1.3 × 10−5).
To investigate potential maternal transmission of the minor allele of SPTB-associated rs116461311 (in SLIT2), we checked the MAF of the variant in maternal samples. The frequency of C-rs116461311 was 0.059 in SPTB mothers (n = 230) and 0.035 in mothers with term delivery (n = 378). Transmission analysis was not feasible because of the low minor allele counts.
SNPs with suggestive association signals in the region of the top GWAS gene (SLIT2) were examined for association with SPTB in a European population (n = 9890) [20]. In a European replication population of very preterm and term-born controls, rs12503652 and rs79034379, which correlate with the best GWAS SNP of the SLIT2 region (rs116461311), were associated with SPTB (Table 2). The two SNPs were in high LD with one another (r2 = 0.85, D′ = 0.95) in the Finnish individuals in 1000 Genomes phase3 data.
To investigate if the SNPs within regions with the most promising signals in the GWAS (p < 10−5) have functional consequences, we screened the GTEx data to examine whether any of the suggestively associated SNPs colocalize with cis eQTLs; that is, whether they correlate with mRNA levels in the analyzed tissues. In the GTEx data, 26 of the SNPs with p < 10−5 for association with SPTB overlapped with significant eQTLs (Table 3). These SNPs were located within five different regions; the majority were in the region encompassing EXOSC1, ZDHHC16, MMS19, and UBTD1 (Tables 1, S3 and S5). These SNPs correlated with mRNA levels of the following genes in different tissues: ZDHHC16, MMS19, FRAT1, ANKRD2, UBTD1, and RRP12. Associated SNPs within the top GWAS region, SLIT2, were not associated with mRNA levels. According to the GWAS catalog, none of the suggestively associated SNPs had been significantly associated with any phenotype.
We further functionally annotated the SNPs (p < 10−4) within the SLIT2 region with HaploReg, v 4.1. Some of the most promising SNPs within SLIT2 (including the best associating variant rs116461311, as well as rs60126904, rs115707845, and rs16869667) were located within regions that contain histone marks and DNase-hypersensitive sites in several tissues. Furthermore, SLIT2 SNPs (rs60126904 and rs115707845) mapped to predicted enhancers in several cell types and tissues, including neuronal cells, different cells in the brain, lung, spleen, adipose cells, colon and duodenum smooth muscle cells, fetal lung and kidney, and fetal membranes (amnion). Thus, there is some evidence of the putative regulatory effects of SLIT2 SNPs in several tissues, including fetal tissues such as placenta.
To identify biological pathways associated with SPTB, we performed pathway analysis of the GWAS data to search for gene set enrichment. SPTB was associated with 16 Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways [corrected (S7 Table) and FDR adjusted p < 0.05, Table 4]. The most significant pathways were axon guidance (p = 8.6 × 10−10), focal adhesion (p = 6.6 × 10−7), and vascular smooth muscle contraction (p = 1.4 × 10−6). Axon guidance, the most significant pathway, included SLIT2 and the gene encoding its receptor, ROBO1. A Gene Ontology (GO) search revealed 35 GO terms associated with SPTB, with a false-discovery rate (FDR) of <0.05 (S8 Table). GO sets that included SLIT2 and ROBO1 are listed in S9 Table. The three most significant GO sets that included SLIT2 were retinal ganglion cell axon guidance, telencephalon development, and negative chemotaxis. The three most significant GO sets that included ROBO1 were telencephalon development, neuron recognition, and negative chemotaxis. These results led us to a more-detailed investigation of the roles of SLIT2 and ROBO1 in SPTB.
We detected suggestive association signals for SNPs in a region that encompasses SLIT2 and in a region downstream of ROBO1. Protein Slit2 binds to Robo proteins specifically and with high affinity [23,24]. Therefore, we analyzed the localization of SLIT2 and its receptor ROBO1 in human placenta by immunohistochemical staining of placentas from SPTB and spontaneous term birth (STB) with anti‐human SLIT2 and ROBO1 antibodies (Fig 3). Both SLIT2 and ROBO1 localized to cytotrophoblasts, syncytiotrophoblasts, and decidual trophoblasts. In addition, we observed strong ROBO1 and faint SLIT2 staining in capillary endothelial cells. We also detected both proteins in the basal and chorionic plates of the placenta. We did not see apparent differences in staining intensities or cellular localization between placentas from SPTB and STB. This indicates that SLIT2 and ROBO1 are expressed in the placenta at the interface between mother and fetus during pregnancy.
Immunohistochemistry demonstrated SLIT2 and ROBO1 in different types of trophoblasts in human placenta (Fig 3). To obtain more quantitative data about placental expression of these proteins, we analyzed SLIT2 and ROBO1 mRNA levels by qRT-PCR in samples collected from the basal and chorionic plates of placentas from SPTB (n = 23), STB (n = 23), and elective preterm birth (EPTB) (n = 34).
We first compared SLIT2 and ROBO1 expression levels between SPTB (n = 23) and STB (n = 23) placentas (Fig 4). Both SLIT2 and ROBO1 mRNA levels were higher in the basal plate of SPTB placentas (SLIT2 fold change [FC] = 1.679, SD = 0.667; ROBO1 FC = 1.387, SD = 0.670) compared to those of STB (SLIT2 p = 0.004, ROBO1 p = 0.013; Fig 4A). There were no differences in mRNA levels of SLIT2 (p = 0.173) and ROBO1 (p = 0.297) between the chorionic plates of SPTB and STB placentas (Fig 4B).
To explore the effects of mode of delivery and GA on SLIT2 and ROBO1 mRNA levels, we compared SPTB with EPTB placentas. SLIT2 and ROBO1 mRNA levels were significantly higher for SPTB (SLIT2 FC = 1.595, SD = 0.580; ROBO1 FC = 1.282, SD = 0.577) compared to EPTB (SLIT2 p = 0.005, ROBO1 p = 0.031) in basal plate samples (Fig 4A). There were no significant differences in these levels in EPTB and STB placentas (SLIT2 p = 0.216, ROBO1 p = 0.328; Fig 4A). These results suggest that higher SLIT2 and ROBO1 expression levels are associated with SPTB.
SLIT/ROBO are involved in many processes that involve cell migration, including axon guidance; thus, they could affect trophoblast cell invasion and decidualization. To this end, we looked at whether SLIT2 or ROBO1 expression in the basal plate of the placenta is associated with fetal growth. We compared mRNA levels with birth weight-for-GA Z-scores (weight Z-score), which included age and gender standardization of the infants. Deliveries with intrauterine growth restriction or other growth disorders were excluded. SLIT2 mRNA levels correlated with Z-scores (p = 0.023, rs = 0.351) in term and preterm fetuses delivered after spontaneous onset of labor (SPTB and STB samples together, S3 Fig). This suggests that SLIT2 expression is associated with variations in fetal growth.
To investigate potential functions of SLIT2 and ROBO1 in placental cells, we silenced SLIT2 and ROBO1 expression separately in the HTR‐8/SVneo human trophoblast cell line with small interfering RNAs (siRNAs) (Fig 5). qRT-PCR revealed that silencing percentages were 60% and 85%, for SLIT2 and ROBO1, respectively. Corresponding percentages revealed by RNA sequencing were 75% and 74% for SLIT2 and ROBO1, respectively. Next, we characterized the transcriptomes of trophoblasts in which SLIT2 or ROBO1 was silenced, as well as of cells treated with siRNA Universal Negative Control #1. Transcriptomic data analysis identified 14 upregulated (S10 Table) and 12 downregulated (S11 Table) genes after SLIT2 knockdown compared to samples treated with siRNA Universal Negative Control. The threshold was an FDR-adjusted p value of ≤0.01 and an FC of ≥2.0. By the same criteria, there were 216 upregulated (S12 Table) and 610 downregulated (S13 Table) genes after ROBO1 knockdown. KEGG pathway database analyses (S14 Table, S15 Table) identified the top pathways affected after SLIT2 and ROBO1 knockdown as inflammation-related pathways such as cytokine-cytokine receptor interaction (KEGG.ID 4060).
Far fewer genes were affected by SLIT2 knockdown than by ROBO1 knockdown, probably because transfection reagent alone upregulated SLIT2 mRNA levels up to 4-fold (p = 0.002) compared to untreated cells. Therefore, knockdown of SLIT2 expression brought SLIT2 mRNA levels close to the levels of intact cells. Transfection reagent by itself seemed to activate inflammation-related pathways. The ROBO1 expression level was not affected by transfection reagent.
ROBO1 knockdown particularly affected genes encoding membrane receptors and other membrane proteins. KEGG pathway analysis revealed that hematopoietic cell lineage (KEGG ID 4640) had the lowest p value (6.57 × 10−8) (S15 Table). ROBO1 knockdown affected 14 of the 42 genes in this pathway: KIT, IL7R, IL1R1, HLA-DRB1, IL1A, ITGB3, TFRC, ANPEP, IL1B, CD22, CD24, KITLG, CSF3, and CD14. SLIT2 (FC = 2.0, p = 0.005) was among the genes upregulated after ROBO1 knockdown (S12 Table). One of the gene families highly affected by ROBO1 knockdown was pregnancy-specific glycoproteins (PSG), a complex gene family that regulates maternal–fetal interactions [25]. Of the ten protein-coding human PSG genes, six were upregulated after ROBO1 knockdown. These data indicate that ROBO1 is an important regulator of the HTR8/SVneo cell transcriptome and suggest a role for ROBO1 in modulation of PSG gene expression. In addition, ROBO1 appears to have immunomodulatory functions in trophoblast-derived cells.
To verify our suggestive findings from the RNA sequencing data, we used qRT-PCR to analyze the effect of ROBO1 knockdown on expression levels of selected genes from a larger number of specimens. Because many inflammation‐related pathways were involved, we investigated TGFA, CXCL6, and total PSG expression levels (Table 5). TGFA was downregulated after ROBO1 knockdown, while CXCL6 was downregulated by both ROBO1 and SLIT2 knockdown. Select members of the PSG family were upregulated when ROBO1 was silenced (Table 5): PSG1, PSG2, PSG4, PSG6, PSG7, and PSG9. To verify, we measured total mRNA expression of different PSGs (PSG1, PSG2, PSG3, PSG4, PSG5, PSG6, PSG7, PSG8, PSG9, and PSG11) at the same time in one qRT-PCR reaction, as described previously [26]. We also tested the effect of knockdown on IL6, TNFA, and SRGAP3 (Table 5).
The results of qRT-PCR were generally in line with those of transcriptome sequencing (Table 5). Knockdown of ROBO1 downregulated mRNA expression of TGFA (S4A Fig). Knockdown of either ROBO1 or SLIT2 downregulated CXCL6, whereas PSGs were upregulated by ROBO1 knockdown, similar to the results of RNA sequencing. Inflammatory cytokines IL6 and TNFA were both downregulated after ROBO1 knockdown (S4B Fig). IL6 was also downregulated after SLIT2 knockdown. There was a trend toward downregulation of SRGAP3 by ROBO1 knockdown. Thus, the RNA sequencing and qRT-PCR results are in accordance and confirm that ROBO1 knockdown affects the expression of immune response–modifying genes in a cell culture model.
Both maternal and fetal genetic factors contribute to SPTB. Nevertheless, variants in the fetal genome associated with SPTB predisposition are not well known. Our aim was to identify fetal genetic factors associated with SPTB. To this end, we performed a case–control GWAS study in a Finnish population that is known to be relatively genetically homogeneous. Based on these results, we identified a SLIT2 variant as a plausible factor for SPTB susceptibility. This led to further investigations to define high-risk populations and characterize SLIT2 and ROBO1 expression in the placenta and in trophoblast cells.
The association of the SLIT2 variant with SPTB was strongest in the population of very premature births (<32 weeks gestation), and this association was replicated in a European population with SPTB fetuses from 24 to 30 weeks gestation. We detected association signals for SNPs in a region that encompassed SLIT2 and for SNPs in a region downstream of ROBO1, which encodes the receptor for SLIT2. SLIT2 and ROBO1 encode proteins of the SLT2-ROBO1 signaling pathway; previous studies have indicated that this pathway is associated with different types of pregnancy complications, including preeclampsia [27,28], impaired placentation of missed and threatened miscarriage in early pregnancy [29], trophoblast invasion, and vascular remodeling during ectopic tubal pregnancy [29,30]. Our hypothesis-free GWAS study provides evidence that these genes have a role in another pregnancy complication, SPTB.
We did not detect associations that would reach the stringent level of genome-wide significance (p<10−8). This may be due to a relatively small sample size in the GWAS, which was one of the limitations in the study. However, we did identify signals below the generally used threshold of suggestive association (p <10−5). Moreover, the SPTB-associated SNPs in the SLIT2 locus showed association in an independent data set. In the future, there is a need to validate our findings in larger data sets to detect signals that may have gone undetected in the current sample size.
Previous study presented that fetal de novo mutations in genes that are involved in brain development are associated with preterm birth [31]. In line with this notion, SLIT2-ROBO1 signaling has a well-documented role in axon guidance during the development of the nervous system [32]. Therefore, it appears that at least in part the same fetal genetic factors are involved both in the onset of preterm birth and in brain development. It is known that preterm birth increases the risk of compromised brain development [33,34]. How much of the risk can be explained by shared genetic risk factors remains to be determined.
KEGG pathway analysis of GWAS data showed that the top-ranking pathway was SLIT2/ROBO1 signaling–regulated axon guidance. Previous studies have indicated that brain and placental development may share common pathways [35–37]. Although SLIT and ROBO were originally identified as axon guidance cues, they interact in many other cellular processes, including regulation of cell migration, cell death, and angiogenesis. As such, they have an essential role in the development of tissues such as lung, liver, kidney, breast, and tissues of the reproductive system [38,39].
Trophoblasts cover a large portion of the placenta and have multiple roles in the maintenance of pregnancy. Invading trophoblasts have a critical function in biogenesis of the placenta [40,41]. Later during pregnancy, decidual trophoblasts may have a role in silencing immune cells in the decidua. The lining of placental villae consists of the syncytiotrophoblast layer and cytotrophoblasts, which have roles ranging from immune protection to uptake of nutrients from maternal blood [42]. In addition to their functions in biogenesis of the placenta and maintenance of pregnancy, trophoblasts may participate in labor induction [43]. Two‐thirds of early pregnancy failures may present with reduced trophoblast invasion [29]. SLIT-ROBO signaling may play autocrine and/or paracrine roles in trophoblast functions, such as differentiation and invasion, by influencing the migration of trophoblastic cells [29,44]. Thus, SLIT2-ROBO1 signaling may be involved in the pathogenesis of pregnancy failures via their effect on trophoblastic cell functions. Indeed, our immunohistochemical experiments demonstrated that villous and decidual trophoblasts from preterm and term placentas were strongly positive for SLIT2 and ROBO1. These results are in line with those of a previous study, which demonstrated that villous syncytiotrophoblasts express high levels of SLIT2 and ROBO1. The study also found that trophoblastic endothelial cells highly coexpress multiple SLIT ligands (SLIT2, SLIT3) and ROBO receptors (ROBO1, ROBO2, and ROBO4) in full-term placenta. Thus, SLIT-ROBO signaling may also have an important role in the regulation of normal placental functions [27].
An earlier study found that levels of several SLIT/ROBO mRNAs and proteins are higher in preeclamptic placentas compared to normal controls [27]. Our data indicate that mean SLIT2 and ROBO1 mRNA levels were higher in SPTB placentas compared to placentas from spontaneous term deliveries and to placentas from elective preterm births. Consequently, increased levels of SLIT2 and ROBO1 mRNAs were associated with SPTB. In addition, SLIT2 may have a role in term spontaneous labor, as SLIT2 mRNA and protein expression are decreased in the myometrium after spontaneous term labor [45].
In addition to the associations of both the SLIT2 variant and mRNA expression of SLIT2 and ROBO1 in basal plate of placenta with SPTB, SLIT2 mRNA levels in placentas were associated with the birth weight of fetuses born after spontaneous labor. A GWAS of beef cattle identified SLIT2 as a candidate gene that affects the weight of internal organs [46]. As the fetal growth and intrauterine distention negatively associates the duration of pregnancy [47] influence of SLIT2 on fetal size is a potential mechanism that remains to be studied as a cause of SPTB.
To understand more about the function of SLIT2 and ROBO1 in trophoblast cells, we silenced their expression separately in immortalized extravillous invading trophoblasts. Altogether, 26 and 826 genes were affected by SLIT2 or ROBO1 siRNA knockdown, respectively. The low number of genes affected by SLIT2 knockdown was probably because the transfection reagent alone upregulated SLIT2 mRNA levels compared to untreated cells. Consequently, knockdown of SLIT2 expression only brought SLIT2 mRNA levels back to the levels of control cells. However, the mRNA expression level of ROBO1 was not affected by the transfection reagent. Genes affected by ROBO1 knockdown were mostly related to infection, inflammation, and immune response. These results correspond with those of previous studies, suggested that members of the SLIT and ROBO families act as regulators of the inflammatory response [45,48]. Both pro‐inflammatory [45,49,50] and anti‐inflammatory [48,51] functions have been reported. Our results from invading trophoblast cells support a proinflammatory role for SLIT2-ROBO1 signaling, since the genes downregulated by ROBO1 knockdown included proinflammatory cytokines and chemokines such as IL1A, IL1B, CXCL8, CCL2, and CXCL6. It is widely acknowledged that IL1 in particular, as well as other proinflammatory cytokines and chemokines, is associated with preterm labor [52–54]. IL1B is a primary secretory product of the inflammasome and as such is thought to have central roles in initiation of preterm labor, such as in the induction of prostaglandin synthesis. Both polymorphisms of IL1A and IL1B, as well as increased levels of IL1B, are associated with preterm birth [19,54–56]. CXCL6 is increased in amniotic fluid from patients with preterm labor complicated by intra-amniotic inflammation and from patients with SPTB without intra-amniotic infection/inflammation [57]. We propose that in trophoblast cells ROBO1 has a role in regulation of proinflammatory mediators. Genes involved in vascular formation (vasculogenesis) or development (angiogenesis) were not affected by SLIT2 or ROBO1 knockdown in trophoblasts.
The PSG family was one of the immune response–associated gene families affected by ROBO1 knockdown. PSGs include ten placental trophoblast–synthetized glycoproteins that belong to the immunoglobulin superfamily [58]. Of the six PSGs upregulated by siRNA-induced knockdown, PSG1 was ranked among the top three upregulated genes (S12 Table). PSGs are essential in the maintenance of normal pregnancy [58]; thus, altered PSG expression patterns could influence pregnancy complications. Over the years, complications such as abortion, preeclampsia, intrauterine growth retardation, fetal distress, and preterm delivery have all been linked to low PSG levels [58–63]. As ROBO1 was upregulated in SPTB placentas and knockdown of ROBO1 upregulated expression of PSG genes, we propose that ROBO1 signaling is important in downregulation of the expression of PSGs. In addition, PSG1 activates TGF-B 1 and TGFB2 [25,64,65]; TGFB1 suggestively associated with SPTB [13] and TGFB2 prevented preterm birth in experimental inflammatory stress [66].
The innate immune response and inflammation contribute to labor and delivery, particularly in preterm pregnancies [3,22,67,68]. Upregulation of proinflammatory cytokines stimulates and potentiates uterine contractions in the myometrium [69–71]. In preterm labor and delivery, it is mostly inflammatory signals that spread to the placenta, fetal membranes, and fetal compartment. It is plausible that SLIT2 and ROBO1 expressed by trophoblasts are associated with SPTB via regulation of inflammation-related factors. SLIT2-ROBO1–guided activation and propagation of inflammatory mediators throughout the fetal–maternal trophoblast interface of the uterine wall would likely influence the tissues actively involved in labor and delivery. As knockdown of ROBO1 downregulated many of the genes that encode cytokines and chemokines, it is probable that upregulation of ROBO1 in SPTB placentas compared to term placentas would also affect expression of these genes. There is both epidemiological and experimental evidence that untimely expression of cytokines and chemokines by either fetal or maternal tissues upregulates the activity of mediators, which leads to premature initiation of the parturition process [72].
In conclusion, the GWAS detected fetal association signals for SPTB and duration of pregnancy in the vicinity of SLIT2 and ROBO1. SLIT2 and ROBO1 were upregulated in SPTB placentas, and further functional studies confirmed that this signaling pathway has a role in regulation of the pathways associated with infection, inflammation, and immune response in trophoblast-derived cells. These results suggest that SLIT2 and ROBO1 play specific roles in increasing susceptibility to SPTB. SLIT2-ROBO1 signaling is associated with complications in early pregnancy and it is possible that it influences invading trophoblasts during placentation. Based on the currently available evidence, we propose that activation of SLIT2-ROBO1 expression and signaling in trophoblast cells contributes to inflammatory and immune activation, which in turn leads to early labor and preterm birth.
The present studies received ethical approval from the participating centers (Oulu University Hospital 79/2003, 14/2010, and 73/2013). Informed consent was obtained from study participants or their parents.
Characteristics of the Finnish study populations are summarized in S1 Table. The discovery GWAS study population consisted of singleton SPTB and term infants sampled in Oulu and Tampere University Hospitals. The study subjects were recruited prospectively during 2004–2014 and retrospectively from the 1973–2003 birth diaries of Oulu University Hospital. For replication, we downloaded the summary statistics of a European population described in a recent study by Rappoport et al. [20] through ImmPort (http://www.immport.org/: SDY1205, DOI: 10.21430/M37N6PJEQT). This population includes 260 SPTB cases (139 male and 121 female infants) and 9,630 controls (4,055 males and 5,575 females). The cases were very preterm infants born between 25 and 30 weeks of gestation and were clinically defined as SPTB in 2005–2008. The control population consisted of adults, originally from the Health and Retirement Study (HRS) [73], who were matched for ethnicity with the European cases.
In the Finnish cohorts, SPTB was defined as birth occurring after spontaneous onset of labor at <36 completed weeks + 1 day of gestation. All medically indicated preterm births and deliveries that included known major risk factors were excluded. These criteria led to exclusion of preterm deliveries that involved the following conditions or characteristics: multiple gestation, preeclampsia, intrauterine growth restriction, placental abruption, polyhydramnios, fetuses with anomalies, clinical chorioamnionitis or acute septic infection in the mother, diseases in the mother that could influence timing of delivery, alcohol/narcotic use, and accidents. Term birth was defined as birth occurring at 38–41 weeks (38 wk + 0 d to 41 wk + 6 d) of gestation. The following conditions were used as exclusion criteria for the control population: multiple gestation, intrauterine growth restriction, placental abruption, polyhydramnios, fetuses with anomalies, and requirements for special care of the newborn. All control infants were from families with at least two term deliveries without any preterm deliveries in the family.
Umbilical cord blood, umbilical cord tissue, or saliva was obtained from the study subjects. Commercial kits were used to extract genomic DNA from blood (UltraClean Blood DNA Isolation Kit; MO BIO Laboratories, Inc., Carlsbad, CA, USA or Puregene Blood Core Kit; Qiagen, Hilden, Germany) and cord tissue (Gentra Puregene Tissue Kit, Qiagen). OraGene DNA collection kits (DNA Genotek, Ontario, Canada) were used for collecting saliva, and DNA was extracted with the prepIT-L2P kit (DNA Genotek). Genome-wide SNP genotyping was performed with the Infinium HumanCoreExome BeadChip (Illumina, San Diego, CA, USA) by the Technology Centre, Institute for Molecular Medicine Finland (FIMM), University of Helsinki.
Genome-wide SNP data were processed with PLINK, v. 1.9 [74]. SNPs with minor allele frequency (MAF) < 0.01, genotyping rate < 0.9, or deviation from Hardy–Weinberg equilibrium (p < 0.0001) were excluded. Individuals with > 0.1 missing genotypes were excluded. Identical by descent (IBD) clustering and multidimensional scaling (MDS) analyses were performed with a linkage disequilibrium–pruned SNP set; population outliers and close relatives (pihat > 0.2) were excluded. Prephasing of genotypes was performed with SHAPEIT2 [75], followed by statistical imputation with IMPUTE2 [76] using the 1000 Genomes Phase 3 variant set (October 2014) as the reference panel. Before association analysis, SNPs with impute info score < 0.8 or MAF < 0.05 in cases or controls were excluded. Altogether 6,778,521 SNPs or short insertions/deletions remained for analysis after these quality control steps.
Associations between SPTB or GA and SNPs were assessed with the frequentist test under the additive model with SNPtest, v. 2.5.2 [77]. After the primary analysis, the following subgroups of SPTB infants were assessed: (1) very preterm infants (GA 23–31 wk + 6 d), (2) moderate-to-late SPTB infants (GA 32 wk + 0 d to 36 wk + 0 d), (3) PPROM before onset of labor, and (4) no PPROM before onset of labor. To account for population substructure in the GWAS, the first two MDS dimensions were included as covariates. In the GWAS, the effect of population stratification was minimal (λ = 1.03). Gene set analysis (GSA)-SNP was used to search for gene set enrichment in pathway analysis [78]. We included only genotyped SNPs located within genes in this analysis to avoid the complicating effects of SNPs in linkage disequilibrium. R, v. 3.2.2 (https://www.r-project.org) was used to create Manhattan plots. LocusZoom [79] was used to create regional association plots.
We annotated SNPs with three approaches: (1) We used Genotype-Tissue Expression (GTEx) data to analyze whether the SNPs overlap with cis expression quantitative trait loci (eQTLs) [80]; 2) we screened whether the SNPs had been associated with any phenotypes in previous GWA studies using the GWAS catalog[81]; and (3) we assessed whether the SNPs were located within putative regulatory regions using HaploReg, v. 4.1 [82].
Samples from human placenta were collected at Oulu University Hospital during 2010–2016 as described [16]. The placental samples used in immunohistochemical staining and in qRT-PCR analysis of SLIT2 and ROBO1 expression were subject to similar inclusion criteria as the samples used in GWAS. The inclusion criteria of gestational age for preterm placental samples was from 25 weeks to 36 weeks+ 6 days and 39 weeks to 41 weeks + 6 days for term samples. The same conditions (multiple gestation, intrauterine growth restriction, placental abruption, polyhydramnios, fetuses with anomalies or requirements for special care of the newborn) as in GWAS were used as exclusion criteria for term controls. Spontaneous preterm samples had almost the same exclusion criteria except the population included few cases with chorioamnionitis or oligohydramnion. The control group of elective preterm samples included cases with various pregnancy complications like IUGR or pre-eclampsia resulting in elective preterm delivery without labor.
Specifically, in total, 18 placental samples were analyzed by immunohistochemistry. Twelve samples were from SPTB deliveries (GA from 25 wk + 2 d to 35 wk + 2 d), and six were from spontaneous term deliveries (GA from 39 wk + 4 d to 41 wk +1 d). Samples from both basal and chorionic plates were included in the study.
RT-qPCR was performed with 23 placental samples from SPTB (GA from 25 wk + 2 d to 36 wk + 0 d), 34 from elective preterm birth (EPTB) (GA from 25 wk + 1 d to 36 wk + 6 d), and 23 from spontaneous term birth (STB) (GA from 39 wk + 1 d to 41 wk + 6 d).
Localization of encoded proteins was visualized in placental tissues by immunohistochemical staining. Samples were embedded in paraffin and cut into 4-μm slices, deparaffinized, and rehydrated. Antigen retrieval was done in Tris-EDTA buffer. Endogenous peroxidase activity was blocked in blocking solution (Agilent, Santa Clara, CA, USA). Samples from the chorionic plate were incubated with mouse anti-human SLIT2 antibody (1:4000 dilution, PA5-3113; ThermoFisher Scientific, Waltham, Massachusetts, USA) or mouse anti-human ROBO1 antibody (1:2000, PA5-34931; ThermoFisher Scientific). Samples from the basal plate of the placenta were incubated in a 1:5000 dilution of mouse anti-human SLIT2 antibody and 1:1000 dilution of mouse anti-human ROBO1 antibody. Bound antibodies were detected with the Envision kit (Agilent).
Tissue samples were homogenized, RNA was isolated with the RNeasy Mini Kit (Qiagen), and cDNA was synthetized as described previously [16]. After the RT-PCR, cDNA samples were diluted 1:2 using Rnase-free H20.
SLIT2 and ROBO1 were relatively quantified by intron spanning assays with Light-Cycler96 (Roche Diagnostics, Risch-Rotkreuz, Switzerland) and cytochrome c-1 (CYC1) as a reference gene. CYC1 was chosen as a reference gene because it is one of the most stably expressed genes in the placenta [16,83–85]. Primers and probes were: forward 5′-CTTCCAGAGACCATCACAGAAA-3′ and reverse 5′-CGTCTAAGCTTTTTATATGGTGAGAA-3′ for SLIT2 (with UPL probe #79), forward 5′-CGCAGAGAAACCTACACAGATG-3′ and reverse 5′-GGATTGGGCAGTAGGTGACT-3′ for ROBO1 (with UPL probe #31), and forward 5′-ATAAAGCGGCACAAGTGGTCA-3′ and reverse 5′-GATGGCTCTTGGGCTTGAGG-3′ for CYC1 (with UPL probe #47). Probes were from the Universal Probe library (UPL) Set (Roche Diagnostics). Each qPCR measurement was done in triplicate. Levels of SLIT2 and ROBO1 were normalized against the CYC1 level, and relative quantifications were then assessed with the ΔΔ cycle threshold method. A few randomly chosen qPCR products were also verified by agarose gel electrophoresis and Sanger sequencing.
Primers and probes genes for transforming growth factor alpha (TGFA), C-X-C motif chemokine ligand 6 (CXCL6), and SLIT-ROBO Rho GTPase activating protein 3 (SRGAP3) were: forward 5′-CCCTGGCTGTCCTTATCATC-3′ and reverse 5′-GGCACCACTCACAGTGTTTTC-3′ for TGFA (with UPL probe #74), forward 5′-CCAGAAAATTTTGGACAGTGG-3′ and reverse 5′-GGGATCTCCAGAAAACTGCTC-3′ for CXCL6 (with UPL probe #61), and forward 5′-GAAGGGCACTCGATGAGGT-3′ and reverse 5′-GCTCATGGTCTTCTCGATGTC-3′ for SRGAP3 (with UPL probe #66).
Total mRNA expression of different PSGs was measured with PCR primers: forward 5′-CCTCTCAGCCCCTCCCTG-3′ and reverse 5′-GGCAAATTGTGGACAAGTAGAAGA-3′ (with UPL probe #15), which are complementary to sequences conserved in all but two PSG transcript variants that both lack the N domain [26]. IL6 and TNFA were analyzed with primers forward 5′-GCCCAGCTATGAACTCCTTCT-3′ and reverse 5′-GCGGCTACATCTTTGGAATC-3′ for IL-6 (with UPL probe #43) and forward 5′-CAGCCTCTTCTCCTTCCGAT-3′ and reverse 5′-GCCAGAGGGCTGATTAGAGA-3′ for TNFA (with UPL probe #40).
Human placental trophoblast cells HTR-8/SVneo (CRL-3271™; ATCC, Manassas, Virginia, USA). were grown in RPMI-1640 culture media (Thermo Fisher Scientific) supplemented with 10% fetal bovine serum (FBS; Sigma-Aldrich, St. Louis, MO, USA) and 1× penicillin/streptomycin (Sigma-Aldrich). Cells were cultured under standard culturing conditions (37°C, 5% CO2, humidified atmosphere), and subculturing was performed with 0.05% trypsin/0.02% EDTA.
siRNAs targeting SLIT2 (s GUCAUAUCAAGAACUGUGAdTdT, a UCACAGUUCUUGAUAUGACdTdT) and ROBO1 (s CAUACCUAUGGCUACAUUUdTdT, a AAAUGUAGCCAUAGGUAUGdTdT) (Sigma-Aldrich) were reverse transfected and then forward transfected in HTR-8/SVneo cells with Lipofectamine 3000 reagent (Invitrogen, Carlsbad, CA, USA) [86]. MISSION siRNA Universal Negative Control #1 (Sigma-Aldrich) was used as a negative control for siRNA transfection and was transfected in the same manner as siRNAs targeting SLIT2 and ROBO1. In the reverse transfection, the cells (70,000 cells/well) were incubated with siRNAs at a final concentration of 30 nM. The cells were transfected again after 24 h of incubation. The second transfection was done as a forward transfection in the presence of 40 nM siRNAs. Cells were incubated with siRNAs for 24 h after the second transfection, and then fresh medium was added and cells were incubated for an additional 24 h. Cells were harvested with 1× Trypsin-EDTA (Sigma-Aldrich).
Cells were disrupted with a 25 G needle and 1 ml syringe, and RNA was isolated in accordance with the manufacturer’s instructions (RNeasy Micro Kit, Qiagen). The quality of isolated RNA was checked with an Agilent 2100 Bioanalyzer system at the Biocenter Oulu Sequencing Center, Finland.
Samples containing total RNA were sent for transcriptomic analysis to the Finnish Functional Genomics Centre (FFGC; Turku, Finland), where transcriptional profiles of SLIT2 (n = 3) and ROBO1 (n = 3) -silenced cells and negative-control cells (n = 3) were detected with the Illumina HiSeq high‐throughput sequencing system. Whole cell RNA sequencing data was analyzed by the Bioinformatics Unit Core Service at the Turku Centre for Biotechnology, Finland. The transcriptomics data have been deposited in NCBI’s Gene Expression Omnibus [87] and are accessible through GEO Series accession number GSE119101 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE119101)
Knockdown of SLIT2 and ROBO1 with siRNAs, total RNA isolation, cDNA synthesis, and qPCR were done as described above, except 30 nM siRNAs were used in the forward transfection instead of 40 nM.
Differences in mRNA expression levels among the phenotypes were assessed by nonparametric Mann–Whitney U-test with SPSS Statistics 20.0 (IBM Corporation).
|
10.1371/journal.ppat.1006329 | The host ubiquitin-dependent segregase VCP/p97 is required for the onset of human cytomegalovirus replication | The human cytomegalovirus major immediate early proteins IE1 and IE2 are critical drivers of virus replication and are considered pivotal in determining the balance between productive and latent infection. IE1 and IE2 are derived from the same primary transcript by alternative splicing and regulation of their expression likely involves a complex interplay between cellular and viral factors. Here we show that knockdown of the host ubiquitin-dependent segregase VCP/p97, results in loss of IE2 expression, subsequent suppression of early and late gene expression and, ultimately, failure in virus replication. RNAseq analysis showed increased levels of IE1 splicing, with a corresponding decrease in IE2 splicing following VCP knockdown. Global analysis of viral transcription showed the expression of a subset of viral genes is not reduced despite the loss of IE2 expression, including UL112/113. Furthermore, Immunofluorescence studies demonstrated that VCP strongly colocalised with the viral replication compartments in the nucleus. Finally, we show that NMS-873, a small molecule inhibitor of VCP, is a potent HCMV antiviral with potential as a novel host targeting therapeutic for HCMV infection.
| Viruses are obligate intracellular pathogens, meaning that they are completely dependent on the host cellular machinery to replicate. Identifying which host genes are necessary for virus replication extends our understanding of how viruses replicate, how cells function and provides potential targets for novel antivirals. Here, we show that a cellular factor called valosin containing protein (VCP) is essential for human cytomegalovirus replication. We demonstrate that VCP is required for the expression of an essential virus gene called IE2. Finally we show that a chemical inhibitor of VCP is a potent antiviral against human cytomegalovirus, demonstrating the potential for VCP inhibitors as novel therapeutics against this virus.
| Human cytomegalovirus (HCMV) is a highly prevalent herpesvirus, infecting 30 to 100% of the global population depending on the socio-economic status. Although normally asymptomatic in healthy individuals, HCMV infection is a significant cause of morbidity and mortality in immunocompromised populations, individuals with heart disease and recipients of solid organ and bone marrow transplant. HCMV is also the leading cause of infectious congenital birth defects [1–9].
During infection, HCMV initiates a programmed cascade of gene expression, resulting in production of infectious virus. Two of the first genes to be expressed are the major immediate early (MIE) genes IE1 (IE72) and IE2 (IE86). The MIE proteins have multiple roles during infection including transactivation of viral genes, which drives replication and virus production [10–12]. Because of this, they are thought to play a pivotal role in controlling the switch between latent and productive infection in HCMV [13,14]. While IE1 is required for efficient virus replication at low multiplicity of infection [13,14], IE2 expression is essential, with deletion resulting in non-viable virus [15]. IE1 and IE2 are generated from the same primary transcript by differential splicing and alternative polyadenylation [10,12,16]. They share the first three exons, with splicing to the fourth or fifth exon determining expression of IE1 or IE2 transcript, respectively (Fig 1). Independent polyadenylation signals exist downstream of both exon four and exon five. Such genomic arrangements, that require terminal exon skipping, are considered relatively unusual in the host cell, with specific factors and mechanisms involved in regulating the process not fully understood [17].
Valosin containing protein (VCP) belongs to the hexameric AAA ATPase family and plays a pivotal role in ubiquitin mediated signaling through remodeling target proteins, often leading to proteosomal degradation [18]. VCP contains two ATPase domains, which hydrolyze ATP to generate the energy required to remodel or unfold target proteins. Through this action, VCP is able to segregate target proteins from associated cellular membranes or larger protein complexes. Once segregated, the target protein is relocalised or degraded via the proteosomal complex. VCP can also affect which proteins are modified through its interaction with multiple ubiquitin regulatory co-factors, making VCP a central signalling hub for ubiquitin mediated regulation. In addition to ubiquitin, VCP has been linked to other post-translational modifiers such as SUMO and Nedd8 [19,20]. As such it is linked to a wide range of biological functions, including protein quality control, autophagy, chromatin remodeling, DNA damage response and more recently RNA processing [21]. Functionally, VCP plays a central role in protein homeostasis by facilitating proteosomal degradation of misfolded or damaged proteins as part of the endoplasmic reticulum associated degradation pathway (ERAD). In addition to its role in protein degradation, VCP has been linked to non-degradative functions involving removal and relocation of proteins from membranes and protein complexes. Examples include removal of Aurora B protein from mitotic chromosomes, removal of transcription factors from chromatin and disassembly of RNP complexes [22,23]. It has also been linked to a number of functional roles in virus replication including entry of Sindbis virus and replication of poliovirus [24–26]. While VCP plays a role in US11-specific degradation of MHC class I protein during HCMV infection [27] a direct role for VCP in the replication of HCMV has not previously been reported.
Using a focused siRNA screen, we identified VCP as an essential factor for HCMV replication. We show that VCP is essential for the onset of virus replication and knockdown of VCP results in changes in alternative splicing of the MIE transcripts, loss of IE2 expression and ultimately failure in virus replication.
Previously, we identified the host gene ATP6V0C as a critical factor in HCMV virus production [28]. ATP6V0C is a component of the vacuolar ATPase, and among other functions, is involved in membrane organization [29]. To identify additional host membrane organization factors involved in HCMV replication we performed a focused siRNA screen against 160 host genes involved in membrane organization using pools of four siRNAs against each target. Primary human fibroblast cells were double transfected with each siRNA pool in 96 well format. Two days post-transfection, the cells were infected with a low passage HCMV strain, TB40E, expressing GFP from an SV40 promoter. GFP levels were monitored over a period of seven days by fluorometry to determine virus infection and replication levels (S1 Table). Analysis of the screen indicated both increased and decreased virus replication ranging from 88% reduction in GFP levels to a 84% increase (Fig 2A). Z-score analysis based on three biological repeats of the siRNA screen identified five clear outliers that resulted in reproducible reductions in virus replication, as measured by GFP fluorescence intensity (Fig 2B). Of these, knockdown of VCP resulted in the largest negative Z-score. To validate that the observed phenotype was due to knockdown of VCP, rather than potential off-target effects, the original siRNA pool was deconvoluted to test the four individual siRNAs against VCP. All four siRNAs resulted in reduced VCP expression and reduced HCMV replication based on GFP fluorescence (Fig 2C). The third siRNA of the pool generated less efficient VCP knockdown, and corresponded with less inhibition of HCMV in the siRNA assay, providing additional evidence that knockdown of VCP results in a direct reduction in virus replication. These results indicate that the initial observed phenotype is due to specific knockdown of VCP rather than off-target artefactual effects of the siRNAs. Although GFP expression serves as an effective read-out for virus replication, loss of GFP signal may not necessarily reflect reduced virus production. To determine the direct effect of VCP knockdown on HCMV virus production, one-step growth curves were generated. Primary human fibroblast cells were transfected with siRNA pools targeting VCP or a negative control non-targeting siRNA. Two days post transfection cells were infected at a multiplicity of infection (MOI) of three with TB40E-GFP. Supernatant was collected at 24-hour time points for seven days total. Knockdown of VCP resulted in significant reduction in HCMV replication, with viral titers over four log10 lower in cells depleted of VCP compared to negative control cells at day seven (Fig 2D). No amplification of infectious virus was observed in cells knocked down for VCP, suggesting a complete block in virus replication. Cell viability assays showed a moderate decrease in cell viability at 5 days post transfection with VCP siRNA compared to negative control siRNA, although visual inspection of cells did not indicate gross cytotoxicity and the reduction is not sufficient to account for the substantial reduction in virus replication (S1 Fig).
Further characterization was performed to determine at what stage of the HCMV replication cycle VCP is required. Analysis of GFP fluorescence at 48 hours post infection (HPI) indicated that all cells were infected, demonstrating that VCP expression is not required for virus entry, translocation of genome to the nucleus or initial transcription of the viral genome (Fig 3A). Western blot analysis was performed to measure the levels of each of the major kinetic classes of HCMV genes, including immediate early (IE1 and IE2), early (pp52) and late (pp28), to define when disruption in virus replication occurs (Fig 3B). In control cells, protein expression of each class of virus gene is clearly observed. However, in cells knocked down for VCP, neither IE2 nor downstream viral gene expression was detected. Strikingly, despite being encoded by the same primary transcript, IE1 expression levels were higher relative to control cells, up until day four post infection. This data indicates that VCP expression is required for the onset of HCMV replication and suggests knockdown of VCP affects expression of the major immediate early genes.
Because knockdown of VCP resulted in specific loss of IE2 but not IE1, we hypothesized that VCP may play a role in IE2 protein stability, transcript stability or regulation of alternative splicing and polyadenylation. To determine whether VCP was required for IE2 protein stability we inhibited the two main pathways of protein degradation using MG132 (proteasome inhibitor–ubiquitin dependent) and pepstatin A or E64 (protease inhibitors–ubiquitin independent) in the context of VCP knockdown (Fig 4A). If VCP regulation of IE2 was due to protein degradation, inhibition of these pathways should rescue IE2 protein levels. Neither MG132, pepstatin A, nor E64 rescued the loss of IE2 proteins levels following knockdown of VCP. Instead, higher concentrations of MG132 phenocopied VCP knockdown in control cells, indicating that proteosomal function is required for IE2 but not IE1 gene expression. As VCP targets ubiquitinated proteins for proteosomal degradation, this result suggests that an intermediate protein, targeted by VCP, may inhibit IE2 expression and is consistent with previous publications showing inhibition of the proteasome reducing IE2 expression and HCMV replication [30].
To determine whether regulation of the MIE genes was occurring at the RNA level, Northern blot analysis was performed. Fibroblast cells transfected with siRNA against VCP or control siRNA were infected 48h post transfection at high MOI with TB40E and total RNA harvested at 24, 48 and 72 HPI. As shown in Fig 4B, knockdown of VCP resulted in a substantial reduction of IE2 RNA levels at 48 and 72 HPI, including the smaller IE60 and IE40 species. This result was confirmed by qRT-PCR, using primers specific for IE1 and IE2 transcript. Supplemental Fig 2A demonstrates the clear difference in the kinetics of IE1 and IE2 transcript accumulation in control cells, with IE1 transcript levels rapidly increasing following infection, while IE2 levels initially increase, then plateau between 12 and 24 HPI, before accumulating to substantially higher levels from 24 HPI onwards. This accumulation fails to occur following knockdown of VCP, where levels of IE2 do not substantially increase after 24 HPI, in contrast to control cells (S2B and S2D Fig). MIE transcription is also affected at earlier time points with levels of both IE1 and IE2 reduced at 6 HPI and increased at 24 HPI (S2C and S2D Fig). An increase in IE1 and IE2 RNA levels at 24 HPI may be a result of increased MIE promoter activity due to the associated loss of IE2 protein expression at early time points, as IE2 is a negative regulator of the MIE promoter. IE1 transcript levels decrease rapidly 24 to 48 HPI, while protein levels remain high, suggesting that IE1 protein is stable and the increase in IE1 transcription at 24 hours may contribute to the prolonged increase in IE1 protein levels. These results suggest that VCP may be affecting other aspects of MIE transcription. However, Northern, Western and qRT-PCR data are all consistent with a substantial effect on IE2 expression from 24 HPI onwards.
To determine whether loss of VCP results in specific destabilization of IE2 transcript, RNA levels were measured following treatment of cells with actinomycin D 48 HPI. Actinomycin D inhibits RNA polymerase, blocking new transcription of IE1 and IE2, allowing for monitoring of transcript stability over time. Total RNA was harvested at the indicated time points and IE1 and IE2 transcript levels determined by Northern blot analysis. Calculation of the half-life of IE1 and IE2 transcripts indicated that stability of both IE1 and IE2 modestly increased by 1.8 and 1.5 fold respectively, suggesting loss of VCP does not result in IE2 transcript destabilization (Fig 5).
Defining changes in alternative splicing of IE1 and IE2 is challenging as multiple factors, including promoter activity, RNA stability and viral genome amplification can all contribute to an increase or decrease in total IE1 and IE2 mRNA levels. To determine whether knockdown of VCP altered differential splicing of the MIE region, RNAseq analysis was performed. This allows a direct comparison of read counts from the first three shared exons of the MIE transcripts to the IE1 and IE2 specific exons four and five, respectively. This controls for changes in promoter activity and effects of genome amplification. Therefore, based on our current understanding of RNA processing, changes in the absolute ratio of exon four or five to the shared exons must be due to changes in splicing or changes in RNA stability. Cells were transfected with VCP siRNA or a negative control siRNA and infected with HCMV at an MOI of three. Total RNA was harvested at 24, 48 and 72 HPI and subjected to strand-specific, paired-end, Illumina sequencing. The annotated exons of the major immediate early region were clearly apparent from the mapped reads in all six samples and reads corresponding to the splice junctions were consistent with previous data for the region, indicating splicing between shared exons one to three and splicing between exon three and four, resulting in IE1 transcript, and three and five resulting in IE2 transcript (S3 Fig). Consistent with the Northern blot and qRT-PCR data, MIE transcription increased at 24 HPI but decreased at 48 and 72 HPI following knockdown of VCP (S4 Fig). Furthermore, splicing of the MIE primary transcript is heavily biased towards IE1 at 24 hours, with IE2 splicing increasing as the infection progresses (S4A Fig). To determine whether knockdown of VCP alters the balance of splicing between IE1 and IE2, the proportion of total reads mapping to the five exons of the IE region were calculated (S2 Table). Fig 6A represents the absolute difference in exon frequencies from VCP knockdown samples compared to negative control samples. At 24 HPI, knockdown of VCP has no effect on the relative proportion of reads mapping to exon four and five. However, by 48 HPI there is a clear increase in the relative proportion of reads mapping to exon four in VCP knockdown samples, with a corresponding decrease in exon five read frequencies. In contrast knockdown of VCP has no effect on the relative proportion of reads mapping to the shared exons, indicating reduced splicing from exon three to five (IE2) and a corresponding increase in splicing from exon three to four (IE1). Knockdown of VCP also resulted in lower frequencies of read pairs spanning exon three to five splice junction (IE2) and a corresponding increase in read pairs spanning exon three to four splice junction (IE1) (Fig 6A and 6B, S4B Fig and S3 and S4 Tables). These results show that knockdown of VCP alters splicing of MIE transcripts resulting in a failure to switch from IE1 splicing to IE2 splicing. To determine whether knockdown of VCP resulted in more generalized effects on virus transcript splicing, we analysed read counts for other well-characterized HCMV spliced transcripts (S5 Fig). Based on exon counts, only UL37 showed a similar pattern to MIE transcripts in response to VCP knockdown where absolute exon frequencies for UL37 exon 1 increased with a corresponding decrease in exon 3 read frequencies. However, unlike the MIE transcripts, analysis of paired-end reads did not support substantial splicing between UL37 exon 1 to exons 2 and 3 (S3 Fig). Instead, exon 1 and exons 2 and 3 of UL37 are likely to be predominantly independent transcriptional units with differing kinetics, with exon 2 and 3 requiring IE2 expression for transactivation.
To determine the effect of VCP knockdown on global viral gene expression, read counts were mapped to the entire viral genome. As shown in Supplemental Fig 6 total read counts mapping to the viral genome were relatively similar at 24 HPI. However, while viral gene expression increased in control cells by approximately 6.5 fold between 24 and 72 HPI, expression in VCP knockdown cells only increased by 3-fold, indicating a general reduction in viral gene expression following VCP knockdown. This is unsurprising given the vital role IE2 protein plays in transactivation of viral gene expression.
To determine whether knockdown of VCP results in a similar effect on all viral genes or differential effects, total reads were mapped to individual viral open reading frames and viral gene expression compared between control cells and VCP knockdown cells (Fig 7 and S5 Table). At 24 HPI moderate reductions in viral gene expression can be observed, with a subset of transcripts expressed at relatively higher levels following VCP knockdown. In particular, expression levels of UL36 and UL37, UL112/113 and the MIE transcripts UL122 (IE2) and UL123 (IE1) were relatively higher at 24HPI in VCP knockdown cells compared to control cells. However, at 48 and 72 HPI there is more profound global suppression of virus gene expression in VCP knockdown cells. This is particularly apparent for the non-coding genes RNA 4.9 and 5.0 and other genes that are highly up regulated at later stages of infection, including UL85, UL86, UL100 and UL102. Despite the loss of IE2 expression, the expression of a subset of viral genes remained equivalent to control levels following knockdown of VCP (Fig 7 and S7 Fig). These include UL21A, UL36 and UL38, UL111A and UL112/113, UL123 (IE1) and UL144. This is particularly surprising for UL112/113 which has previously been identified as a target for IE2 transactivation [31,32].
Previous studies have shown that the cell cycle impacts on the regulation of MIE alternative splicing with IE2 splicing and virus replication blocked at the G2/M phase [33]. Furthermore, cyclin dependent kinases are required for efficient IE2 expression and cyclin A2 overexpression inhibits IE2 splicing and virus replication [34]. To determine whether knockdown of VCP results in alterations in cell cycle, which in turn regulates MIE splicing, cells were stained with propidium iodide following VCP knockdown. The results clearly show that knockdown of VCP has no gross effect on cell cycle control (Fig 8). Consistent with the majority of cells being in G0/G1 phase, cyclin A2 levels remained undetectable by western blot in control and VCP knockdown cells (S8 Fig). These results show that regulation of MIE splicing by VCP is not a consequence of alterations in cell cycle control and instead represents a novel regulatory pathway. To determine whether the effect of VCP on MIE expression could be the result of a simple delay or inhibition of progression in virus replication, we compared MIE protein and transcript levels following treatment of cells with Ganciclovir, a well characterized drug, which inhibits viral DNA synthesis. While both Ganciclovir treatment and VCP knockdown result in substantial reductions in IE2 gene expression, only knockdown of VCP results in a corresponding increase in IE1 gene expression and relative changes in IE1 and IE2 splicing (S9 Fig). Furthermore, in contrast to VCP knockdown, IE2 and pp52 expression could be detected four days post infection following Ganciclovir treatment. These data suggest alterations in MIE levels, caused by VCP knockdown are not an artifact of delayed virus replication.
Previous studies have demonstrated that VCP plays an important role in the cytoplasm of HCMV infected cells, mediating US11-dependent degradation of MHC class I, by stripping the protein from the ER membrane in a ubiquitin dependent manner [27]. To determine the cellular localisation of VCP during HCMV infection, primary fibroblast cells were infected at high MOI with HCMV, fixed at 24-hour time points and stained for VCP. As shown in Fig 9, VCP displayed dynamic temporal cellular localisation during HCMV infection. In uninfected cells, VCP staining resulted in diffuse signal throughout the cell. However following infection with HCMV, distinct puncta can be observed in the nucleus of infected cells. By 48 HPI VCP was consistently found within two large puncta within the nucleus, which increased in size by 72 hours. Such staining is characteristic of the viral replication compartments. Following infection with HCMV the viral genomes are deposited adjacent to ND10 domains before forming two distinct replication compartments within the nucleus, characterized by colocalisation of viral DNA, DNA replication proteins and IE2 protein [35]. Therefore, VCP localisation to the virus replication compartments in the nucleus coincides with increased IE2 expression.
Given that VCP is essential for HCMV replication, we investigated whether small molecule inhibitors against VCP are viable antiviral candidates. Because VCP is a potential anti-cancer target, a number of effective small molecule inhibitors of VCP have been developed. NMS-873 is a highly potent and selective inhibitor of VCP [36]. To determine whether treatment of infected cells with NMS-873 results in the same phenotype as VCP siRNA knockdown, cells were pretreated with NMS-873 or DMSO and infected at high MOI with HCMV. Total protein was harvested every 24 hours for a total of five days and subjected to western blot analysis. As shown in Fig 10A, treatment with 1μM of NMS-873 recapitulated the phenotypic effects of VCP siRNA, with loss of IE2 transcript and protein expression. In contrast to siRNA knockdown of VCP, treatment with NMS-873 did not result in increased IE1 expression. This may be due to additional off target effects of the drug compared to siRNA knockdown or a consequence of inhibiting the activity of VCP compared to directly knocking down the protein. Next, we compared the antiviral activity of NMS-873 to Ganciclovir, currently the most commonly used HCMV antiviral. Primary human fibroblast cells were pretreated with different concentrations of Ganciclovir, NMS-873 or DMSO control, with cells and supernatant harvested seven days post infection for plaque assay. As shown in Fig 10B, NMS-873 and Ganciclovir clearly inhibited the production of infectious virus. However, NMS-873 displayed a higher level of potency at ten-fold lower concentrations, with an IC50 of 0.13 μM compared to 1.3 μM for Ganciclovir. Co-treating cells with NMS-873 at the same time as infection also resulted in reduced IE2 levels and a similar reduction in infectious virus (Fig 10C and S10A Fig). In contrast, treating cells 24 HPI did not reduce IE2 expression or virus production to the same extent (Fig 10C and S10B Fig). While immediate early and early viral gene expression was not substantially affected by treatment of cells with NMS-873 at 24 hour post infection, expression of the late gene pp28 was substantially reduced, with infectious virus reduced 100 fold, suggesting VCP may also be involved in late processes of virus replication. This is consistent with the pleiotropic nature of VCP and its potential to affect HCMV replication in a multitude of ways. Treatment with NMS-873 also showed little toxicity at effective concentrations (Fig 10D), suggesting NMS-873 and small molecule inhibitors of VCP show significant potential as HCMV antiviral therapies.
VCP plays a pivotal role in ubiquitin-dependent signaling through remodeling target proteins, often leading to proteosomal degradation. As such it is linked to a wide range of biological functions, including protein quality control, autophagy, chromatin remodeling and more recently RNA processing [21]. Following systematic screening of host genes involved in membrane organization, we identified VCP as a critical host factor for HCMV replication. Further characterization demonstrated that knockdown of VCP resulted in an initial increase in MIE transcription, followed by a substantial reduction in the expression of the major immediate early transcript IE2. Strikingly, despite being expressed from the same primary transcript, IE1 levels increased following VCP knockdown. Our data also suggests that VCP plays a role during late stages of virus replication as inhibition of VCP 24 HPI did not substantially affect IE2 expression but still resulted in a two-log reduction in infectious virus. Given the functionally pleotropic nature of VCP, it is unsurprising that knockdown results in multifactorial effects on the virus and potentially reflects the diverse roles of ubiquitin modulation and signaling during virus infection.
Viral gene expression during HCMV replication occurs through a tightly regulated cascade mechanism with expression of the MIE genes IE1 and, in particular, IE2, required for transactivation of early viral genes and subsequent late viral gene expression. Due to the central role of the MIE proteins in driving replication of the virus they have been suggested to play a pivotal role in regulating the balance between productive and latent infection. IE1 and IE2 are generated from the same primary transcript by alternative splicing, sharing the first three exons, while splicing to independent final exons through the relatively unusual process of terminal exon skipping [10,12]. Consistent with previous publications [33,34], our data shows that IE1 expression levels increase more rapidly than IE2, due to preferential splicing to the IE1 terminal exon 4, resulting in lower levels of IE2 transcription, which plateaus between 12 and 24 HPI (S2, S3 and S4 Figs). As infection progresses, splicing to IE2 terminal exon 5 increases, with a corresponding decrease in IE1 splicing. Regulation of alternative splicing of the MIE transcripts is therefore crucial for determining whether replication of the virus progresses, as expression of IE2 is dependent on this switch in splicing. Although the timing and precise splicing pattern of the MIE region has been well documented, how the splicing is regulated and the downstream consequences for virus replication are poorly understood [10]. However, it is likely to depend on a complex interplay of cellular and viral factors and the architecture of the MIE transcriptional region [37].
Expression of the MIE transcripts and their splicing are intimately linked to multiple factors including cell cycle, cell type and differentiation status [13,38,39]. Suppression of MIE transcripts during HCMV infection of myeloid cultures is thought to be key for establishing latent infection. Conversely, activation and differentiation of latently infected cells results in up-regulation of MIE gene expression and subsequent reactivation. During infection of fibroblast cells, expression of MIE transcripts and replication are closely linked to cell cycle status, with MIE expression suppressed during S/G2 phase [40,41]. This suppression is due in part to cyclin A2, which alters splicing of the MIE transcripts in a manner similar to knockdown of VCP, resulting in loss of IE2 expression and failure in virus replication [33]. While expression of cyclin A2 results in a similar phenotype, our data shows that VCP knockdown does not cause gross changes in the cell cycle or increased cyclin A2 expression, indicating an alternative mechanism of action.
Although the majority of research on VCP has focused on its cytoplasmic activity, recent studies have identified equally important and diverse roles for VCP in the nucleus. The association of VCP with the viral replication compartments early in infection suggests that the nuclear functions of VCP may be playing a critical role in HCMV replication. Studies in yeast and mammalian cell lines have demonstrated that VCP modulates the association of factors with chromatin, leading to regulation of cell cycle, transcription, DNA replication and DNA damage response, all of which could potentially affect virus replication and MIE expression [21]. Previous reports have shown that HCMV induces the host DNA damage response, with response factors including ATM, γH2AX and E2F1 required for efficient HCMV replication and IE2 expression [42]. γH2AX also colocalises with the viral replication compartments in the nucleus. VCP acts downstream of ATM and is required for the removal of ubiquitinated factors associated with dsDNA break regions, allowing repair to proceed [21]. If knockdown of VCP disrupts the DNA damage response during HCMV infection, this could contribute to the observed loss of IE2 expression and inhibition of virus replication.
Reports are also emerging that VCP plays a role in transcript stability and regulation of splicing. HuR is a RNA binding protein that stabilizes transcripts by binding to canonical AU rich regions within 3’UTRs. Following ubiquitination of HuR in response to stress signals, VCP removes HuR from associated transcripts resulting in transcript destabilization [43]. In Drosophila, VCP has been linked to dendrite pruning in neuronal cells [44]. Expression of mutant forms of VCP results in incorrect splicing of a gene important to dendritic pruning, MICAL, possibly through mis-localisation of the RNA binding protein TDP-43. Given the known mechanism of action of VCP it is likely that any effects on MIE splicing are indirect, occurring through targeting of a ubiquitinated intermediate protein. One possible model is that a negative regulator of IE2 splicing exists at early stages during HCMV infection, which promotes IE1 expression at the expense of IE2 expression. Based on the transcriptional architecture of the major immediate region, such a factor could regulate MIE splicing by promoting polyadenylation of the IE1 transcript, thereby inhibiting read-through to the IE2 final exon. The factor would be ubiquitinated and removed by VCP between 12 and 24 HPI, allowing read-through of the first polyadenylation site and splicing of the IE2 transcript. This would be consistent with the initial plateau in IE2 transcription and localisation of VCP to the replication compartments at 12 to 24 HPI, where spliceosome and proteasome components have also been reported to colocalise [30,45]. The NMS-873 data also supports this model, with inhibition of IE2 expression only effective with either pre-treatment or addition of the drug at the same time as infection. Treatment at 24 HPI was ineffective at blocking IE2 expression, suggesting a window in which VCP activity is required for regulation of MIE splicing. Interestingly, blocking de novo protein translation by cycloheximide treatment, effectively rescues IE2 transcription following NMS-873 treatment, suggesting the inhibitory factor is induced by virus infection (S11 Fig).
Although not common, there are a number of examples of terminal exon skipping in the human genome, which would be analogous to the proposed model. Maturation of B cells to differentiated plasma cells results in a switch from predominantly membrane bound IgM to secreted IgM. Studies have shown that increased expression of the polyadenylation factor cstF-64 promotes polyadenylation upstream of the exons required for membrane bound IgM resulting in increased expression of the secreted form [46,47]. Tissue specific regulation of calcitonin peptide is also regulated through differential polyadenylation of an embedded terminal exon [48]. Here, the splicing factor SRSF3 (SRp20) binds to a polyA enhancer sequence and directs polyadenylation of the upstream polyA site through recruitment of polyA factors. Restricting SRSF3 activity results in removal of the upstream terminal exon by splicing and enhanced expression of the downstream terminal exon. In these examples splicing is directly linked to cellular differentiation, in the case of IgM, and tissue specificity, in the case of calcitonin. These factors are fundamental to the regulation of HCMV latency with cell type specificity determining the site of latency and differentiation of latently infected cells linked to reactivation [13,14]. Given the central role of IE2 in determining virus replication, linking regulation of MIE splicing to cellular differentiation and tissue specificity would be a potent mechanism of regulating the establishment, maintenance and reactivation of latency in HCMV. Experiments are currently underway to determine whether regulation of MIE splicing by VCP follows a similar paradigm to cellular examples and whether such regulation is involved in HCMV latency.
Finally, as VCP is clearly essential for virus replication, small molecule inhibitors of VCP are potentially attractive antiviral candidates. HCMV-related illness accounts for more than 60% of diseases associated with solid organ transplant patients. Prolonged treatment, especially in patients with severely suppressed immune systems, greatly increases the risk of antiviral resistance [49–51]. Very few antivirals have been developed for use against HCMV since the licensing of Ganciclovir, and of these, the same viral genes are targeted, reducing the likely usefulness of these drugs against resistant strains [52]. An alternative strategy for the development of novel antivirals involves targeting of host genes required by the virus for successful replication. VCP has been identified as a potential anti-cancer target and as such a number of small molecule inhibitors have been developed, the most potent of which is NMS-873 [36]. We show that NMS-873 is 10 times more potent than Ganciclovir at equivalent concentrations, with little sign of toxicity at active levels. Whether the drug is toxic in vivo remains to be determined. Deletion of VCP in transgenic mice is non-viable and naturally occurring mutations in humans are linked to severe developmental defects [21,53]. However, transient inhibition of VCP by small molecule inhibitors may be a viable treatment option, especially in patients where resistance to current antivirals poses a significant risk to health.
Normal Human Dermal Fibroblasts (NHDF; Gibco) were maintained in Dulbecco’s modified high glucose media (DMEM; Sigma) supplemented with 10% fetal bovine serum (FBS; Gibco) and 1% penicillin-streptomycin (Invitrogen). A low passage HCMV strain TB40E, which expresses GFP from an SV40 promoter was used for siRNA library screening, western blot analysis, and northern blot analysis. Laboratory adapted HCMV strain AD169 was used for immunofluorescence experiments.
The Human siGENOME siRNA Library that targets 140 membrane trafficking genes (4 siRNAs per gene; Dharmacon, Inc.) and 20 other selected genes were included in the primary screen. In brief, NHDFs were seeded in 96-well plate a day before siRNA transfection. Next day, cells reached 90–95% confluency and were transfected with siRNA twice (4 hours apart between first and second transfections) using Lipofectamine RNAiMAX (Invitrogen) according to the manufacturer’s protocol. Transfected NHDFs were incubated for 48 hours and then infected with GFP expressing TB40E virus at an MOI of three. GFP intensity was monitored every 24 hours with Synergy HT microplate reader (Biotek). The entire screen was performed in duplicate and repeated three times. VCP gene identified from the primary screen was subsequently silenced with four individual siRNAs to eliminate the possibility that the phenotype was associated with off target effects.
Cells were lysed with RIPA buffer (0.1% SDS, 1% Triton X-100, 1% deoxycholate, 5 mM EDTA, 150 mM NaCl, and 10 mM Tris at pH 7.2) containing protease inhibitor cocktail (Roche). Ten micrograms of the total lysate was separated in 10% SDS-polyacrylamide gels and transferred to PVDF membranes (Millipore). Primary antibodies used in this paper are mouse anti-CMV IE1/2 monoclonal antibody (MAB8131, Millipore), mouse anti-CMV pp52 monoclonal antibody (CH16, Santa Cruz Biotechnology), mouse anti-CMV pp28 monoclonal antibody (CH19, Santa Cruz Biotechnology), rabbit anti-VCP polyclonal antibody (H-120, Santa Cruz Biotechnology), and mouse anti-β-Actin monoclonal antibody (Abcam). Blots were probed with primary antibody (1:500–1:2000) diluted in 5% dehydrated milk in Tris Buffered Saline (TBS) and subsequently the HRP-conjugated secondary antibodies (Pierce) at 1:5000. Blots were washed in TBS three times, incubated with chemiluminescent substrate (SuperSignal West Pico; Thermo Scientific) according to the manufacturer’s protocol, and exposed in G:Box (Syngene) for visualization of bands.
Total RNA was isolated by using Trizol solution according to the manufacturer’s protocol. Northern blot analysis for IE1 and IE2 mRNAs was conducted using total RNA that was separated on a 1.2% agarose formaldehyde gel and then transferred using Whatman TurboBlotter Rapid Downward Transfer Systems. IE1 and IE2 probes were generated by PCR using cDNA from TB40E infected NHDFs (as template) and labelled with Amersham Rediprime II DNA labelling system (GE Healthcare) with the following primers (IE1: TCAAACAGATTAAGGTTCGAGTGG, and ATCCACATCTCCCGCTTATCCTCG; IE2: TCATGGTGCGCATCTTCTCCACC, and TTACTGAGACTTGTTCCTCAGGTCC). After hybridization and wash, the membranes were exposed to autoradiography film with an intensifying screen (Biomax Transcreen HE, Kodak) for visualization of bands.
Cells were transfected and infected as described above. Total RNA was harvested using Trizol according to manufacturers guidelines. Total RNA was submitted to Edinburgh Genomics for generation of TruSeq stranded libraries and subjected to HiSeq high output v4 125PE sequencing. The strand-specific RNA-seq reads were mapped to a combined version of the human (hg38) and human herpesvirus 5 (KF297339) genomes using the HISAT spliced read mapper (PMID:25751142). Only valid read pairs mapped together in the correct orientation were retained (i.e. carrying a SAM flag of one of 83, 99, 147 or 163) and a custom Perl script was used to count the number of these reads mapping to each exon as well as the number of read pairs spanning different combinations of exons.
NMS-873 (Xcess Biosciences) and Ganciclovir (Cambridge Bioscience) were dissolved in DMSO and added to the cell culture at a working concentration 24 hours before HCMV infection. Viability for NMS-873 treated cells was assessed using the CellTiter-Glo luminescent cell viability assay kit (Promega) and CellTiter-Blue (Promega) 120 hours after NMS-873 addition. MG132 (Cambridge Bioscience), Pepstatin A (Sigma), and E64 (Sigma) were dissolved in DMSO and added to the cell culture 24 hours after HCMV infection. 100 μg/ml cycloheximide (in DMSO, Sigma) was added to the cell culture 30 minutes before HCMV infection to block protein biosynthesis.
Laboratory adapted HCMV strain AD169 infected cells were fixed and permeabilized in Methanol:Acetone solution (1:1) for 7 minutes, and then blocked with 5% human serum in PBS for 30 minutes. Primary and secondary antibodies were diluted with 5% human serum in PBS. Cells were washed with PBS after primary and after secondary antibody incubations. Primary antibodies used in this paper are mouse anti-CMV IE2 monoclonal antibody (12E2, Santa Cruz Biotechnology), rabbit anti-VCP polyclonal antibody (H-120, Santa Cruz Biotechnology) at 1:500. Alexa-fluor-647 conjugated goat anti-mouse or Alexa-fluor-488 conjugated goat-anti-rabbit IgG secondary antibodies were diluted 1:1000. All images were acquired with Zeiss LSM 710 confocal microscope fitted with 63X/1.4 oil-immersion objective lens.
Total RNA was isolated by using Trizol solution according to the manufacturer’s protocol followed by DNase (Turbo DNA-free kit, Ambion) treatment, and then reverse transcribed with poly T primers using High Capacity cDNA Reverse Transcription Kit (Invitrogen). Real-Time PCR was carried out using by Taqman assays with pre-designed gene-specific primer/probe set (Applied Biosystems) on Rotor gene 3000 (Corbet Research). Custom primer/probe set are CGTCAAACAGATTAAGGTTCGAGTGG, CCACATCTCCCGCTTATCCTCG, and 56-FAM/CATGCTCTG/ZEN/CATAGTTAGCCCAATACACTTCATCTC- CTCG/3IABkFQ for IE1, and ATGGTGCGCATCTTCTCCACC, TTACTGAGACTTGTTCCTCAGGTCCTG, and 56-FAM/CAGGCTCAG/ZEN/GGTGTCCAGGTCTTCGGGAGG/3IABkFQ for IE2.
Cells were harvested with trypsin, washed with PBS, and followed by fix with 70% cold ethanol for 30 minutes at 4°C. Then the fixed cells were washed twice with PBS, treated with RNaseA, and labelled with 10 μl of propidium iodide (1 mg/ml stock solution, Sigma). Cell cycle analysis was first carried out using BD LSRFortessa X20, and then further analyzed using FlowJo software.
|
10.1371/journal.pgen.1006063 | Antimicrobial Functions of Lactoferrin Promote Genetic Conflicts in Ancient Primates and Modern Humans | Lactoferrin is a multifunctional mammalian immunity protein that limits microbial growth through sequestration of nutrient iron. Additionally, lactoferrin possesses cationic protein domains that directly bind and inhibit diverse microbes. The implications for these dual functions on lactoferrin evolution and genetic conflicts with microbes remain unclear. Here we show that lactoferrin has been subject to recurrent episodes of positive selection during primate divergence predominately at antimicrobial peptide surfaces consistent with long-term antagonism by bacteria. An abundant lactoferrin polymorphism in human populations and Neanderthals also exhibits signatures of positive selection across primates, linking ancient host-microbe conflicts to modern human genetic variation. Rapidly evolving sites in lactoferrin further correspond to molecular interfaces with opportunistic bacterial pathogens causing meningitis, pneumonia, and sepsis. Because microbes actively target lactoferrin to acquire iron, we propose that the emergence of antimicrobial activity provided a pivotal mechanism of adaptation sparking evolutionary conflicts via acquisition of new protein functions.
| Immunity genes can evolve rapidly in response to antagonism by microbial pathogens, but how the emergence of new protein functions impacts such evolutionary conflicts remains unclear. Here we have traced the evolutionary history of the lactoferrin gene in primates, which in addition to an ancient iron-binding function, acquired antimicrobial peptide activity in mammals. We show that, in contrast to the related gene transferrin, lactoferrin has rapidly evolved at protein domains that mediate iron-independent antimicrobial functions. We also pinpoint signatures of natural selection acting on lactoferrin in human populations, suggesting that lactoferrin genetic diversity has impacted the evolutionary success of both ancient primates and humans. Our work demonstrates how the emergence of new host immune protein functions can drastically alter evolutionary and molecular interactions with microbes.
| Genetic conflicts between microbes and their hosts are an important source of evolutionary innovation [1]. Selective forces imposed by these antagonistic interactions can give rise to dramatic bouts of adaptive gene evolution through positive selection. J.B.S. Haldane originally speculated on the importance of infectious disease as an “evolutionary agent” over 60 years ago [2], and the Red Queen hypothesis later posited that predators and their prey (or pathogens and their hosts) must constantly adapt in order to sustain comparative fitness [3,4]. More recent studies have demonstrated how evolutionary conflicts progress at the single gene or even single nucleotide level, as molecular interfaces between host and microbial proteins can strongly impact virulence and immunity [5–7]. Host-pathogen interactions thus provide fertile ground for studying rapid gene evolution and acquisition of novel molecular traits [8].
Lactoferrin presents a compelling model for investigating adaptation from an ancestral “housekeeping” function to a specialized immunity factor. Lactoferrin arose from a duplication of the transferrin gene in the ancestor of eutherian mammals roughly 160 million years ago [9]. A fundamental and shared feature of these proteins is the presence of two evolutionary and structurally homologous iron binding domains, the N and C lobes, each of which chelates a single iron ion with high affinity. Iron binding by these proteins can effectively starve microbes of this crucial metal, a protective effect termed nutritional immunity [10,11]. Microbes in turn actively scavenge iron from these and other host proteins in order to meet their nutrient requirements [12,13]. The importance of iron in human infectious disease is highlighted by genetic disorders of iron overload, such as hereditary hemochromatosis, which render affected individuals highly susceptible to bacterial and fungal infections [14,15]. In addition to its role in nutritional immunity, lactoferrin has acquired new immune functions independent of iron binding following its emergence in mammals. Lactoferrin is expressed in a variety of tissues and fluids including breast milk, colostrum, saliva, tears, mucous, as well as the secondary granules of neutrophils and possesses broad antimicrobial activity [16]. Portions of the lactoferrin N lobe are highly cationic, facilitating interaction with and disruption of microbial membranes. Two regions of the lactoferrin N lobe in particular, lactoferricin and lactoferrampin, can be liberated from the lactoferrin polypeptide by proteolytic cleavage and exhibit potent antimicrobial activity against bacteria, fungi, and viruses [17,18]. Lactoferrin, as well as lactoferricin alone, can directly bind the lipid A component of lipopolysaccharide (LPS) as well as lipoteichoic acid, contributing to interactions with surfaces of Gram-negative and Gram-positive bacteria [19,20]. Lactoferrin thus poses a unique challenge for microbes—while its ability to bind iron makes it an attractive target for “iron piracy,” lactoferrin surface receptors could render cells more susceptible to associated antimicrobial activity. Despite a growing appreciation for lactoferrin’s immune properties, the evolutionary implications of these unique functions remain unclear. In the present study we decipher recent signatures of natural selection acting on lactoferrin in primates as well as modern humans to understand the evolutionary consequences of a newly acquired antimicrobial activity from a distinct ancestral function.
To assess the evolutionary history of lactoferrin in primates, we assembled gene orthologs from publicly available databases and cloned lactoferrin complementary DNA (cDNA) prepared from primary cell lines. In total, we compared 15 lactoferrin orthologs from hominoids, Old World, and New World monkeys, representing roughly 40 million years of primate divergence (Fig 1A and S1 Fig). We then used maximum likelihood-based phylogenetic approaches (performed with the PAML and HyPhy software packages) to calculate nonsynonymous to synonymous substation rate ratios (dN/dS) across this gene phylogeny [21–23]. For our study we included the N-terminal 19 amino acid positions of the full-length lactoferrin protein, which are removed during processing of the mature polypeptide in humans. Our analysis indicated that lactoferrin has evolved under episodic positive selection in the primate lineage, consistent with a history of evolutionary conflict with microbes (Fig 1A and S1–S7 Tables). These findings are also in line with previous genome-wide scans for positive selection in primates which identified the lactoferrin gene (LTF) among other candidate loci [24]. We next determined signatures of selection across individual codons in lactoferrin. In total, 17 sites displayed strong evidence of positive selection (posterior probability >0.95 from Naïve Empirical Bayes and Bayes Empirical Bayes analyses in PAML), with 13 of the 17 sites found in the N lobe (Fig 1B and 1C and S1 Fig and S2, S4, S5 and S6 Tables). This observation was notably dissimilar from a parallel analysis of primate serum transferrin, where sites under positive selection were restricted to the C lobe (Fig 1B and 1C and S3 Table). These results are further consistent with our previous work indicating that rapid evolution in primate transferrin is likely due to antagonism by the bacterial iron acquisition receptor TbpA, which exclusively binds the transferrin C lobe [25–28]. Thus, while lactoferrin and transferrin both exhibit signatures of positive selection in primates, patterns of selection across the two proteins are highly discordant.
Evidence of episodic positive selection in primate lactoferrin led us to more closely investigate variation of this gene across human populations. Data from the 1000 Genomes Project revealed six nonsynonymous polymorphisms at greater than 1% allele frequency in humans (S8 Table). Of the 17 sites we identified as rapidly evolving across primate species, amino acid position 47 overlapped with a high frequency arginine (R) to lysine (K) substitution in the N lobe of lactoferrin in humans (Fig 2A and S8 and S9 Tables). This position is markedly polymorphic between populations; while individuals of African ancestry carry the K47 allele at about 1% frequency, this variant is found in non-African populations at roughly 30–65% allele frequency, with the highest frequencies observed among Europeans (Fig 2B and S9 Table). The presence of R47 in related great apes combined with its high frequency in African populations suggests that R47 is in fact the ancestral allele in humans. Data from the Neanderthal genome browser (http://neandertal.ensemblgenomes.org) further revealed lysine to be the consensus residue at position 47 in recently sequenced Neanderthals. The presence of the lactoferrin K47 allele in Neanderthal and non-African human populations and its near absence in Africans suggests one of several intriguing genetic models for the history of this variant, including long-term allelic diversity in hominins, convergent evolution, or introgression from Neanderthals into modern humans.
Given the shared variation at position 47 between primate species and among human populations, we sought to determine whether lactoferrin exhibits signatures of positive selection in modern humans. Calculation of pairwise FST between a subset of human populations identified an elevated signal of differentiation between European (CEU) and African (YRI) populations [29], consistent with observed differences in allele frequencies between these groups (S2 Fig). The FST at rs1126478 was 0.70 (empirical p-value < 0.001), 0.30, and 0.03 for CEU-YRI, CEU-CHB, and CEU-FIN, respectively. Single nucleotide variants neighboring rs1126478 also showed signs of elevated FST suggesting that a shared CEU haplotype was driving the signal of differentiation (S2 Fig).
We next applied measures of haplotype homozygosity to assess the possibility that the K47 haplotype has been subject to natural selection in humans. Linkage around R47 alleles breaks down rapidly within a few kilobases, while the K47 variant possesses an extended haplotype (homozygosity of 0.5 at 21,913 bases), consistent with the possibility of an adaptive sweep in this genomic region (Fig 2C). A selective sweep is also consistent with bifurcation plots around position 47, where the K47 haplotypes possess increased homogeneity relative to R47 haplotypes (Fig 2D). We observed a slight an elevation of the genome-wide corrected integrated haplotype score (iHS) for the K47 allele (-1.40136) and a depletion of observed heterozygosity (S2, S3 and S4 Figs). We also examined the patterns of cross population extended haplotype homozygosity (XP-EHH). Consistent with the FST and EHH results, the XP-EHH score was elevated at the K47 position when CEU individuals were compared against YRI (1.1; p-value: 0.129) or CHB (3.1; p-value: 0.003)(S5 Fig). While XP-EHH between CEU and YRI was moderate, surrounding SNPs less than 3 kilobases away had values as high as 2.89 (rs189460549; p-value: 0.01). Genome-wide, the K47 XP-EHH signal is moderate compared to other loci. Next we compared the joint distribution of the p-values from dN/dS analyses [24] with the empirical p-values from the CEU-CHB XP-EHH analyses (S6 Fig). The previous genome-wide rank for lactoferrin, from dN/dS analyses, was 226 before considering the joint distribution and 156 after. The top 20 genes with the greatest change in rank (dN/dS p-value < 0.01) include BLK, DSG1, FAS, SLC15A1, GLMN, SULT1C3, WIPF1, and LTF. This meta-analysis highlights candidate genes that have undergone species-level as well as population-level selection in primates and humans, respectively. By integrating molecular phylogenetic analyses and population genetics approaches, we pinpointed signatures of positive selection associated with an abundant human lactoferrin polymorphism.
Signatures of positive selection in the lactoferrin N lobe among diverse primates, including position 47 in humans, led us to more closely investigate evolutionary pressures that have influenced variation in this region. After gene duplication from ancestral transferrin, lactoferrin gained potent antimicrobial activities independent of iron binding through cationic domains capable of disrupting microbial membranes. Two portions of the lactoferrin N lobe in particular, termed lactoferricin (amino acids 20–67 in full-length protein; 1–48 in mature protein) and lactoferrampin (amino acids 288–304 in full-length protein; 269–285 in mature protein), have been implicated in these antimicrobial functions [18,30].
Phylogenetic analysis revealed that several sites corresponding to lactoferricin and lactoferrampin display signatures of positive selection (Fig 3A and 3B). Notably, positive selection in lactoferricin localized to sites harboring cationic (lysine, arginine) or polar uncharged residues (asparagine), which could mediate membrane disruption and regulate antimicrobial activity. Position 47, which exhibits signatures of selection in humans as well as other primates, also lies within the lactoferricin peptide region. In contrast, hydrophobic tryptophan residues proposed to mediate insertion into microbial membranes are completely conserved among primates, as are cysteine residues that participate in intramolecular disulfide bond formation (Fig 3A). We also observed rapid evolution of the position immediately C-terminal to the pepsin cleavage site in lactoferrampin (Fig 3A), suggesting that the precise cleavage site in this peptide may be variable among species. Notably, the proteases responsible for lactoferrin processing in mucosal secretions and neutrophils remain elusive; identification of such factors will assist in revealing the consequences of genetic variation proximal to cleavage sites. Expanding our phylogenetic analysis to other mammalian taxa, we found that lactoferrin also exhibits signatures of positive selection in rodents and carnivores (S7 Fig and S10 Table). While the specific positions that contribute most strongly to these signatures could not be resolved with high confidence, N-terminal regions corresponding to lactoferricin in primates are absent in several rodent and carnivore transcripts, suggesting that this activity may have been lost or modified in divergent mammals. These observations are further consistent with previous work which identified signatures of positive selection in lactoferrin antimicrobial peptide domains across diverse mammals [31]. Together these results demonstrate that lactoferrin-derived cationic peptides of the N lobe are rapidly evolving at sites critical for antimicrobial action.
While rapid evolution of the lactoferrin N lobe may reflect selection for improved targeting of microbial surfaces, it could also represent adaptations that prevent binding by inhibitors encoded by bacteria. For example, pneumococcal surface protein A (PspA) is a crucial virulence determinant of Streptococcus pneumoniae, and several studies have demonstrated that PspA specifically binds and inhibits antimicrobial portions of the lactoferrin N lobe [32]. Consistent with an important evolutionary impact for this interaction, numerous sites under positive selection in the lactoferrin N lobe lie proximal to the PspA binding interface [33], including those corresponding to the lactoferricin peptide (Fig 3C). These data suggest that adaptive substitutions in lactoferrin could negate PspA binding, leading to enhanced immunity against S. pneumoniae or related pathogens.
Many strains of pathogenic Neisseria, which cause the sexually transmitted disease gonorrhea as well as acute meningitis, encode lactoferrin binding proteins (LbpA and LbpB) which mediate iron acquisition from lactoferrin [34,35]. Of four sites identified under positive selection in the lactoferrin C lobe, at least two appear proximal to the proposed Neisseria LbpA binding interface based on recent molecular modeling studies (S8 Fig) [36]. One of these, position 589, also aligns to a region under strong positive selection in transferrin (position 576 in humans) which directly contacts the related bacterial receptor TbpA (Fig 1B) [28]. These findings suggest that, similarly to transferrin, antagonism by bacterial Lbp proteins may have promoted natural selection in the lactoferrin C lobe. Signatures of selection at distinct lactoferrin-pathogen interfaces thus highlight the diverse conflicts that have arisen during the evolution of this unique immunity factor.
Together our results suggest that the emergence of novel antimicrobial activity in the N lobe of lactoferrin strongly influenced host-microbe interactions in primates, including modern humans (Fig 4). High disparity in sites under positive selection between the N and C lobes of lactoferrin and transferrin indicate that distinct selective pressures influenced these proteins during primate evolution. We previously demonstrated that primate transferrin has been engaged in recurrent evolutionary conflicts with the bacterial receptor, TbpA [25]. This receptor is an important virulence factor in several Gram-negative opportunistic pathogens including Neisseria gonorrhoeae, Neisseria meningitidis, Haemophilus influenzae, as well as related animal pathogens [26,37–39]. Notably, TbpA binds and extracts iron exclusively from the C lobe of transferrin, and signatures of positive selection in transferrin are almost entirely restricted to the TbpA binding interface (Fig 1) [25]. The fact that transferrin family proteins are recurrently targeted by microbes for iron acquisition may have provided the selective advantage for antimicrobial functions that arose in the lactoferrin N lobe.
Our results suggest at least two non-mutually exclusive scenarios for evolutionary conflicts involving the lactoferrin N lobe. Positive selection in this region could reflect adaption of lactoferrin for enhanced targeting of variable pathogen surfaces. Lactoferricin is capable of binding the bacterial LPS, which itself is heavily modified in many human-associated bacteria to mediate immune evasion and could provoke counter-adaptations at this interface. Conversely, variation in the lactoferrin N lobe could negate interactions with bacterial inhibitory proteins such as PspA encoded by S. pneumoniae. Lactoferrin binding activity has also been identified in several other important bacterial pathogens including Treponema pallidum [40], Staphlococcus aureus [41], and Shigella flexneri [42], raising the possibility of multiple independent evolutionary conflicts playing out at the lactoferrin N lobe. Iron-loaded lactoferrin could further be viewed as a “Trojan horse,” where microbes that target it as a nutrient iron source may be more susceptible to antimicrobial peptides. Consistent with this hypothesis, recent work has suggested that Neisseria encoded LbpB recognizes the lactoferrin N lobe, in contrast to its homolog TbpB which selectively interacts with the iron-loaded C lobe of transferrin [35,43,44]. LbpB binding to the lactoferrin N lobe could thus provide a counter-adaptation with dual benefits by neutralizing lactoferrin antimicrobial activity through negatively charged protein surfaces while simultaneously promoting iron acquisition by its co-receptor, LbpA [43]. These observations point to adaptations involving de novo protein functions on both sides of this molecular interface.
It is important to note that many “pathogenic” bacteria that routinely encounter lactoferrin in the respiratory mucosa are generally commensals that rarely cause disease. For example, H. influenzae colonizes a huge proportion of the human population but typically only causes disease in young children who lack a robust immune response. In addition, the dual functions of lactoferrin likely have pleiotropic effects on complex microbial communities in the host mucosa, with inhibition of some members creating new niches for others. Thus, the evolutionary forces acting on lactoferrin and the consequences for positive selection are likely more nuanced than a two-dimensional host-pathogen arms race. Future studies aimed at understanding the functional impact of lactoferrin variation will assist in understanding such complex biological effects.
Our results raise the possibility that the lactoferrin K47 variant introgressed into humans from Neanderthals at some point after the out-of-Africa expansion [45]. An alternative explanation could be convergent evolution of lactoferrin in distinct lineages of early hominins for enhanced immune function. Recent reports indicate that the human lactoferrin K47 variant, within the N lobe lactoferricin peptide, may have a protective effect against dental cavities associated with pathogenic bacteria [46]. Moreover, saliva isolated with patients homozygous for the K47 variant possesses enhanced antibacterial activity against oral Streptococci relative to homozygous R47 individuals [47]. Future analysis of lactoferrin sequence in archaic humans could provide additional insight on the history and functional properties of this variant. Together these studies provide a direct link between variation in the lactoferrin N lobe and protection against disease-causing bacteria, consistent with adaptive evolution of lactoferrin in humans and other primates.
Notably, the lactoferrin gene, LTF, is located only ~60 kilobases away from CCR5, a chemokine receptor which is also an entry receptor for HIV [48–52]. A 32-base pair deletion in CCR5 (CCR5-Δ32) confers resistance to HIV infection, and is present at a high frequency in northern Europeans while absent from African populations [53]. Although early evidence suggested that CCR5-Δ32 might itself be subject to positive selection in humans, more recent studies have concluded that these signatures are more consistent with neutral evolution [54]. It is intriguing that, like CCR5-Δ32, the lactoferrin K47 variant exhibits increased allele frequency in European populations relative to Africans. However, the presence of the K47 variant at high frequencies in Asian and American populations points to a much earlier origin for this variant than CCR5-Δ32. Moreover, EHH and bifurcation analyses indicate that the haplotypes associated with the lactoferrin K47 variant do not encompass CCR5, suggesting that variation at the CCR5 locus is unlikely to contribute to signatures of selection in LTF (Fig 2B and 2C and S9 Table). The proximity of the LTF and CCR5 genes combined with their high degree of polymorphism and shared roles in immunity suggest the potential for genetic interactions relating to host defense. Future studies could reveal functional or epidemiological links between these two factors in human immunity.
In summary, we have discovered that lactoferrin constitutes a crucial node of host-microbe evolutionary conflict based on signatures of natural selection across primates, including humans. Our findings suggest an intriguing mechanism for molecular arms race dynamics where adaptations and counter-adaptations rapidly emerge at the level of new protein functions in addition to recurrent amino acid substitutions at a single protein interface (Fig 4). Our evolutionary analyses highlight how the process of gene duplication and subfunctionalization can drastically alter the progression of host-microbe genetic conflicts.
RNA was obtained from the following species via the Coriell Cell Repositories where sample codes are indicated: Homo sapiens (human; primary human foreskin fibroblasts; gift from A. Geballe), Gorilla gorilla (western lowland gorilla; AG05251), Papio anubis (olive baboon; PR00036), Lophocebus albigena (grey-cheeked mangabey; PR01215), Cercopithecus aethiops (African green monkey; PR01193), Colobus guereza (colobus monkey; PR00240), Callithrix geoffroyi (white-fronted marmoset; PR00789), Lagothrix lagotricha (common woolly monkey; AG05356), Saimiri sciureus (common squirrel monkey; AG05311). Gene sequences from additional primate, rodent, and carnivore species were obtained from Genbank.
RNA (50 ng) from each primate cell line was prepared (RNeasy kit; Qiagen) and used as template for RT–PCR (SuperScript III; Invitrogen). Primers used to amplify lactoferrin cDNA were as follows: GTGGCAGAGCCTTCGTTTGCC (LF-forward; oMFB256) and GACAGCAGGGAATTGTGAGCAGATG (LF-rev; oMFB313). PCR products were TA-cloned into pCR2.1 (Invitrogen) and directly sequenced from at least three individual clones. Gene sequences have been deposited in Genbank (KT006751 –KT006756).
DNA multiple sequence alignments were performed using MUSCLE and indels were manually trimmed based on amino-acid comparisons. A generally accepted primate species phylogeny [55] (Fig 1A) was used for evolutionary analysis. A gene tree generated from the alignment of lactoferrin corresponded to this species phylogeny (PhyML; http://atgc.lirmm.fr/phyml/). Maximum-likelihood analysis of the lactoferrin and transferrin data sets was performed with codeml of the PAML software package [21]. A free-ratio model allowing dN/dS (omega) variation along branches of the phylogeny was employed to calculate dN/dS values between lineages. Two-ratio tests were performed using likelihood models to compare all branches fixed at dN/dS = 1 or an average dN/dS value from the whole tree applied to each branch to varying dN/dS values according to branch.
Positive selection in lactoferrin was assessed by fitting the multiple alignment to either F3X4 or F61 codon frequency models. Likelihood ratio tests (LRTs) were performed by comparing pairs of site-specific models (NS sites): M1 (neutral) with M2 (selection), M7 (neutral, beta distribution of dN/dS<1) with M8 (selection, beta distribution, dN/dS>1 allowed). Additional LRTs from the HyPhy software package that also account for synonymous rate variation and recombination (FUBAR, REL, FEL, MEME, BUSTED) were performed [22,23]. Molecular structures of lactoferrin, transferrin and associated proteins were visualized using Chimera (http://www.cgl.ucsf.edu/chimera/).
For variant-based analyses we used genotype calls from the 1000 Genomes project (release: 20130502, shapeit2 phased). Weir and Cockerham’s Fst estimator [29] was used for the population comparisons, implemented in GPAT++. EHH and the bifurcation diagrams were calculated using the [R] package REHH [56]. Genome-wide iHS scans were performed using GPAT++ and XPEHH plots were generated previously published datasets [57,58].
|
10.1371/journal.pgen.1002752 | Adaptive Introgression across Species Boundaries in Heliconius Butterflies | It is widely documented that hybridisation occurs between many closely related species, but the importance of introgression in adaptive evolution remains unclear, especially in animals. Here, we have examined the role of introgressive hybridisation in transferring adaptations between mimetic Heliconius butterflies, taking advantage of the recent identification of a gene regulating red wing patterns in this genus. By sequencing regions both linked and unlinked to the red colour locus, we found a region that displays an almost perfect genotype by phenotype association across four species, H. melpomene, H. cydno, H. timareta, and H. heurippa. This particular segment is located 70 kb downstream of the red colour specification gene optix, and coalescent analysis indicates repeated introgression of adaptive alleles from H. melpomene into the H. cydno species clade. Our analytical methods complement recent genome scale data for the same region and suggest adaptive introgression has a crucial role in generating adaptive wing colour diversity in this group of butterflies.
| Hybridisation occurs between many animal species, however its evolutionary relevance is still a matter of great debate. While some argue that hybridisation leads to maladaptive gene combinations, and therefore to an evolutionary dead end, others consider interspecific hybridisation as a process with great potential to fuel evolution. We examine this question by exploring the origins of red wing colouration, a trait under natural selection, in the adaptive radiation of closely related species of Heliconius butterflies. By sequencing genetic regions both linked and unlinked to the red wing pattern locus, we found experimental evidence supporting multiple hybridisation events that have mediated the acquisition of colour adaptations from H. melpomene to H. timareta. This introgression has allowed H. timareta to colonise new fitness peaks in the Müllerian mimicry landscape. In this way, our results support the idea that interspecific hybridisation in animals constitutes a source of genetic variation that promotes diversification.
| Closely related species often hybridise through incomplete barriers to gene flow, but the evolutionary consequences of such genetic interchange remain a matter of debate [1], [2], [3], [4], [5], [6]. This is primarily because hybridisation is considered unlikely to introduce useful genetic variation [1], [4], [5], [7]. Alleles that cross species boundaries may be neutral in their effects [7] or, perhaps most commonly, natural selection will prevent the introgression of foreign genetic material into a genetic background that is already well adapted [8]. However, sometimes, introgression may be favoured if the region gained confers advantages to the recipient species [5]. Although such favourable gene combinations may be produced only rarely, they might still contribute important variation for adaptive change. Importantly, hybridisation is a potential source of novel alleles already tested by natural selection that would be unlikely to arise through mutation alone.
In organisms other than bacteria, evidence for adaptive introgression in nature is scarce [9], [10]. Nonetheless, several remarkable examples in plants have demonstrated adaptive introgression, for example in transferring herbivore resistance in Helianthus [5], flood tolerance in Iris [11] and the gene controlling rayed flowers in Senecio vulgaris [12]. In animals, examples include adaptive introgression of melanism from domestic dogs into North American wolves [13] and warfarin pesticide resistance in European house mice, gained from the Algerian mouse [14]. Nonetheless, these examples all represent a single instance of transfer of a trait, often in association with environments showing significant levels of human intervention. A more pervasive role for introgression in recent adaptive radiations has been postulated, for example in Darwin's finches and sailfins [15], [16], but convincing genetic evidence for introgression of specific adaptive traits is still missing in these systems.
Heliconius butterflies display a striking radiation in adaptive wing patterns, facilitated by Müllerian mimicry between distantly related species and coupled with divergence between closely related species [17]. These butterflies frequently hybridise across species boundaries [18], [19], and it has been hypothesised that introgression might play an important role in speciation and adaptive radiation. In particular two closely related species groups, Heliconius melpomene and Heliconius cydno are known to hybridise occasionally, and genetic evidence indicates a low level of ongoing gene flow [20], [21]. H. melpomene has radiated into almost 30 geographical colour pattern races across Central and South America [22], broadly falling into two main phenotypes, which we here refer to as the red-banded type (presence of a red band or patch in the forewing controlled by the B allele, regardless of hind wing phenotype) and the rays type (orange forewing basal patch and orange rays in the hind wing). The sister clade to H. melpomene includes the species Heliconius cydno, H. pachinus, H. timareta and H. heurippa, jointly referred to hereafter as the H. cydno clade [23]. The former two species are typically black with white or yellow elements [22], while the latter two species exhibit patterns similar to those of H. melpomene [23], [24]. We have previously suggested that the presence of red phenotype elements in these H. cydno affiliates, that is H. heurippa and H. timareta, could be the result of the acquisition of mimicry colour patterns via adaptive introgression from H. melpomene [3], [19], [24], and in the case of H. heurippa have provided DNA sequence evidence in support of this transfer [25]. However, these phenotypic patterns could also be explained if red variants were either ancestral, with multiple subsequent trait losses in the H. cydno clade, or if they had independent origins in both H. melpomene and the red H. cydno affiliates, specifically H. timareta and H. heurippa [26].
In H. melpomene the HmB locus controls variation in red colour patterns [27], [28], a trait under strong natural selection [29], [30]. Genomic analysis of this region has identified clear peaks of genetic divergence between adjacent races of H. melpomene associated with variation in red phenotypes [25], [28], [31]. In H. melpomene, the strongest divergence lies in a non-coding region in between a kinesin gene and the transcription factor optix [31]. The latter is the strongest candidate gene so far for the red locus [32], and its expression shows a perfect association with red wing colour elements in a wide range of geographical races of H. melpomene and its co-mimics H. erato, prefiguring in both species the forewing red band, the dennis orange patch and the hind wing rays [32].
Having such information provides an excellent opportunity to explicitly test the introgression hypothesis for red wing patterns across the broader H. melpomene/H. cydno species complex. Here, we specifically examine the phylogenetic history of divergent and convergent colour pattern races of H. melpomene, H. cydno, H. timareta and H. heurippa and ask how this history varies between loci linked and unlinked to colour pattern. The data allows us to understand the origins of adaptive colouration and ask whether similar wing patterns have multiple independent origins, or arose once within the complex and crossed species boundaries. Thus, we provide an explicit test of the hypothesis that hybridisation has repeatedly contributed to an adaptive radiation. This study was carried out alongside a genome-wide study of a subset of the taxa included here [33]. The analyses presented here on smaller gene regions, sequenced across a much larger set of taxa, permit a different set of analytical tools to be used to test for the extent and direction of introgression.
We analysed 221 haplotypes from nine loci (Table S1), sampled from 111 individuals in five species (Figure 1). Three loci (the mitochondrial fragment COI and nuclear GAPDH and Hsp90) were unlinked to colour pattern, whereas the remaining six loci were sampled across the genomic interval modulating red pattern variation, specifically where the highest genetic divergence peaks associated with variation in red phenotypes have been found in H. melpomene [31]. Analysis of molecular variance in the mitochondrial fragment COI showed population structure largely explained by species relationships (∼47%) and geography (∼30%) but less by colour phenotype (Table 1). Phylogenetic analysis supports three monophyletic clades: (i) H. cydno-H. timareta, (ii) H. melpomene from the Pacific and the Atlantic coast, and (iii) H. melpomene from the Amazonas and the Andes (Figure 2).
In previous studies, nuclear markers showed varying degrees of clustering by species, with some loci showing mutual monophyly between the H. melpomene and H. cydno clade species, while others showed substantial allele sharing among species [20], [34]. Here, both unlinked nuclear markers (GAPDH and Hsp90) showed little population structure either by colour phenotype, species or geography (Table 1) with only about 15% of the variation explained by species and much less by colour pattern phenotype (Table 1). This result was corroborated by phylogenetic analysis (Figure 2), where similar alleles were spread broadly among species, wing pattern phenotypes, and across major biogeographic boundaries. Even among some loci within the red pattern interval, for instance kinesin and Hm01012, there was a poor correspondence with either species boundaries, geography or colour pattern, with each factor explaining less than 10% of the molecular variation at these markers (Table 1). Similarly, phylogenies of these two markers did not exhibit clear clustering by any of these categories (Figure 3).
Other markers across the red locus showed an increasing tendency to partition variation by colour pattern phenotype. Coding sequence of the transcription factor optix clustered most of the red-banded phenotypes of H. melpomene together, but H. melpomene individuals with rayed phenotypes were scattered across the genealogy. Optix also failed to show a clear phenotype association for H. timareta and H. heurippa (Figure 3). Nonetheless, colour pattern explained 28% of the variation within optix alleles (Table 1). A similar result was observed at HmB449k and HmB520k, which were 75 kb downstream and 2 kb upstream from optix, respectively, in the red interval (Table 1). Both loci grouped H. melpomene red-banded phenotypes into a monophyletic lineage (Figure 3), but failed to show a phenotype association in H. heurippa, H. timareta and rayed H. melpomene.
HmB453k was the striking exception to these patterns and showed strong population structure based on colour phenotype when analysed both by Neighbour-Joining and Maximum Likelihood (Figure 3, Figure 4). Over 60% of the segregating variation at this locus was explained by colour pattern phenotype (Table 1). Moreover, the allelic genealogy of this locus clearly defined three major clades, which largely corresponded to three major colour pattern phenotypes (Figure 4). The first clade contained red-banded type taxa (H. melpomene, H. timareta subsp. nov from Peru and H. heurippa), the second grouped rayed species (H. melpomene, H. timareta florencia and H. timareta contigua), and the third containing the species with no dorsal red wing colouration (H. cydno, H. timareta subsp. nov from Colombia and H. timareta timareta). Strikingly, individuals of the polymorphic population of H. timareta from eastern Ecuador were separated by phenotype, with rayed and non-rayed individuals sampled from the same locality falling into their respective phenotypic clades.
There were some exceptions to the complete clustering by phenotype in HmB453k (Figure 4). For example, the east Andean race, H. m. plesseni possesses white and red spots on the forewing and is typically considered a red-banded pattern. However, here all individuals from this race form a distinct monophyletic group on the HmB453k genealogy (Figure 4). This perhaps indicates that this phenotype shows an independent origin as compared to other red-banded patterns, consistent with its distinct white and red band phenotype. In addition, six haplotypes from the rayed race H. m. malleti did not cluster in the same clade as other rayed individuals, but similarly formed a separate monophyletic clade nested within the broader genealogy (Figure 4). This might also represent an independent origin of rayed phenotypes within H. melpomene, but is perhaps more likely a result of recombination between the HmB453k marker and nearby functional sites.
In order to address alternative explanations for the strong colour pattern signal within the HmB453k genealogy [26], we tested three alternative tree topologies for this fragment. The first alternate topology assumed that mtDNA topology correctly reflected the relationship among the three species; the second, a topology that considers independent phenotypic convergence in H. melpomene, H. timareta and H. heurippa and the third, a topology where H. melpomene constitutes a red polymorphic ancestral taxon and H. cydno/H.timareta/H.heurippa are derived with multiple losses of red patterns (Figure S1). According to the Shimodaira–Hasegawa (SH) test, the ML tree was better supported than any of the three alternative topologies (p<0.05 in all cases) [35]. These same three alternative tree topologies (Figure S1) were also tested against a ‘perfect’ ML HmB453k genealogy where the non-clustering alleles of H. m. malleti and H. m. plesseni were removed. In this case, again the SH test showed that the ML tree was better than any of the three alternatives (p<0.05 in all cases). Thus, we can rule out the alternative hypotheses proposed for pattern sharing across this group, namely multiple independent origins of red patterns, or ancestral red patterns subsequently lost multiple times [26].
To determine whether introgression is the cause of the shared DNA sequence variation observed among species, we applied the Isolation with Migration model in H. melpomene, H. timareta and H. heurippa using the program IM [36]. In order to obtain non-recombining blocks of sequence for this analysis, the taxa were separated into rayed and red-banded groups (see methods). In both datasets, IM estimated a population size of H. timareta smaller than that of H. melpomene (Table S3) and a time of divergence between these two species of ∼700,000 years. Maximum-likelihood estimates for introgression (2Nm), in general showed evidence of gene flow between species in the four markers analysed (Table 2). Models invoking gene flow in both directions were a significantly better fit than any model with no gene flow in any or in both directions (Table 3, models ABC0D, ABCD0, ABC00). We also found evidence for significant asymmetry in gene flow, as the model with unequal gene flow between species was significantly better than the model with similar gene flow in both directions (Table 3, model ABCDD).
When gene flow parameters were estimated for individual genes, nuclear genes Hsp90 and GAPDH, together with the mitochondrial fragment COI, showed evidence for ongoing gene flow between the study species (Table 2). However, the fragment HmB453k was the only marker consistently showing the strongest unidirectional introgression from H. melpomene to H. timareta in both phenotype datasets, thus suggesting that HmB453k alleles of H. timareta are derived from those of H. melpomene (Figure 5). Most notably in the rayed data set, this marker showed the highest magnitude of gene flow seen at any of the markers (Table 2). As the HmB453k fragment is located in the genomic region controlling the red wing phenotypes that is known to be under selection, one of the IM model assumptions is violated. Previous IM analysis on simulated scenarios with divergent selection in early stages of divergence have showed underestimated gene flow rates (2Nm) [37]. It could be argued therefore, that if selection is having an effect on our estimates we might be underestimating migration rates.
To further explore and confirm the signatures of introgression between these species we also used a linkage disequilibrium (LD) test for gene flow [38].
Briefly, the difference (x = DSS−DSX) between the magnitude of LD among all pairs of shared polymorphisms (DSS; Disequilibrium Shared-Shared) and that among all pairs of sites for which one member is a shared polymorphism and the other is an exclusive polymorphism (DSX; Disequilibrium Shared-Exclusive), is indicative of whether or not polymorphisms in the populations are the result of gene flow (positive x value) or retained ancestral polymorphism (negative x value) [38]. This because polymorphisms that are shared due to ancestral polymorphism are expected to be older on average, having more time to recombine and break down associations, than polymorphisms acquired via post-divergence gene flow [38].
We applied this test to the same phenotypic groups analysed with IM, and additionally, to pairs of H. melpomene and H. timareta populations found in sympatry. In general, the LD analysis showed values consistently positive across all comparisons and loci (Table 4), suggesting onging gene flow between H. melpomene and H. timareta. Notably, HmB453k was the only locus with significant gene flow in both the phenotypic and sympatric datasets, where H. timareta was always the recipient species (highest positive value) of H. melpomene alleles (p<0.001, Table 4). This analysis therefore provides strong confirmation of the IM results.
Adaptive novelty can arise de novo from mutations, from standing variation within populations or through gene flow among related populations or species, and the relative importance of these factors remains an open question in evolutionary biology. In Heliconius butterflies, the recent identification of the optix transcription factor as the locus of selection for red wing phenotypes offers the opportunity to address this question [32]. In a parallel study, we demonstrated that the distantly related H. melpomene and H. erato radiations use independently derived optix alleles to generate mimetic red patterns, implicating de novo mutations at the same locus [39]. Here, in contrast, we show that mimicry between more closely related species has involved multiple instances of allele sharing through adaptive introgression. Thus, the allelic variants that fuel adaptation do not necessarily need to be generated de novo, but can also be derived from introgression, accelerating the evolutionary process.
In this and previous studies, putatively neutral markers have shown that H. melpomene and the H. cydno clade are two distinct species assemblages that occasionally exchange genes [20], [21], [40]. Despite evidence for gene flow at neutral markers, H. melpomene and the H. cydno clade species often coexist in Central America and the Andes, and are ‘good’ species with distinct ecologies and strong barriers to gene flow, including both strong pre- and post-mating isolation [22], [41], [42], [43]. Here, we also found pervasive gene flow among H. melpomene, H. timareta and H. heurippa, similar to that previously observed in comparisons involving H. melpomene and H. cydno [21], [40]. The gene flow observed at markers unlinked to the wing pattern locus is bi-directional and not correlated with any obvious phenotypic trait. In contrast, the HmB453k marker, located within the red colour locus in a non-coding region downstream of optix, shows a striking association with wing phenotype and unidirectional introgression from H. melpomene to H. timareta.
The functional sites driving phenotypic variation within Heliconius are almost certainly cis-regulatory elements of optix and perhaps other adjacent protein coding regions, which act as a phenotypic switch for red pattern elements [32]. Notably, optix shows no amino acid substitutions between divergent colour pattern forms of the same species or between convergent forms of distantly related species [32]. To date, HmB453k shows the strongest association with wing pattern phenotype, much stronger than kinesin, which showed evidence for adaptive introgression of red phenotypes into H. heurippa [25], and even stronger than the optix coding region [32]. The strong signal we observe at HmB453k argues that it must be very close to the functional region(s) regulating colour pattern variation. Nonetheless, the fact that two races (H. m. plesseni and H. m. malleti) do not fall into the expected clades in this marker, might suggest that HmB453k does not itself contain functional sites. It is also likely that multiple functional sites across the genomic region control different aspects of the phenotype. Indeed, linkage disequilibrium analysis shows at least three sites in optix and HmB453k that consistently co-segregate (Figure S2), and in general there is substantial linkage disequilibrium across the HmB locus.
The lack of a strong association at the kinesin locus was surprising given the strong association seen at this locus in our previous study of H. heurippa, albeit with a much more taxonomically restricted sample [25]. However, we have considerably smaller sequence coverage of kinesin here, which might affect the signal we recovered from this gene. We believe that the previous study identified a genuine signal of introgression, but that the functional sites controlling the phenotype, which are likely to be regulatory in nature, are located in the non-coding sequence between kinesin and optix. Unpublished expression data indicate that there is evidence for functional involvement of both genes in wing pattern specification (Pardo-Diaz, unpublished data).
Previous criticism of the hypothesis of adaptive introgression in Heliconius and these species in particular, has focused on two alternative hypotheses, that either red variants were ancestral, with multiple subsequent trait losses, or that they have independent origins in these closely related lineages [26]. We have explicitly ruled out these alternatives, both by coalescent analysis using IM and LD analysis that indicate strong and significant evidence for directional gene flow, and by tree topology tests. In addition, it could also be hypothesised that natural selection might drive independent and convergent evolution of the sequence variants seen in the HmB453k region, if these were directly responsible for regulation of optix expression. Under this hypothesis however, one would expect multiple divergent haplotypes to be associated with this region in the surrounding sequence. Instead, we clearly observe a single haplotype at the centre of the associated region, with a decline in association with genetic distance, consistent with a single origin for each phenotype in this clade.
Alongside a parallel study involving a complete genomic sampling of the red colour region in a subset of the taxa used here [33], our data provide the first evidence for adaptive introgression driven by mimicry in Heliconius. The introgression previously documented in H. heurippa established a novel non-mimetic phenotype in eastern Colombia [24], [25]. In contrast, the additional cases of introgression documented here represent convergence due to mimicry selection, rather than establishment of an entirely novel pattern, albeit with a common genetic origin for the shared patterns. The direction of the asymmetrical gene flow is consistent with mimicry theory. First, H. melpomene is generally more locally abundant in the eastern Andes as compared to H. timareta (CJ, pers. obs.), so mimicry theory would predict that rare species should experience stronger selection to converge onto abundant models. Thus it seems likely that H. timareta adopted the local H. melpomene wing pattern, rather than vice-versa. Second, H. cydno and its co-mimics H. sapho and H. eleuchia are almost entirely restricted to the western side of the Andes [44]. One plausible scenario is therefore that the ancestors of H. timareta migrated along the eastern slope of the Andes and were faced with the absence of a white/yellow co-mimic. It seems likely that this imposed an additional selection pressure to mimic H. melpomene and H. erato instead. This eventually led to the establishment of H. timareta as a replacement of H. cydno distributed along the eastern slopes of the Andes in sympatry with H. melpomene.
The data provide evidence for multiple independent introgression events. H. t. florencia shares a rayed pattern with its co-mimic, H. m. malleti, in south-eastern Colombia [45], while the very different phenotype of the red-banded race H. t. ssp. nov. is mimetic with H. m. amaryllis in the Tarapoto region of Peru. A likely additional case is represented by the polymorphic population of H. timareta in Ecuador. Although the rayed phenotype in this population may share a common origin with that of H. t. florencia in Colombia, their distribution is disjunct, separated by the red banded H. tristero found in Mocoa, Colombia. Thus, the acquisition of red patterns by H. timareta has been driven by natural selection for mimicry, and has occurred multiple times (at least once for each red colour element) in the last 700,000 years.
The introgression of regions controlling red wing colouration from H. melpomene to the H. cydno clade has facilitated mimicry and has also played a role in speciation. In H. heurippa the red/yellow hybrid pattern is used as mating cue, which contributes to reproductive isolation from its closest relatives [24], [25], [46]. Although barriers to gene flow within H. timareta have not been investigated, it is possible that similar isolation might be found between red-banded and rayed races of this species, such that these might represent incipient species generated through hybridisation.
In this and previous work we are beginning to piece together a more complete picture of the history of this complex adaptive radiation. It seems likely that the red-banded pattern in H. erato spread and diversified early in the history of the radiation, followed by emergence of the H. erato rayed pattern that spread across Amazonia interrupting the geographical continuity of the ancestral red-banded phenotype [39]. In the H. melpomene lineage there was a speciation event in which H. cydno colonised the yellow/white phenotypic niche to mimic the H. eleuchia and H. sapho clade, and H. melpomene diversified to mimic the phylogenetically distant H. erato [39]. Reproductive isolation between the species is partly due to colour pattern mate choice, which arose between closely related taxa such as H. melpomene and H. cydno. Then divergence of the H. timareta/heurippa lineage from the rest of H. cydno, around 700,000 years ago, arose as a result of adaptive introgression of wing patterning alleles from H. melpomene in the eastern Andes.
In summary, we provide evidence that contributes to resolving the longstanding debate over the evolutionary importance of hybridisation in animals. Our data allow statistical tests for the incidence of introgression based on both coalescent patterns and linkage disequilibrium, with consistent results, and indicate the direction of introgression. The results imply that interspecific hybridisation facilitates adaptability and diversification, not only when the selection pressure is human-mediated, but also allows the colonisation of either empty or under-utilised fitness peaks in animal adaptive radiations. In other adaptive radiations such as Darwin's finches [15], Daphnia waterfleas [47] and African cichlids [48], rapid diversification may similarly be mediated by introgression [1]. The evolutionary impact of such transfers might be higher if the traits interchanged are also involved in reproductive isolation, thus contributing to speciation.
Our sample set consisted of 111 individuals from 4 different species, namely H. melpomene, H. cydno, H. timareta and H. heurippa (Table S1). In total 14 races of H. melpomene, 8 races of H. cydno, 5 races of H. timareta and one of H. heurippa were sampled covering most of the geographic distribution of each species from Central to South America (Figure 1). H. numata was included as outgroup. DNA was extracted using the QIAGEN DNeasy 96 Blood & Tissue Kit. One mitochondrial and eight nuclear fragments (Table S2) were amplified with QIAGEN Taq DNA Polymerase, purified using ExoSAP and sequenced with ABI Big Dye Terminator. Two of the nuclear markers are unlinked to the HmB red locus whereas the remaining six are all located across the region (Table S2). From these colour-linked fragments, optix and kinesin have previously been implicated in red wing pattern determination. The remaining four were identified as regions under divergent selection with high levels of population differentiation associated with red colouration [31].
Sequences were aligned and cleaned using Codon Code Aligner. Haplotype inference for heterozygous calls was conducted with the PHASE algorithm implemented in DNAsp v5.10.01 [49], with 5000 iterations and allowing for recombination. Inferred haplotypes with a confidence higher than 95% were accepted. In the case of the fragments HmB449k and HmB453k cloning was necessary due to the presence of considerable indel variation. PCR products of these two markers were ligated into the pGEM-T easy vector and five to ten clones per individual were selected and sequenced. Sequences were deposited in GenBank under accession numbers JX003980–JX005837.
For each fragment, phased haplotypes were used to construct phylogenetic trees using the Neighbour-Joining method under the P model of uncorrected distance in PAUP* 4.0b10. Node support in the resulting trees was estimated by 1000 bootstrap replicates using a heuristic search. To confirm the phylogenetic groupings obtained by Neighbour-Joining for HmB453K, a maximum likelihood phylogeny was also constructed with PhyML [50], using the GTR+I+G substitution model selected by Modeltest [51] and with branch support values obtained by 1000 bootstrap replicates. The stability of the inferred phylogeny for HmB453k was examined using the Shimodaira-Hasegawa test (SH test) [35] in PAUP* 4.0b10. For all phylogenetic inferences trees were rooted with H. numata as outgroup.
Analysis of molecular variance (AMOVA) with 1000 permutations, implemented in ARLEQUIN v.3.5 [52], was used to assess population structure by species, geography and phenotype. For species, four groups were set, corresponding to H. melpomene, H. cydno, H. timareta and H. heurippa. In the geography analysis, haplotypes were grouped into six geographic regions: (i) the Guiana shield, (ii) Amazon, (iii) Pacific, (iv) East Andes foothills, (v) Cauca Valley and (vi) Magdalena Valley. These geographic clustering matches the biogeographic provinces (i) Humid Guyana, (ii) Napo+Imeri, (iii) Choco+Wester Ecuador+Arid Ecuador, (iv) North Andean Paramo, (v) Cauca and (vi) Magdalena defined by Morrone [53]. When compared by phenotype, haplotypes were grouped in three groups: the red-banded type [presence of red forewing band], the rayed type [presence of orange rays in the hind wing] and the non-red type [absence of any dorsal red element on the wings]. The outgroup H. numata was excluded from these analyses.
In order to estimate the role and direction of historical gene flow between H. melpomene and H. timareta (H. heurippa was included in H. timareta for the purposes of this analysis), we used the Isolation-Migration (IM) Bayesian model [36]. IM uses Markov chain Monte Carlo (MCMC) sampling to obtain maximum-likelihood estimates of six parameters: current population sizes, ancestral population size, rates of migration between two populations (m1 and m2) and the timing of divergence (t). IM assumes both free recombination between loci and no recombination within them, therefore the software SITES [54] was used to select genetic blocks with no recombination within each locus. To fulfill the assumption of free recombination between loci, only the unlinked colour loci and one of the fragments linked to red colouration [HmB453k] were selected for this analysis. We ran IM on two modified datasets for each species pair: (i) H. melpomene rayed type-H. timareta and (ii) H. melpomene red-banded -[H. heurippa and H. timareta]. These groups constituted the maximal units in which we could get enough data without recombination, with the rayed dataset being a block of 379 bp and the postman dataset one of 313 bp. Unfortunately, pairwise comparisons involving all the species' alleles were not possible nor were comparisons involving species in parapatry because such groupings contained small non-recombining blocks that lacked enough informative sites. However, since our main interest was to determine the direction and magnitude of introgression (m) within phenotypes, these datasets are sufficient for addressing this question. For all datasets, after searching for the parameter range using preliminary runs, 30 million steps were sampled from the primary chain after a 300,000 burn-in period under the HKY model with 10 chains per set. Mixing properties of the MCMC were assessed by visual inspection of the parameter trend plots and by examining that the effective sample size (ESS) was higher than 50, as recommended [36], [37].
To get biologically meaningful units of gene flow, the maximum likelihood estimates and 90% highest posterior density (HPD) interval for the migration rates (m) were converted into the effective number of gene migrations received by a population per generation (2Nm, in Table 2). For this conversion, we used a generation time of 35 days and a mutation rate per gene calculated with the calibration time proposed by Wahlberg et al. for Nymphalidae [55] coupled with the divergence between the melpomene/cydno clade per locus estimated with the software SITES. Our estimates of mutation rate per locus per year were: 6.2×10−6 for COI, 7.1×10−8 for GAPDH, 1.2×10−6 for Hsp90 and 3.4×10−5 for HmB453k.
We finally compared the model including all six parameters to simpler demographic models in order to statistically test the hypothesis of zero or equal gene flow between populations (m1 = 0, m2 = 0, m1 = m2 = 0, m1 = m2>0). These analyses were conducted using IMa [36] by running the initial M-mode output with identical settings in the L-mode and sampling 5×105 genealogies.
We further tested the presence, significance and direction of gene flow per locus between H. melpomene and H. timareta using a method based on linkage disequilibrium (LD) developed by Machado et al in 2002 [38]. In this test, a positive difference between the LD among all pairs of shared polymorphisms (DSS) and the LD among all pairs of sites for which one member is a shared polymorphism and the other is an exclusive polymorphism (DSX), is indicative of gene flow. The magnitude of the difference directly measures the direction of the introgression, with the species with the highest positive value being the recipient [38]. The same phenotypic datasets used in the IM analysis and also groups of species in complete sympatry, were subjected to independent runs, each of them with 30000 simulations. The input files were prepared with the program SITES [56] calculating D' as a measure of linkage disequilibrium (as suggested by Machado et. al [38]) and analysing linkage disequilibrium among shared polymorphism between groups (by choosing the -s and -p options in the LD string).
Linkage disequilibrium across the HmB region was calculated for all populations using the software MIDAS [57] only considering sites with allele frequency higher than 5%, and visualised with the R package LDheatmap [58].
|
10.1371/journal.pgen.1005093 | Male-Biased Aganglionic Megacolon in the TashT Mouse Line Due to Perturbation of Silencer Elements in a Large Gene Desert of Chromosome 10 | Neural crest cells (NCC) are a transient migratory cell population that generates diverse cell types such as neurons and glia of the enteric nervous system (ENS). Via an insertional mutation screen for loci affecting NCC development in mice, we identified one line—named TashT—that displays a partially penetrant aganglionic megacolon phenotype in a strong male-biased manner. Interestingly, this phenotype is highly reminiscent of human Hirschsprung’s disease, a neurocristopathy with a still unexplained male sex bias. In contrast to the megacolon phenotype, colonic aganglionosis is almost fully penetrant in homozygous TashT animals. The sex bias in megacolon expressivity can be explained by the fact that the male ENS ends, on average, around a “tipping point” of minimal colonic ganglionosis while the female ENS ends, on average, just beyond it. Detailed analysis of embryonic intestines revealed that aganglionosis in homozygous TashT animals is due to slower migration of enteric NCC. The TashT insertional mutation is localized in a gene desert containing multiple highly conserved elements that exhibit repressive activity in reporter assays. RNAseq analyses and 3C assays revealed that the TashT insertion results, at least in part, in NCC-specific relief of repression of the uncharacterized gene Fam162b; an outcome independently confirmed via transient transgenesis. The transcriptional signature of enteric NCC from homozygous TashT embryos is also characterized by the deregulation of genes encoding members of the most important signaling pathways for ENS formation—Gdnf/Ret and Edn3/Ednrb—and, intriguingly, the downregulation of specific subsets of X-linked genes. In conclusion, this study not only allowed the identification of Fam162b coding and regulatory sequences as novel candidate loci for Hirschsprung’s disease but also provides important new insights into its male sex bias.
| Hirschsprung’s disease (also known as aganglionic megacolon) is a severe congenital defect of the enteric nervous system (ENS) resulting in complete failure to pass stools. It is characterized by the absence of neural ganglia (aganglionosis) in the distal gut due to incomplete colonization of the embryonic intestines by neural crest cells (NCC), the ENS precursors. Hirschsprung’s disease has an incidence of 1 in 5000 newborns and a 4:1 male sex bias. Although many genes have been associated with this complex genetic disease, most of its heritability as well as its male sex bias remain unexplained. Here, we describe an insertional mutant mouse line (“TashT”) in which virtually all homozygotes display colonic aganglionosis due to defective migration of enteric NCC, but in which only a subset of homozygotes develops megacolon. Surprisingly, this group is almost exclusively male. The TashT ENS defect stems, at least in part, from the disruption of long-range interactions between evolutionarily conserved elements with silencer activity and Fam162b, resulting in NCC-specific upregulation of this uncharacterized protein coding gene. Global analysis of gene expression further revealed that several hundreds of genes are significantly deregulated in TashT enteric NCC. Interestingly, this dataset includes multiple X-linked candidate genes potentially underlying the male sex bias. Taken together, our data pave the way for a clearer understanding of the intriguing male sex bias of Hirschsprung’s disease.
| The enteric nervous system (ENS) is the intrinsic neural network of the gastrointestinal tract. One of its essential roles is to regulate intestinal motility. The ENS is made up of interconnected neural ganglia, themselves composed of neurons and supporting glial cells, forming two main parallel networks: the submucosal plexus and the myenteric plexus. The muscles of the bowel wall that ensure peristaltic movements are controlled by the myenteric plexus of the ENS.
The ENS is constructed during embryo development by derivatives of migrating neural crest cells (NCC) [1]. These multipotent cells originate from the dorsal part of the neural tube, undergo an epithelial-mesenchymal transition, and migrate extensively to contribute to numerous embryonic structures. Among several different cell types, NCC generate melanocytes as well as all enteric neurons and glia. The developing bowel is mainly colonized by NCC derivatives originating from the vagal region of the neural tube. Such colonization proceeds as a rostro-caudal wave lasting more than 5 days in the mouse (from embryonic day (e) 9.0 to 14.5), with NCC derivatives first entering the foregut, passing through the midgut (prospective small intestine) and finally populating the hindgut (prospective colon) either by migrating through the intestinal mesenchyme [2] or by taking a shortcut via the mesentery [3]. The hindgut is the last part of the intestines to be colonized and, therefore, the most susceptible to enteric NCC (eNCC) developmental defects. Sacral NCC also contribute to the ENS, but this later contribution is minor and cannot compensate for a lack of vagal NCC [4].
Defects in hindgut colonization by eNCC result in a lack of neural ganglia in the colon, leading to intestinal blockage due to absence of peristalsis. This phenotype is generally described as “aganglionic megacolon” because of the subsequent massive accumulation of fecal material and severe distention of the colon. In humans, this condition is called Hirschsprung's disease (HSCR) and, depending on the length of aganglionosis, is clinically subdivided in short-segment (i.e. restricted to the rectosigmoid colon) and long-segment forms. Short-segment HSCR represents the vast majority of cases and is more common in males than females, with an overall ratio of ~4:1 [5]. In patients displaying longer segments of aganglionosis, the sex bias is much less pronounced or absent altogether. Although mutations in at least 15 genes have been implicated in HSCR, heritability is unexplained for the majority of cases [6]. HSCR is thus a classic example of a complex disease involving multiple genes, incomplete penetrance, variable expressivity and an intriguing male bias.
Most known HSCR-associated genes encode players from two signaling pathways: the GDNF ligand/ RET receptor and EDN3 ligand/ EDNRB receptor pathways. In fact, RET is the main gene associated with HSCR [5]. For both pathways, the receptor is found at the surface of eNCC while the ligand is dynamically secreted from the surrounding mesenchyme during the colonization phase. The role of GDNF/RET and EDN3/EDNRB signaling in ENS formation has been well conserved evolutionarily and studies in animal models have revealed that both pathways profoundly influence every key aspect of eNCC development such as proliferation, survival, differentiation and, most especially, migration [7]. Mouse models have been particularly informative in this regard and multiple lines bearing mutation—either spontaneous or targeted—of genes encoding members of Gdnf/Ret and Edn3/Ednrb pathways have been studied [8–13]. However, the incomplete penetrance and, above all, the male bias observed in human HSCR have been poorly replicated in current animal models [14].
Here, we report the creation of a new insertional mutant mouse model for HSCR that displays, for the first time, incomplete penetrance of the aganglionic megacolon phenotype with a very strong male bias. Extensive characterization of this mouse line and independent validation via transient transgenesis indicate that this outcome is, at least in part, initiated by the specific upregulation of Fam162b in NCC.
The TashT mouse line was obtained from an insertional mutagenesis screen for genes involved in NCC development. This screen was based on the random insertion of a Tyrosinase (Tyr) minigene in the FVB/n genetic background. Owing to its specific expression in melanocytes, the Tyr minigene rescues the albino phenotype of FVB/n mice and thus provides a visible—and generally uniform—pigmentation marker for transgenesis [15]. Since melanocytes are derived from NCC, this genetic tool also proved to be a potent indicator of abnormal NCC development via identification of non-uniform pigmentation patterns. This approach yielded several transgenic mutant lines (to be described elsewhere) among which TashT (Tachetée, in French) was identified due to its variegated pigmentation (Fig. 1a).
In addition to the pigmentary anomalies that are similar in both heterozygous and homozygous mutants, a subset of TashTTg/Tg animals suffer from aganglionic megacolon around weaning age. These animals are smaller than unaffected TashTTg/Tg siblings (reaching about 74% of littermate weight) and exhibit bowel obstruction concomitant with lack of myenteric ganglia in the distal colon (Fig. 1a,c). The most striking and interesting feature of this lethal phenotype is the fact that the vast majority of affected animals are male (Fig. 1b). In addition, we found via histological analyses that colonic aganglionosis can not only be detected in megacolon-suffering but also in non-affected TashTTg/Tg animals of both sexes (Fig. 1c). To determine whether megacolon expressivity could be correlated with extent of aganglionosis, we undertook a systematic analysis of the length of the colonic ENS for a random group of TashTTg/Tg animals of weaning age via staining of acetylcholinesterase activity (Fig. 1d). In accordance with such correlation, quantification results first revealed that—regardless of the sex of the animal—a minimal length of the colon (~ 80%; critical region in Fig. 1e) has to be properly innervated in FVB/n mice in order to avoid blockage. No intermediate phenotype was noted in the course of these analyses as none of the non-affected animals showed signs of abnormal accumulation of feces in the colon. Furthermore, these results confirmed that most TashTTg/Tg animals exhibit aganglionosis in the distal colon and that males (ganglionated on average over 79% of total colon length) are more affected than females (ganglionated on average over 89% of total colon length) (Fig. 1e). Importantly, a similar statistically significant difference between male and female animals (74% vs 90%) was also observed in neonatal colons, thus confirming the developmental origin of this male-biased defect.
To analyze ENS formation in TashT embryos, we took advantage of the fact that this line bears a second co-injected transgene (pSRYp[1.6kb]-YFP) that labels migrating NCC derivatives (including eNCC of vagal and sacral origin) with YFP fluorescence [16] (see also S1 Fig). Whole-mount detection of fluorescence in dissected stage-matched embryonic intestines revealed that, in comparison to TashTTg/+ or G4-GFP control embryos [17] (S2 Fig), a colonization delay by eNCC of vagal origin is clearly observed for TashTTg/Tg embryos starting around e11.0 (Fig. 2a). Although ectopic fluorescence in the cecum and proximal hindgut regions impeded precise visualization of the migration front between e11.5 and e14.5, we found that this delay persists through the time at which normal colonization of the digestive tract is overtly completed (e15.5). The presence of scattered fluorescent cells beyond the chains of vagal-derived eNCC at this stage also suggests that the contribution of sacral-derived eNCC to the distal hindgut is not abrogated in TashTTg/Tg embryos.
Closer inspection of the migration front at e11.0 showed that cell protrusions in the form of filopodia were not overtly affected in TashTTg/Tg eNCC (S3 Fig), suggesting that cells could still investigate, and had the capacity to respond to, their environment. To further characterize the colonization defect, leader cells at the tip of eNCC chains were then visualized during several hours in ex vivo cultures of e11.0 intestines from littermate control (TashTTg/+) and mutant (TashTTg/Tg) embryos (Figs. 2b, S4 and S1–S2 Videos). It is noteworthy that TashTTg/Tg intestines were selected for these analyses on the basis of the severity of their colonization defect in order to increase the odds of detecting differences between control and mutant eNCC. Using these conditions, we found that the average migration speed, and therefore travel distance, is almost halved in TashTTg/Tg leader eNCC while directionality of migration is not noticeably affected (Fig. 2c). Given the selection bias towards more affected embryos, it is important to bear in mind that this severe effect is most likely not representative of the true average migration speed of mutant eNCC. It should also be noted that the small difference in eNCC location along the intestines—before cecum (TashTTg/Tg) vs entry of cecum (TashTTg/+)—cannot account for the observed difference in speed since the eNCC migration front normally displays a fairly stable net speed of ~35 micron/hour (~0.58 micron/min) between e10.5 and e12.5 [18].
Migration of eNCC is dependent on multiple signaling pathways among which GDNF/RET and EDN3/EDNRB are recognized as the most critical regulators [19–22]. To evaluate the status of these signaling pathways in TashT embryos, we made use of a recently described quantitative migration assay using e12.5 midgut explants [23]. With control TashTTg/+ or G4-GFP tissues, collagen gels containing either GDNF or EDN3 increased the number of cells coming out of the explants (Fig. 2d). Interestingly, when these ligands were used in combination, a synergistic increase in eNCC numbers invading the collagen was observed. However, little reaction to these extracellular ligands was detected in eNCC derived from TashTTg/Tg embryos (Fig. 2d). In fact, eNCC from homozygous embryos migrated out of intestinal explants even in the absence of exogenous ligands, suggesting they lost some sensitivity to their endogenous microenvironment and distinguished poorly between intestinal tissue and collagen gel.
Premature differentiation or a scarcity of progenitor cells can disturb eNCC colonization and lead to incomplete ENS formation [19,22,24,25]. To verify whether these processes might contribute to the TashT phenotype, we performed a detailed marker analysis of embryonic intestines in order to quantify neuronal and glial differentiation as well as proliferation and cell death of eNCC. Quantification of proliferation and cell death in e12.5 stage-matched bowel tissues failed to reveal any significant difference between TashTTg/Tg and control TashTTg/+ embryos (S5 Fig). Assessment of neuronal differentiation at the same stage also failed to reveal any significant difference (S6a-S6b Fig). On the other hand, glial differentiation at e15.5 was found to be less prevalent in TashTTg/Tg distal bowel tissues, to the benefit of undifferentiated progenitors (S6c-S6d Fig). This, however, is most likely a consequence of the delay in rostro-caudal colonization, and goes contrary to the idea that premature differentiation is the cause of the TashT migration defect. Taken together, these results thus highlight the eNCC migration defect (concomitant with insensitivity towards GDNF and EDN3) as the principal cause of the TashT aganglionosis phenotype. Moreover, since no clear sex bias was observed in these analyses, the relatively modest male bias in phenotype severity is most likely the result of an accumulation of subtle differences during the whole ENS developmental time window.
Breeding of the TashT line revealed systematic co-segregation of pigmentation with YFP fluorescence, meaning co-integration of both transgenes into a single autosomal locus which is frequent when an equimolar mixture of each transgene is micro-injected [15]. FISH analysis first allowed a rough estimate of the localization of the transgene insertion site on chromosome 10 at bands B2–B3 (S7a Fig). To obtain a more precise localization, we sequenced the whole genome of a TashTTg/Tg mouse. Mapping of high-throughput paired sequencing reads allowed us to localize the transgenic insertion around the middle of a 3.3Mb gene desert between Hace1 and Grik2 (Fig. 3a,b). Twice as many reads were observed in a 26kb non-coding region of chromosome 10B2, indicating a duplication. Flanking this duplicated region were paired reads with one end mapping to chromosome 10 and the other end mapping to sequences corresponding to either one or the other transgene (Fig. 3a). A schematic representation of the inferred organization of the TashT transgene insertion site is shown at the bottom of Fig. 3a. The number of transgene copies was estimated from the mapping data and the total size of the insertion calculated to be about 700kb.
The TashT locus contains several blocks of evolutionary conserved non-coding sequences (Fig. 3b). To evaluate their regulatory potential, we cloned seven ~1kb fragments containing most of these constrained elements (CE)—named CE1 to CE7—in a luciferase expression vector bearing a minimal thymidine kinase promoter. Transcriptional activity was then assessed in various cell lines (Neuro-2a, P19, Cos7 and NIH 3T3) via luciferase assays (Figs. 3c and S9). Overall, this analysis revealed very strong repression activity in a cell type-independent manner for two of the cloned regions (CE5 and 6) (Figs. 3c and S9). These luciferase assays thus suggest that the TashT transgenic insertion has disrupted at least one important long-range regulatory element that normally represses expression of a surrounding gene. However, expression of the two most proximal neighboring genes on each side of the gene desert (Lin28b and Hace1 as well as Grik2 and Ascc3) (Fig. 3b) was found to be similar in TashTTg/Tg and control G4-GFP e12.5 eNCC recovered by FACS (S8 Fig).
To cast a wider net and detect transcript variation in an unbiased manner, we sequenced the rRNA-depleted transcriptome of FACS-recovered eNCC from anterior intestinal tissues of stage-matched control (G4-GFP) and TashTTg/Tg e12.5 embryos. This stage was chosen because it combines ease of intestine dissection with clear presence of the eNCC colonization defect in TashTTg/Tg tissues. It is also important to note that this analysis was restricted to anterior intestinal tissues only (prospective oesophagus, stomach and small intestine) because endogenous YFP labelling in the TashT line is strictly specific to eNCC in these regions (S1b Fig). As mentioned above (see Fig. 2a), the TashT line exhibit ectopic YFP fluorescence in non-NCC derivatives in more posterior regions (prospective cecum and colon) and, therefore, these regions were excluded from both control and mutant cell preparations.
Analysis of RNAseq data revealed that over 1200 coding and non-coding genes are differentially expressed in a significant manner (≥2-fold and p <0.001) in TashTTg/Tg eNCC, among which upregulated and downregulated genes are equally represented (S1 Dataset and S10c Fig). The most deregulated genes (≥4-fold) are listed in Table 1 and this shorter list now indicates a strong enrichment for upregulated genes in TashTTg/Tg eNCC (41 downregulated vs 188 upregulated). Each of these 229 genes was manually assigned to a category based on function and/or localization of their gene product. Assembling these categories in “super-categories” reveals that TashT affected genes are mostly involved in the control of cell signaling (categories: Ligand-receptor and Signal transduction) and gene expression (category: Transcription factor) as well as in the composition of, and interaction with, the cell microenvironment (categories: Extracellular matrix, Cell adhesion as well as Channel and transmembrane transport). Especially notable examples within each of these super-categories include, respectively, genes encoding Gdnf and Edn3 ligands, many Hox transcription factors as well as various Collagen members. Another category worth mentioning is the Metabolic pathway which notably contains many players of retinoid signaling (Aldh1a1, Aldh1a2, Aldh1a7 and Rdh10). We verified the expression level of selected genes by semi-quantitative RT-PCR. Our selection criteria included genes known as playing a major role in HSCR (Table 2 and S10a Fig) as well as genes located on a sex chromosome whose change in expression could explain the observed male bias (i.e. upregulated on chromosome Y or downregulated on chromosome X) (Table 3 and S10b Fig). All genes tested followed the trend set by the RNAseq data.
Given that intra-chromosomal regulatory chromatin contacts are much more prevalent than inter-chromosomal ones [26], the selection criteria for identification of the TashT causative gene was its location relative to the repressive elements disrupted by the transgene insertion. We therefore focused on the handful of upregulated genes located on chromosome 10 (Table 4) in our search for a gene directly regulated by the conserved elements. The closest candidate, Fam162b, is ~3.6 Mb telomeric to the conserved elements (Fig. 3b) and its overexpression in TashTTg/Tg eNCC was validated via semi-quantitative RT-PCR (S11a Fig). Note that Fam162b mRNA can also be detected in eNCC FACS-sorted from control embryos, indicating its low level expression in normal eNCC populations (see S1 Dataset). Analyses of Fam162b open-reading frame sequences using ExPASy (www.expasy.org) and Uniprot (www.uniprot.org) resources suggest that this uncharacterized gene encodes a single pass transmembrane protein localized to mitochondria.
We subsequently verified that the conserved silencer elements near the transgene insertion site normally interact with the Fam162b locus using chromosome conformation capture (3C) techniques [27]. These analyses first revealed that such interaction can be detected in wild-type whole embryonic intestines using different primer pairs (Fig. 4a,b). We also found that this interaction is cell type-specific as it is detected in NCC-derived Neuro-2a cells but not in undifferentiated P19 cells (Fig. 4c), the same murine embryonic cell lines used in our luciferase assays (Fig. 3c). Importantly, we further found that this interaction is lost in TashTTg/Tg embryonic gut tissues (Fig. 4b). Therefore, these results strongly suggest that the TashT transgenic insertion disrupts intra-chromosomal contacts that normally repress Fam162b expression in NCC (Fig. 4d).
To independently validate the candidacy of Fam162b as being involved in the TashT ENS defect, we generated transgenic e15.5 embryos specifically overexpressing Fam162b in NCC and analyzed the impact on ENS formation using specific markers. The transgenic construct consisted of a bicistronic cassette containing Fam162b and eGFP coding sequences driven by a previously described Sox10 enhancer (U3, also known as MCS4) fused to the Hsp68 minimal promoter [28,29]. Western blotting confirmed that the cloned Fam162b sequences express a protein of expected size (S11b Fig). Using fluorescence from eGFP as a surrogate marker for transgene expression, we obtained a total of three Fam162b transgenic embryos (all females) exhibiting variable levels of transgene expression. In contrast to littermate controls for which bowel tissues were fully colonized, all three transgenic embryos displayed incomplete colonization of the distal hindgut by vagal-derived eNCC and, as a result, this region was found to only contain scattered Sox10-positive eNCC of presumably sacral origin in a way similar to what is observed in TashTTg/Tg tissues (compare Fig. 4e with Fig. 2a). In addition to having used a different genetic background for these experiments (B6C3), we believe that the fact that all Fam162b overexpressers were female likely explains, at least in part, the more modest effect observed in comparison to TashTTg/Tg embryos. Although we cannot currently exclude the possibility that other gene(s) might also be primary target(s) of the TashT mutation and might thus also contribute to the TashTTg/Tg phenotype, these transgenesis data support a role for Fam162b overexpression in TashTTg/Tg pathogenesis.
We have identified and characterized a novel mouse model for aganglionic megacolon, called TashT, allowing us to describe the eNCC transcriptome during development and to propose two novel genetic loci implicated in the pathogenesis of megacolon—one that is protein coding, the other involving evolutionary conserved non-coding sequences—as well as an ultra-long-range interaction between these loci. In addition, this mouse model allows us to speculate on the mechanistic nature of the hitherto unexplained male bias of the megacolon phenotype observed in humans, as well as its variable penetrance.
To the best of our knowledge, this study is the first to report a transcriptome analysis of sorted eNCC. Previous screens for genes expressed in eNCC were performed on whole embryonic intestines using DNA microarrays and based on the differential expression between normal and Ret-null aneural tissues [30,31]. In addition to analyzing eNCC directly, we took full advantage of the RNAseq technology and included non-coding RNA in our analyses. This resulted in a much more extensive list of genes known to be expressed in eNCC, from a few hundred to several thousand (S1 and S2 Datasets).
Importantly, the transcriptional signature of TashTTg/Tg eNCC not only highlighted Fam162b as a potential causative gene, but also provided mechanistic insights into the identified cell migration defect. In this regard, it is noteworthy that several modulated transcripts in TashTTg/Tg eNCC encode components of the extracellular matrix (ECM) (Table 1 and S10c Fig). NCC have been suggested to modify their extracellular environment during, or perhaps as a requirement for, migration [32,33]. One possibility is thus that the modulated ECM in TashTTg/Tg embryonic intestines is less permissive to cell migration [34,35]. Another, not mutually exclusive possibility is that the reduced sensitivity of TashTTg/Tg eNCC to growth factors/chemoattractants normally present in the intestinal ECM underlies the migration defect. In agreement with this, we have found that TashTTg/Tg eNCC have lost their ability to respond to exogenous GDNF and EDN3 in explant assays, with eNCC appearing unable to distinguish between bowel tissue and collagen gel (Fig. 2d). This latter outcome is most likely due to the surprising robust overexpression of both Gdnf and Edn3 by TashTTg/Tg eNCC (Table 2 and S10a Fig). Indeed, given that Gdnf and Edn3 are both normally heavily secreted from the surrounding mesenchyme, additional oversecretion from eNCC is expected to disturb the dosage of ligands these cells normally encounter and/or to disrupt any gradient that might be present. Our RNAseq data also suggest that the lack of responsiveness to GDNF and EDN3 might be due to an overabundance-induced negative feedback on their cognate receptor. This hypothesis is supported by the fact that expression of both Ret and Ednrb is reduced in TashTTg/Tg eNCC (Table 2 and S10a Fig). Interestingly, in the case of Gdnf/Ret signaling, this hypothesis is further supported by the observed overexpression of Lrig1 and Lrig3 in TashTTg/Tg eNCC (S1 Dataset). These genes encode functionally-redundant transmembrane proteins [36] which, as specifically demonstrated for Lrig1 in neuronal cells, can be induced by Gdnf at the transcriptional level and then physically interact with Ret in order to reduce Gdnf binding and tyrosine kinase activity [37]. Regardless of the exact underlying mechanism, a lack of responsiveness to EDN3 might well be responsible for the slower migration of TashTTg/Tg eNCC (Figs2b-c and S4), as inhibition of Ednrb signaling has been recently shown to primarily affect the speed of eNCC migration [38].
Most of the known and characterized long-range acting regulatory elements do not interact with their nearest promoter but bypass several intervening genes in order to reach their target promoter [39]. Long-range enhancer-promoter interactions are also thought to be more commonly involved in the regulation of tissue-specific genes [40]. Moreover, most studies of intra-chromosomal long-range interactions involve loci up to several hundred kb away from each other, though ultra-long-range events (several Mb) between enhancer and promoter are not uncommon [27,40,41]. Our 3C data are in accordance with these observations and indicate that the interaction between the conserved elements near the TashT transgene insertion site and the Fam162b gene ~3.6 Mb away (Fig. 4b) falls in the ultra-long-range category. Little is known of the spatiotemporal pattern of Fam162b expression. The evidence to date is in agreement with an expression in neural derivatives: it is weakly expressed in the mouse olfactory bulb (Allen Brain Atlas, RP_051012_01_G06) and expressed in the frontonasal prominence of mouse embryos, proximal to the oral cavity [42]—a tissue heavily populated by cranial NCC. These observations are consistent with the idea that Fam162b is poised for active transcription in neural and/or neural crest cells but kept in check through a repressive mechanism.
Using luciferase assays, we demonstrated that a subset of the highly conserved elements near the TashT transgene insertion site has a robust negative regulatory function on transcription in a cell type-independent manner (Fig. 3c and S9 Fig). Given the forced juxtaposition of regulatory elements with the proximal promoter in such assays, the absence of a cell type-specific activity in our analysis thus points to a chromatin conformation-dependent mechanism conferring specificity in vivo. As supported by our 3C data (Fig. 4c), we suggest that an ultra-long-range chromatin loop maintains Fam162b expression at a basal level in a subset of neural-derived cells, including eNCC. Further investigations into the regulation of Fam162b expression will be necessary to confirm and expand this hypothesis. The biological function of the Fam162b gene product is also currently unknown. Characterization of this function will clearly be facilitated by the wealth of information obtained from the RNAseq data as well as by the observed cell migration defect.
Robust correlation between extent of aganglionosis and expressivity of the megacolon phenotype in TashTTg/Tg animals allowed us to identify the minimal distance of myenteric innervation necessary for successful movement of luminal content across the colon in FVB/n mice. This critical region (~80% of colon length; Fig. 1e) represents a threshold level beneath which intestinal blockage occurs systematically, and in a sex-independent manner. The fact that the mean length of the ganglionated region of TashTTg/Tg males ends in this critical region while females typically show a more extensive ENS explains the apparent contradiction between the male bias in megacolon expressivity and the near complete penetrance of distal aganglionosis in both sexes. A common defect of eNCC colonization, slightly exaggerated in males, is thus the source of the observed sex bias of the megacolon phenotype in TashTTg/Tg animals. In this regard, it is noteworthy that a similar link between extent of aganglionosis and megacolon expressivity has been previously described in mice bearing the Ednrbs-l allele [43] as well as in Ret+/-::Ednrbs/s compound mutants [14]. Interestingly, although only a very modest male sex bias in megacolon expressivity was reported in this latter case (~1.5:1), the correlations made with extent of aganglionosis are in agreement with the threshold level revealed by our study. Indeed, full penetrance of megacolon in Ret+/-::Ednrbs/s males was correlated with a mean length of the ganglionated region clearly beneath the threshold (59% of colon length) whereas partial penetrance of megacolon in Ret+/-::Ednrbs/s females was correlated with a mean length of the ganglionated region much closer to the threshold (72%) [14]. However, as evidenced by the fact that Sox10Dom mutants on a C57BL/6J—C3HeB/FeJ mixed background display a shorter aganglionic zone leading to megacolon (~10%), it should also be noted that position of the threshold level may vary as a function of the genetic background [44].
The aganglionic megacolon of TashTTg/Tg animals share striking similarities with both the variable penetrance and male sex bias of short-segment HSCR, the most common form of the disease (~80% of cases). The threshold level identified with the TashT line is also in accordance with the fact that virtually no sex bias is observed in long-segment HSCR. Our analysis of the TashT line thus provides useful insights into the ontogeny of aganglionic colon and the origin of the sex bias, and shows that, though perhaps suffering from chronic constipation, TashTTg/Tg mice are nevertheless able to pass intestinal material when more than 4/5 of their colon is innervated.
Apart from Ret+/-::Ednrbs/s compound mutants, it is interesting to note that a male bias in the extent of aganglionosis—but not megacolon expressivity—has also been reported in other mouse and/or rat models and in each case implicated a mutation in either Ret or Ednrb [13,45,46]. Taken together with our data showing deregulated Ret and Ednrb signaling in TashTTg/Tg animals as well as with the previous description of a RET non-coding mutation that is twice as frequently transmitted in boys than in girls [47], these observations strongly suggest that both pathways are involved in the regulation of expression of a still undefined sex chromosome-linked gene with critical function in the developing ENS. We reasoned that exaggerated defects in males could arise from overexpression of male-specific genetic material (upregulated genes on Y chromosome) or from a deficiency in the expression of genes present as single alleles (downregulated genes on X chromosome). Another interpretation of this second possibility is that females are protected through biallelic expression of some X chromosome genes, provided they escape X-inactivation [48–51]. Analysis of our transcriptome dataset revealed that expression of Y-linked genes in eNCC is limited to a group of four genes (Kdm5d, Eif2s3y, Uty and Ddx3y) (S2 Dataset). Of these clustered genes—also known to be expressed in the developing brain [52,53]–, Ddx3y is the only one that approaches significant upregulation in TashTTg/Tg eNCC (1.5-fold; edgeR p = 0.0011; DESeq p = 0.0055). In marked contrast, multiple X-linked genes were found to be significantly downregulated in TashTTg/Tg eNCC, including Dcx—a previously suggested potential HSCR susceptibility locus (Table 3) [30]. As X-inactivation escapees tend to be found in clusters [48], other interesting candidates include a group of genes—in the vicinity of Dcx—that contains Bex1 and the Plp1-Rab9b gene pair (Table 3). It is noteworthy that the candidacy of these genes is also supported by human cases with Xq22 microdeletions encompassing them. Indeed, while such cases are assumed to be embryonic lethal in males, female patients suffer from Pelizaeus-Merzbacher-like disease with symptoms of gastrointestinal motility problems including constipation [54].
Work with mice was performed in accordance with the guidelines of the Canadian Council on Animal Care (CCAC) and approved by the relevant institutional committee (Comité institutionnel de protection des animaux; CIPA reference #650) of University of Quebec at Montreal (UQAM). Mice were euthanized by gradual-fill carbon dioxide (CO2) gas preceded by isoflurane anesthesia. TashT transgenic mice and Fam162b transgenic embryos were generated via standard pronuclear microinjection [55], using embryos derived from FVB/n albino and B6C3 mice, respectively. For the TashT line, two transgenes were co-injected at equimolar ratio: a Tyrosinase minigene to allow visual identification of transgenic animals via rescue of pigmentation [15] and a pSRYp[1.6kb]-YFP construct that provides fluorescent marking of migrating NCC in the developing embryo [16]. The previously described Gata4p[5kb]-GFP line (G4-GFP) was used as wild-type control [17]. Mice were mated overnight and noon on the day a vaginal plug was observed was designated as embryonic day (e) 0.5. For Fam162b transient transgenics, a transgene carrying the PCR-amplified Fam162b open reading frame under the control of the Hsp68 minimal promoter [56] and a NCC-specific Sox10 enhancer (U3, also known as MCS4) [28,29] was used. In order to provide a positive control for transgene expression, an IRES-GFP cassette (pIRES2-EGFP, Clontech) was also included immediately downstream of the Fam162b ORF, creating a bicistronic message. Fifteen days after microinjection, foster mothers were sacrificed, embryos were collected, individually analyzed for GFP and immunostained for Sox10 and βIII-Tubulin (see immunofluorescence section below).
Antisense and sense digoxigenin-labelled RNA probes were synthesized using a DIG transcription kit (Roche). Mouse spleen lymphocytes were collected and metaphase slides prepared using standard cytogenetic protocols [57,58]. Slides were aged at room temperature for 7 days, and then GTG banded, again using standard protocols [55]. Slides were scanned at 100X using a Nikon Eclipse E800 microscope, and representative metaphases were photographed at 1000X using a Nikon DXM1200 digital camera and SimplePCI software. Slides were de-stained using CitriSolv (Fisher) for 10–15 minutes, fixed for 1 minute in a 1% paraformaldehyde (PFA) solution in 2X SSC then dehydrated in graded ethanol washes and stored for future use in 100% ethanol. Slides were air dried then incubated in a denaturation solution for 5 minutes at 73°C, then again dehydrated in graded ethanol washes and stored in 100% ethanol. Probe was generated using a digoxigenin-labeled Tyr minigene DNA fragment [57]. Just prior to adding the denatured FISH probe, slides were air dried. Five to 10 μl of denatured probe was added per slide, which was then incubated overnight at 37° under a coverslip in a humidified dark atmosphere. Detection was performed using a rhodamine-conjugated anti-DIG antibody (Roche), following the manufacturer’s instructions. Slides were counterstained with DAPI (Sigma) and antifade (P-phenylenediamine; Sigma). Previously photographed G-banded metaphases were re-located using epifluorescence, rephotographed, and compared to FISH images.
Intestines were dissected from e12.5 G4-GFP and TashT embryos without their posterior end (prospective cecum and colon) because, in contrast to more anterior regions in which only eNCC are fluorescently-labelled (see S1b Fig), TashT intestines also contain high amounts of YFP-positive mesenchymal cells that are not derived from NCC in these regions (see Figs. 2a and S4). A MoFlo XDP (Beckman Coulter) cell sorter was used to collect GFP- or YFP-positive single viable cells from the remaining dissociated intestinal tissue. Dissociation was carried out at 37°C in EMEM containing collagenase (0.4 mg/ml; Sigma C2674), dispase II (1.3 mg/ml; Life Technologies 17105–041) and DNAse I (0.5 mg/ml; Sigma DN25).
Whole genome and transcriptome library generation and sequencing was performed by McGill University and Génome Québec Innovation Centre using the HiSeq 2000 platform (Illumina). Fifty to 150 million paired-end sequences (100 bp length) were obtained from 300–500 bp library inserts, resulting in an overall 30x coverage for genome reads. Sequences were filtered based on quality and mapped onto the Mus musculus reference genome (mm9 for genomic DNA, mm10 for RNA). For the transcriptome, total RNA was rRNA-depleted before making the libraries. Three libraries were generated for each cell population, though only two were of sufficient quality for subsequent bioinformatic analysis. DESeq and edgeR differential gene expression analyses (adjusted p-value < 0.001), as well as a minimum 2-fold expression difference, were taken into account to determine significantly deregulated messages between FACS-recovered eNCC from control (G4-GFP) and TashTTg/Tg embryos. Gene Ontology analyses were performed using GOToolBox, and at least 2-fold enriched/depleted categories were selected from a hypergeometric test with a Benjamini-Hochberg corrected p-value threshold of 0.01 (http://genome.crg.es/GOToolBox).
Dissected postnatal intestines were fixed in 4% PFA overnight at 4°C and embedded in paraffin. Sections (7 μm) were stained by immunohistochemistry according to standard techniques. Briefly, the paraffin was removed from slide mounted sections by washing with xylene and ethanol, and sections were treated for antigen retrieval with boiling sodium citrate pH 6.0 for 10 min. Slides were blocked for 1 h in blocking solution (1% bovine serum albumin, 1% milk, in Tris-buffered saline pH 7.5), then incubated overnight at 4°C with mouse anti-βIII-Tubulin (Abcam ab78078, 1:200) primary antibody. Following several Tris-buffered saline (TBS) pH7.5 washes, sections were incubated for 2h at room temperature (RT) in alkaline phosphate-conjugated anti-mouse secondary antibody (Abcam ab97043, 1:200). After two washes in TBS pH 7.5 and one at pH 9.5, the staining was revealed with a nitro blue tetrazolium (NBT, 500μg/ml) and 5-bromo-4-chloro-3-indolyl-phosphate (BCIP, 187.5μg/ml) solution (Roche Applied Science) for 10 to 20 minutes. The reaction was stopped with a solution of TBS pH7.5 containing EDTA (20mM) and the slides were mounted with glycerol mounting medium (DAKO).
For embryonic tissues, freshly dissected intestines were fixed 1 hour at RT with 4% PFA in PBS. Alternatively, whole embryos were fixed overnight at 4°C and the intestines dissected afterwards. For adult tissues, whole intestines were fixed in 4% PFA overnight at 4°C, cut longitudinally along the mesentery and washed in PBS. The outermost muscle layers were then stripped from the mucosa/submucosa. Fixed tissues were dehydrated in methanol. After rehydration, the intestines were incubated 2 hours at RT in blocking solution (10% fetal bovine serum, 0.1% Triton-X100 in PBS). Tissues were incubated with primary antibodies overnight at 4°C. The antibodies used were: mouse anti-βIII-Tubulin (1:200; Abcam ab78078), rabbit anti-S100β (1:500; Dako Z0311), goat anti-Sox10 (1:100, Santa Cruz Biotech. sc-17342) and rabbit anti-Ki67 (1:1000; Abcam ab15580). Secondary antibodies Alexa Fluor 594- or Alexa Fluor 647-conjugated anti-goat, -mouse or -rabbit (1:500, Jackson Immunoresearch) were incubated for 2 hours at RT and counterstained with DAPI. All antibodies were diluted with blocking solution. For the TUNEL assay, tissues were permeabilized 20 min at 37°C in 0.3% Triton-X100, 0.1% sodium citrate in PBS 1x, then stained in a 1:9 mix of enzyme solution:label solution from the in situ cell death detection kit, TMR red (Roche Applied Science 12156792910) 1h at 37°C.
The whole colon (from cecum to anus) was dissected from adults and neonates. It was cut longitudinally along the mesentary (adult tissues only), rinsed, pinned flat and fixed in 4% PFA O/N at 4°C. Staining was performed on tissues as previously described [59]. Following the staining procedure, muscle strips were prepared as described above (adult tissues only).
Live imaging of eNCC was performed using a suspended culture technique adapted from Nishiyama et al., 2012 [3]. The abdomen of TashT e11.0 embryos was opened and the surrounding tissues trimmed just enough to expose the developing intestine. The embryo was placed on a small nitrocellulose filter (Millipore GSWP01300) soaked in PBS, and the extra tissues surrounding the intestine were slightly pressed onto the membrane. The PBS was then blotted off before being replaced with DMEM/F12 media (containing 10% FBS and antibiotics). The filter was flipped on top of a DMEM/F12-filled 2 mm-wide trough in a 1% agarose film covering the round glass bottom of a 35 mm culture dish (Greiner Bio One 627860), so that the intestine would float in media without touching either the agarose or the glass. The dish was incubated at 37°C, 5% CO2 during 6 hours, while 250μm-thick stacks were acquired with a 10x objective and a Nikon A1R confocal unit. Cell morphology viewed by YFP fluorescence allowed us to label the center of the cell located at the tip of a chain of migrating eNCC at each timeframe. Five to 6 chain tip cells were tracked from at least 3 intestines of each genotype. Speed and directionality were calculated from this dataset, with the orientation of the mesentery as a reference angle. We cannot totally exclude the possibility that more than one cell was included per chain tip measurement as individual eNCC at the wavefront sometimes exchange places with one another by a leapfrogging process [38].
Ex vivo cell migration assays were performed as previously described [23]. Collagen gels containing or not GDNF (10 ng/ml, Cedarlane CLCYT305–2) and/or EDN3 (250 ng/ml, Sigma E9137) were prepared at least 1 hour before use and kept at 37°C in a CO2 (5%) incubator. Vibratome transverse sections (200 μm) of embryonic small intestines were put down on the collagen gels for 3 days at 37°C with 5% CO2 to let cells migrate into the gel. Bowel sections were removed and cells within the gels were fixed with 4% PFA for 1 hour at RT before being stained with DAPI (Sigma-Aldrich) to detect cell nuclei. All images were taken at 70X magnification with a Leica M205FA fluorescence stereomicroscope as described below.
YFP expression in TashT tissues or in migrating cells was visualized using a Leica M205FA fluorescence stereomicroscope. IHC slides were observed using a DM2000 Leica upright microscope. Images were acquired with a Leica DFC495 digital camera and Leica Application Suite (LAS) software (Leica microsystems). IF of stained intestines were examined using an inverted Nikon TI microscope. Images were acquired with a Nikon A1 confocal unit and NIS-Element AR4 software, using standard excitation and emission filters for visualizing DAPI, YFP, Alexa Fluor 594 and 647, as well as spectral imaging coupled with linear unmixing in order to distinguish between YFP and GFP fluorescence. All images were processed with ImageJ software [60]. Image J was also used for cell counting with the analyze particles function, or with the cell counter manual function.
Constrained elements around the TashT insertion site were identified using Ensembl’s 35 eutherian mammals multiple alignment track (EPO_LOW_COVERAGE) on the NCBIm37 mouse genome version (www.ensembl.org). Seven regions named CE1 to CE7 (ranging between 600 to 1400bp) and containing one or multiple constrained elements were amplified by PCR (oligo sequences available upon request), cloned in the pGEM-T vector (Promega) and validated by sequencing. Luciferase reporter constructs were generated by subcloning each PCR fragment, in both the sense and antisense orientation, into a modified pXP2 vector containing the 109bp Thymidine kinase minimal promoter [61]. Neuro-2a neuroblastoma cells were propagated in EMEM supplemented with 10% FBS whereas P19 embryocarcinoma cells were propagated in alpha-MEM supplemented with 2.5% FBS and 7.5% CBS. Cos7 and NIH 3T3 cells were propagated in DMEM supplemented with 10% FBS. Transfections in 24-well plates and luciferase assays were performed in triplicate at least three times as previously described [62].
Western blots using whole cell extracts of transfected Cos7 cells were performed as previously described [63]. Cells were transfected with a CMVp-driven expression vector for the same Fam162b-IRES-eGFP bicistronic cassette used to produce transgenic embryos and protein expression was assessed using the following primary antibodies: rabbit anti-Fam162b (1:1000; Abcam ab122309), rabbit anti-GFP (1:5000; Abcam ab290) and rabbit anti-Gapdh (1:2500, Santa Cruz Biotech. sc-25778).
Total RNA was extracted using the RNAeasy Plus purification mini kit (QIAGEN) on FACS-sorted eNCC. The OneStep RT-PCR kit (QIAGEN) was then used on 100 ng of RNA with primers specific to the desired target (sequences available upon request). PCR consisted of 25–30–35 or 30–35–40 cycles of: 20 seconds at 95°C, 30 seconds at 62°C and 30 seconds at 72°C. Amplicons were resolved on a 2% agarose gel and quantified using the densitometry tools of ImageJ. The expression level of the housekeeping gene Gapdh was used for normalization.
3C was performed as previously described [27] with minor modifications. Starting material was either 1x108 cells (for Neuro-2a and P19 cell lines), as recommended, or ~1x106 cells (for e12.5 intestinal material), in which case reaction volumes were divided by a factor of 20. Only one cycle of phenol, then phenol/chloroform extraction was performed. Glycogen (0.05 mg/ml) was added as a co-precipitant prior to ethanol precipitation. For each PCR reaction, 20 ng (BAC library), 50 ng (embryonic intestines libraries) or 100 ng (cell lines libraries) of library DNA was used. Sequencing of the amplicon confirmed the identity of the chimeric fragment amplified.
Data are presented as mean ± standard deviation, with the number of experiments (n) included in the figure and/or legend. Quantification data were subjected to Student's t-test for statistical significance, except for directionality circular data which were compared through an ANOVA. Differences were considered statistically significant when the p value was less than 0.05.
|
10.1371/journal.pgen.1007581 | Cooperation, cis-interactions, versatility and evolutionary plasticity of multiple cis-acting elements underlie krox20 hindbrain regulation | Cis-regulation plays an essential role in the control of gene expression, and is particularly complex and poorly understood for developmental genes, which are subject to multiple levels of modulation. In this study, we performed a global analysis of the cis-acting elements involved in the control of the zebrafish developmental gene krox20. krox20 encodes a transcription factor required for hindbrain segmentation and patterning, a morphogenetic process highly conserved during vertebrate evolution. Chromatin accessibility analysis reveals a cis-regulatory landscape that includes 6 elements participating in the control of initiation and autoregulatory aspects of krox20 hindbrain expression. Combining transgenic reporter analyses and CRISPR/Cas9-mediated mutagenesis, we assign precise functions to each of these 6 elements and provide a comprehensive view of krox20 cis-regulation. Three important features emerged. First, cooperation between multiple cis-elements plays a major role in the regulation. Cooperation can surprisingly combine synergy and redundancy, and is not restricted to transcriptional enhancer activity (for example, 4 distinct elements cooperate through different modes to maintain autoregulation). Second, several elements are unexpectedly versatile, which allows them to be involved in different aspects of control of gene expression. Third, comparative analysis of the elements and their activities in several vertebrate species reveals that this versatility is underlain by major plasticity across evolution, despite the high conservation of the gene expression pattern. These characteristics are likely to be of broad significance for developmental genes.
| Animal development relies on the early delimitation of specific embryonic territories that will later participate in the formation of tissues and organs. This process is governed by sets of so-called developmental genes. The activities of the genes are themselves controlled by associated DNA sequences called enhancers. In vertebrates a part of the embryonic brain is delimited by the activity of the gene krox20. In this study, we have performed a comprehensive analysis of the krox20 regulatory landscape in the zebrafish vertebrate model. We show that 6 enhancers cooperate according to different modes to establish the complete pattern of krox20 activity. Furthermore, these enhancers appear unexpectedly versatile, combining different types of activities. This versatility is underlain by major plasticity across vertebrate evolution, despite the high conservation of the delimitation process. These observations are likely to be of broad significance for developmental genes.
| Enhancers are short, cis-acting regulatory elements that modulate transcription of target genes, relatively independently of their orientation or distance with respect to the promoter. They act as platforms to recruit multiple transcription factors [1] that interact with the transcription machinery at the promoter via cofactors [2]. A single gene can be controlled by multiple enhancers that show different activity profiles, providing both diversity and specificity of expression [3], or redundant profiles that may be required to ensure transcriptional robustness [4]. Interactions between enhancers can occur through different modes of cooperation: additive, synergistic, repressive, hierarchical or competitive [5]. Multiplicity of enhancers is a common feature among developmental genes [6] and is likely to play a major role in the evolution of gene expression, as it provides the necessary flexibility for pattern evolution [7].
For many years, the functions of enhancers have been mainly investigated through analysis of transgenic constructs carrying a reporter gene driven by a minimal promoter and linked to the enhancer [8]. Although fruitful, this approach is based upon the assumption that enhancer function can be recapitulated by the activity profile deduced from such an assay. However, it has not been established that this is always the case. In recent years, the advent of easy and efficient genome editing techniques, in particular those based on the CRISPR/Cas9 system, have facilitated mutation of putative enhancers in their natural genomic context [9,10], enabling the direct dissection of enhancer function in various species, including vertebrates [11].
Hindbrain segmentation is a highly conserved morphogenetic process in vertebrate development [12]. Among the regulatory genes involved in segmentation, Krox20 (also known as Egr2) plays a particularly important role. It encodes a zinc-finger transcription factor and is specifically and precisely expressed in two developing hindbrain segments, rhombomeres (r) 3 and 5 [13–15]. Krox20 is responsible for the formation and specification of these rhombomeres [16–19]. The regulation of Krox20 expression in the developing hindbrain provides an attractive model to study the functions and evolution of cis-acting elements involved in control of patterning in vertebrates. Three evolutionarily conserved enhancer elements active in the hindbrain have previously been identified near the Krox20 gene, termed A, B and C [20]. Analysis of chicken element A revealed that it is active in r3 and r5 and requires Krox20 binding for this activity [20], suggesting that it acts as an autoregulatory element. Indeed, deletion of element A in the mouse leads to a complete loss of Krox20 expression at late stages without affecting early stages, a phenotype very similar to Krox20 loss-of-function [21]. In contrast, chicken element B enhancer activity is Krox20-independent, and is restricted to r5 [20,22], making it a prime candidate for the initiation of Krox20 expression in r5. Finally, chicken enhancer C is active in the r3-r5 region, also in a Krox20-independent manner, suggesting that it might contribute to the initiation of Krox20 expression in r3 and/or r5 [20,23]. Surprisingly, deletion of element C in the mouse does not affect Krox20 expression at early stages, but leads to a loss of maintenance of Krox20 at late stages in r3 [24]. This loss of maintenance is due to cooperation in cis between element A and C, leading to increased accessibility of element A and potentiation of its autoregulatory activity in r3 [24]. This unexpected function of element C, unlike a classical enhancer, clearly illustrates the necessity of mutating putative cis-regulatory elements in their chromosomal context to decipher their true function.
In spite of these observations, previous analyses do not provide a complete global picture of Krox20 regulation in the hindbrain. In particular, they do not explain the basis for early Krox20 expression in r3. We therefore decided to engage in a systematic search and analysis of Krox20 cis-regulatory elements. For this purpose, we turned to the zebrafish, which allows easier identification and functional characterisation of regulatory elements and evolutionary comparisons with existing data from other vertebrates.
This approach has revealed a complex cis-regulatory landscape, with 6 elements controlling zebrafish krox20 expression in the hindbrain. Three of these are the homologues of the previously identified mouse and chicken elements A, B and C. Combining transgenic reporter analyses and CRISPR/Cas9-mediated mutagenesis in the chromosomal context, we assign precise functions to each of these 6 elements and provide a comprehensive view of krox20 cis-regulation in the hindbrain. Three important features of gene regulation emerge. First, cooperation and redundancy between multiple cis-elements play a major role in regulation (for instance, 4 elements cooperate to maintain autoregulation). Second, unexpected versatility of several elements allows them to be involved in different aspects of expression control. Third, this versatility is underlain by major plasticity across vertebrate evolution, despite the highly conserved pattern of Krox20 expression. These characteristics of Krox20 cis-regulation are likely shared by other developmental genes and are therefore of broad significance.
To study krox20 cis-regulation in detail in the zebrafish hindbrain, we first analysed its expression pattern by in situ hybridization, to provide a reference for comparison with the activities of putative enhancers. As krox20 regulation in the hindbrain has been shown to involve a positive feedback loop [20], we examined both wild type embryos and those carrying a homozygous point mutation in the krox20 coding sequence that abolishes Krox20 function and thereby prevents autoregulation (krox20fh227 allele [21,25]). In agreement with previous studies [21], krox20 expression is dynamic between the 95% epiboly and 20-somite stages (20s). A positive feedback loop contributes to the amplification and maintenance of expression, as in absence of active protein, krox20 mRNA disappears from r3 between 5s and 10s, and from r5 between 10s and 15s (Fig 1A). In contrast, in the wild type, the mRNA is maintained in both rhombomeres beyond 20s. krox20 is also expressed in neural crest cells leaving the neural tube from the r5/r6 region (Fig 1A, arrowhead).
To identify the transcriptional enhancers responsible for krox20 expression in the hindbrain, we undertook a systematic approach based on the observation that active cis-regulatory sequences typically show greater DNA accessibility than other sequences. We assessed chromatin accessibility within the krox20 locus and its vicinity by ATAC-seq [26]. ATAC-seq was performed on either 95% epiboly whole embryos or on micro-dissected regions (whole hindbrain, including r3 and r5, or a posterior region devoid of krox20-expressing cells; Fig 2) from 5s and 15s embryos. These conditions correspond to key moments in krox20’s expression dynamics: at the very beginning of gene activation (95% epiboly), after activation with limited (5s hindbrain) or full (15s hindbrain) contributions of the autoregulatory loop, and in regions where the gene remains silent (posterior regions). The analysis revealed 7 major peaks that are present outside of the promoter and coding sequence when krox20 is active (Fig 2). As all 7 peaks were located in non-repetitive regions and additional enhancers may have been missed by ATAC-Seq, we extended our survey of functional enhancers to all non-repetitive intergenic regions. This led to the selection of 22 sequences (ranging from 720 to 1726 bp), 7 containing one of the identified accessibility peaks (Fig 2, blue boxes).
To evaluate transcriptional enhancer activities associated with the 22 selected sequences, each was cloned into the Zebrafish Enhancer Detection (ZED) plasmid [27], upstream of a GFP reporter gene driven by the gata2 minimal promoter. These constructs were co-injected with transposase mRNA into one-cell stage zebrafish embryos and GFP fluorescence was monitored. Among the 22 cloned sequences, 6 led to hindbrain-specific GFP expression (Fig 1B), suggesting that each harboured a transcriptional enhancer. These 6 sequences were named A to F according to their positions along the locus. Each sequence included one of the accessibility peaks, demonstrating that assessment of chromatin accessibility by ATAC-seq is a powerful approach to identify cis-regulatory elements (Fig 2). All of these peaks (with the exception of the one corresponding to element F) were reduced at 15s in the krox20-negative posterior region of the embryo and two of them (corresponding to elements D and E) were very small at 95% epiboly (Fig 2), suggesting that in most of these regions, DNA accessibility is correlated with gene activity. In silico analysis of the 7th accessible region, located close to the promoter, revealed a putative binding site for the architectural protein CTCF [28] that may participate in increasing chromatin accessibility.
Elements A to E are located upstream of krox20, whereas element F is located downstream. Elements A, B, C and F show sequence similarity with the previously identified mouse and chicken hindbrain enhancers A, B and C [20] and the mouse NE element [24], respectively, and occupy the same relative positions along the locus (Fig 2). Sequence conservation between species is relatively high for elements B and F, reduced for element C, and low for element A (Figs 2 and S1). Sequences weakly homologous to element E were also identified in the vicinity of the mouse and chicken krox20 gene, again at the same relative positions (Figs 2 and S1). No sequences homologous to element D were detected in the mouse or chick (Fig 2).
To further investigate the activity of the 6 zebrafish elements, embryos injected with each construct were used to generate stable transgenic lines, whose profiles of GFP expression during hindbrain development were established by in situ hybridization (Fig 1B). At least two independent lines were analysed for each element, with the exception of element F, for which only one line was obtained. The patterns of GFP expression were identical for the different lines corresponding to the same element. We found that element A is weakly active in r3 between 3s and 10s, and much stronger in r5 from 3s to beyond 20s. Element B is active only in r5 between 3s and approximately 10s. At 10s, element B also drives GFP expression in neural crest cells migrating posteriorly to r5 (Fig 1B). Element C activity, first observed in r3 at 3s, later extends into r4 at 5s and then into r5 at 10s, and vanishes thereafter. Elements D and E are both active in r3 and r5 between 3s and 20s; D is more efficient in r3 at early stages, whereas E shows more activity in r5 at late stages. Finally, element F activity is restricted to r3, with very early onset (95% epiboly) but rapid extinction (at around 10s).
This enhancer assay suggests that among the 22 non-repetitive intergenic sequences located within and around the krox20 locus, 6 are likely to have hindbrain enhancer activities that reflect aspects of the normal hindbrain expression of the gene. This conclusion is further supported in that 3 of these elements, A, B and C, appear to show both structural and functional homology to previously characterised mouse and chicken enhancers [20]. Indeed, the patterns of activity of the homologous elements in the three species are very similar: B is restricted to r5, A is active in both r3 and r5, and C is active in a domain extending from r3 to r5. Sequences homologous to elements E and F also occur in the chick and mouse genomes, at the same relative positions as in the zebrafish. Together, the 6 zebrafish cis-acting elements appear to recapitulate all aspects of krox20 expression, in particular early activity in r3 and r5 for F and B, respectively, and intermediate or late activities in both r3 and r5 for all others. Finally, the fact that almost all major accessibility peaks identified by ATAC-seq correspond to hindbrain enhancers constitutes a strong validation of the use of this procedure to identify novel transcriptional cis-acting elements.
To determine the roles played in krox20 hindbrain regulation by the various cis-acting elements identified in the vicinity of the gene, we generated stable zebrafish lines with deletions of each element using CRISPR/Cas9 technology. Mutations were obtained by injecting into one-cell stage embryos the Cas9 protein together with two guide RNAs that targeted sequences flanking each element, resulting in its deletion. Stable lines were then selected; the deletions were characterised by PCR cloning and sequencing (S2 Fig) and the lines were used to obtain homozygous mutant embryos. The generation of stable lines carrying deletions of several elements was sometimes problematic. In such cases, we used an alternative approach that allowed us to obtain mutations in both alleles, directly in the injected embryo. Embryos were injected with the Cas9 protein, together with a mix of 3–4 guide RNAs that targeted evolutionarily conserved short sequences and/or putative binding sites for transcription factors, located within a 150–450 bp region presumably corresponding to the core enhancer (S2 Fig). This procedure was very efficient, allowing the introduction of deletions within both alleles at the same time, as demonstrated by the absence of fragments corresponding to the wild type allele following PCR amplification and further analysis of the DNA sequences (S3 Fig). Although the deletions introduced in both alleles might be different (S3B Fig), in all the cases analysed they led to complete or almost complete inactivation of the element, as judged by the homogeneity of the phenotypes associated with the mutations and their similarity to those corresponding to germ-line deletion of the same element (S4 Fig). Genotypes of mutated embryos generated through this approach (somatic deletion) are noted with the * symbol following the inactivated element.
To grossly map cis-acting elements governing krox20 in the hindbrain, we first generated a line carrying a deletion, ∆(A-E), that completely eliminated a 75 kb intergenic region between krox20 and nrbf2, including the 5 identified upstream elements, but excluding the krox20 promoter region (S2 Fig). The expression of krox20 in embryos carrying a homozygous ∆(A-E) deletion was dramatically affected: krox20 expression was initiated in r3, but krox20 mRNA levels rapidly decreased in this rhombomere and no expression was ever observed in r5 (Fig 3, ∆(A-E)). This result indicates that cis-acting sequences sufficient for initiation of krox20 expression in r3 are located outside of the deleted region. In contrast, cis-acting elements necessary for initiation in r5 and maintenance in r3 are located within this region.
There is an obvious candidate for governing krox20 initiation in r3: the downstream element F, which shows enhancer activity at early stages specifically in this rhombomere (Fig 1). Indeed, in homozygous mutants with a deletion of element F, krox20 expression was completely abolished in r3 at all stages, whereas r5 expression was unaffected (Fig 3, ∆F). Therefore, element F is absolutely required for initiation of krox20 expression in r3, consistent with its enhancer activity there at early stages (Fig 1B). In absence of any initiation, the feedback loop cannot be engaged and so no expression is observed at later stages either.
We next sought to identify the cis-acting sequences involved in the initiation of krox20 expression in r5 that are located within the ∆(A-E) deleted region. For this purpose, we generated zebrafish lines carrying deletions of each of the elements A, B, C, D or E. No phenotype was observed with any deletion in the heterozygous state. When affecting both alleles, two deletions, ∆A and ∆B, appeared to delay initiation of krox20 expression in r5 (Fig 3). In the case of ∆B, r5 expression was completely abolished at 3s and dramatically reduced at 5s, but at later stages, normal levels of expression were gradually reached (Fig 3). There was no effect on r3. Note that B is the only element whose deletion also obliterates krox20 expression in neural crest cells derived from the r5/r6 region (Fig 3, arrowhead). For ∆A, r5 expression was also affected at 3s and 5s, although less severely than in the ∆B mutant. However, ∆A also led to a slight reduction of expression in both r3 and r5 at later stages. The other deletions (∆C, ∆D and ∆E) did not affect krox20 expression in r5 at early stages (Fig 3).
To determine whether elements A and B are the only contributors to the initiation of krox20 expression in r5, we examined the effect of deleting both, by introducing a deletion of B (∆B’) in a ∆A background (S2 Fig). Embryos carrying homozygous deletions of both elements (∆A ∆B’) show a stronger phenotype than embryos with a single mutation: expression in r5 is only detected after 5s and late expression is also severely affected, presumably because the feedback loop cannot be appropriately established due to late and very poor initiation (Fig 4). As there was still limited expression maintained in r5 in the double mutant, we wondered whether a third element might be involved in the initiation step. Three elements show enhancer activity in r5: C, D and E (Fig 1). However, for elements D and E, this activity appears to be totally dependent on the presence of functional Krox20 protein (Fig 5A). This is not the case for element C, raising the possibility that it could cooperate with elements A and B to initiate krox20 in r5. We therefore combined deletions in C with deletions of A and/or B and examined whether any expression remained. The combination of homozygous B and C deletions did not increase the severity of the phenotype associated with B deletions (Fig 4). However, the combination of homozygous A, B and C deletions led to an almost complete loss of expression in r5: only a very low level of mRNA was reproducibly observed at 12s (Fig 4).
In conclusion, this analysis identified the cis-acting elements involved in the initiation phase of krox20 expression [21]: their homozygous mutation affects the hindbrain expression of krox20 at very early stages (at around 3s), before any significant involvement of the autoregulatory loop (Fig 1A). In r3, a single element, F, is absolutely required. In r5, however, the situation is more complex and elements show partial redundancy. Although element B appears as the major contributor, elements A and C are also involved and the mutation of all three elements is required to essentially abolish krox20 r5 expression. Residual expression could be due to very weak activity of a non-characterised fourth element or to the fact that the internal mutations in enhancers A and C do not totally inactivate them (Figs 4 and S2). Finally, among the identified elements, in the neural crest derived from r5/r6, enhancer B is the only one required for krox20 expression.
Three of the krox20 cis-acting elements, A, D and E, appear to share similar characteristics: they act as enhancers in both r3 and r5, and are active at late stages (up to 20s). Furthermore, deletions ∆A and ∆E lead to a slight decrease in krox20 mRNA levels in both r3 and r5 after 5s (Fig 3). These features suggest that they are involved in the maintenance of krox20 expression and possibly in the underlying positive feedback loop [21]. In addition, the chick and mouse orthologues of element A contain Krox20 binding sites that are required for enhancer activity [20,21], and mouse element A is absolutely necessary for krox20 autoregulation [21]. To investigate whether zebrafish elements A, D and E could be involved in direct krox20 autoregulation, we first examined the activity of these elements in the absence of the Krox20 protein. As indicated above, without Krox20, the enhancer activities of elements D and E were completely abrogated in both r3 and r5 (Fig 5A), demonstrating that these elements are Krox20-dependent and are likely to be involved in the feedback loop. In the case of element A, in the absence of Krox20 protein, r3 enhancer activity was completely eliminated, but some r5 activity was maintained, although severely reduced (Fig 5A). These data indicate that element A possesses a dual function: Krox20-dependent enhancer activities in both r3 and r5 and a Krox20-independent enhancer activity specifically in r5. This latter activity is likely to contribute, together with elements B and C, to the initiation of krox20 expression in r5 (Fig 4).
To determine whether the Krox20-dependent activities of elements A, D and E might involve direct binding of the Krox20 protein, we looked for potential binding sites for Krox20 within the enhancer sequences. For each, we identified several putative binding sites (S1 and S2 Figs). Oligonucleotides corresponding to sequences from each enhancer and carrying two of these binding sites were synthesized and used to perform gel retardation experiments in the presence of the Krox20 protein, together with specific or non-specific competitors. In each case, there was at least one strong retarded band, corresponding to a specific complex with Krox20 (S5 Fig), indicating that these elements contain high affinity Krox20 binding sites and supporting the idea that their enhancer activity is dependent on direct binding of Krox20.
As the phenotypes associated with the single homozygous mutation of elements A, D or E are limited, it is likely that these elements cooperate to establish full autoregulation. We tested this hypothesis by combining the different mutations. Indeed, the combination of two homozygous deletions, affecting A and D, A and E, or D and E severely reduced krox20 expression at 12s and 22s (Fig 5B). Furthermore, elimination of the three enhancers, either by introduction of deletions affecting each one (Fig 5B) or by combination of a deletion of element A with a deletion of the D-E region (S2 and S4 Figs) led to complete loss of krox20 expression at 12s and 22s. Note that neural crest expression at 12s, which relies on element B, is maintained in all cases (Fig 5B).
In conclusion, our data establish that elements A, D and E all carry Krox20-dependent enhancer activities. Furthermore, these elements cooperate to generate the positive feedback loop that maintains late expression of krox20. Finally, these activities are likely to involve direct binding of the Krox20 protein to each enhancer.
On the basis of the above analysis, element C appears somehow peculiar. Like elements A, D and E, it shows enhancer activity in r3 and r5, but this activity is Krox20-independent (Figs 1B and 5A). Furthermore, its activity is not restricted to r3 and r5, but also covers r4, with a dynamic anterior-posterior pattern (Fig 1B). Deletion experiments have shown that element C is a minor contributor to initiation of krox20 expression in r5 (Figs 3 and 4). It is also involved in late krox20 expression, as its deletion leads to a slight decrease in krox20 mRNA levels in both r3 and r5 after 5s (Fig 3), although this is not likely to occur via direct autoregulation (Fig 5A). To determine whether element C interacts with other elements at late stages, we combined its deletion with mutations in A, D and E. Inactivation of element C did not exacerbate the late phenotype associated with the elimination of element A (Fig 6, compare ∆A and (∆A C*)). In contrast, the phenotype was more severe when mutation of element C was combined with mutations of elements D and E (Fig 6, compare (D* E*) and (∆C D* E*)). In fact, this latter genotype leads to a phenotype similar to that of (∆A D* E*), although slightly less severe in r5 (Fig 6), probably due to the more significant involvement of element A in the initiation of krox20 in r5, as compared to element C. Together, these data are consistent with element C contributing to autoregulation by modulating the activity of element A.
Similar cooperation was previously observed in the mouse, where the orthologue of element C, although not directly participating in the positive feedback loop, cooperates in cis with element A to potentiate its autoregulatory activity [24]. To investigate whether such a cis-cooperation exists between A and C in zebrafish, we generated embryos homozygous for D and E mutations and heterozygous for A and/or C deletions (Fig 6). The latter were introduced by crossing ∆A and ∆C homozygous lines and were therefore present on different chromosomes. When both heterozygous deletions for A and C were present (∆A/+ +/∆C D* E*), krox20 expression at late stages was affected in a manner similar to the combination (∆A D* E*), where the deletion of element A is homozygous. In contrast, when only the heterozygous deletion of C was introduced in the (D* E*) background (∆C/+ D* E*), it did not significantly increase the severity of the (D* E*) phenotype (Fig 6). These results support the existence of a cis interaction between C and A, required to allow A to participate in the autoregulatory loop.
Together, these data indicate that element A does not take part in autoregulation when a functional element C is not present on the same chromosome. Therefore, element C cooperates with element A to potentiate its autoregulatory activity, just as in the mouse. However, in the zebrafish, two additional cis-regulatory elements, D and E, directly participate in the feedback loop. In contrast to element A, element D and E are not likely to depend on element C to exert their enhancer activities.
Zebrafish element A acts both as a Krox20-independent initiator element in r5 and as an autoregulatory element in r3 and r5. The existence of these dual activities is surprising in view of what we know of its chicken and mouse orthologues. Chicken element A is totally dependent on Krox20 binding for its enhancer activity, as demonstrated by comparison of a reporter transgene in mouse Krox20 null and wild type backgrounds, and by mutation of element A Krox20 binding sites with enhancer activity assessed in chick embryos [20]. In addition, while deletion of mouse element A completely abolishes the positive feedback loop, it has no effect on early expression in r5 in this species [21]. Therefore, element A does not appear to act as an initiator element in chick nor mouse, suggesting its enhancer activity has been modified during vertebrate evolution. To investigate whether a coherent pattern of evolution of the element might be identified, we analysed the activities of orthologues of element A from several key species in the vertebrate phylogenetic tree (Fig 7). We cloned the orthologues of zebrafish element A (zA) identified by sequence alignments from koi carp Cyprinus rubrofuscus (kA), spotted gar (sA), Xenopus tropicalis (xA), chicken (cA) and mouse (mA) into the ZED GFP expression vector, generated stable zebrafish transgenic lines (at least two independent ones for each species) and determined the patterns of GFP expression by in situ hybridization. In a wild type zebrafish background, despite the heterospecific character of the assay, all elements behaved similarly and could direct GFP expression in r3 and r5, although there were some relative variations in the expression level between the two rhombomeres (Fig 7). To determine whether any of these enhancer activities were dependent on the Krox20 protein, we injected transgenic embryos from each line with the Cas9 protein and guide RNAs targeting the sequences encoding the three zinc fingers of the Krox20 protein, which constitute the DNA binding domain (S2 Fig). This treatment effectively abolishes krox20 expression at 12s (S6 Fig), and allows the assessment of Krox20-independent enhancer activity. A large proportion of the activities of the enhancers was Krox20-dependent (Fig 7). However, limited Krox20-independent activities were maintained in some cases. Surprisingly, their patterns appeared different from one species to another and incoherent with the phylogenetic tree: the zebrafish and spotted gar elements remained active in r5 only, whereas the koi carp element was only active in r3; the mouse element was weakly active in both r3 and r5, whereas no activity was detected with the Xenopus and chick elements (Fig 7).
In conclusion, this analysis shows that the features required for Krox20-dependent expression of element A are likely to have been largely conserved during the course of vertebrate evolution. In contrast, the capacity of this element to combine its autoregulatory activity with Krox20-independent initiator functions appears highly contingent, with no clear correlation with the course of evolution. Furthermore, this Krox20-independent activity can occur in r3, in r5 or in r3 and r5, revealing a surprising plasticity of element A for acquiring and losing additional functions during evolution.
In this work, we have performed a comprehensive functional analysis of the cis-regulatory landscape of an important developmental gene, krox20. In the zebrafish, the organisation appears highly complex, since no less than 6 cis-acting elements are required to control the expression of the gene in two rhombomeres. These elements can account for all aspects of krox20 expression in the developing hindbrain, allowing us to propose a global view of its regulation (Fig 8). As previously observed in the mouse, the enhancer activities of these elements can be classified as Krox20-independent or -dependent, the latter underlying the positive feedback loop that ensures amplification and maintenance of krox20 expression at late stages. Apart from the initiation of krox20 expression in r3, the other aspects of the regulation of the gene are controlled by multiple elements: initiation of krox20 expression in r5 is governed by 3 elements (A, B and C), whereas autoregulation is controlled by 4 elements (A, C, D and E). These elements appear to cooperate according to various modes, possibly involving cis-interactions. Surprisingly, two elements (A and C) appear to participate in both regulatory aspects, revealing an intriguing interplay that might originate from the sharing of some binding sites for transcription factors involved in both activities. Finally, comparisons among vertebrates have revealed that the krox20 cis-regulatory landscape is unexpectedly poorly conserved and that particular elements show a remarkable evolutionary plasticity.
Several studies, mostly performed in Drosophila, have recently shown that cooperation between cis-regulatory elements is a common feature in the regulation of developmental genes and can occur according to different modes, including in particular additive, synergistic or hierarchical interactions [3,5,24,29,30]. The present study provides examples of such co-operations in vertebrates, in the initiation of krox20 expression in r5 as well as in the positive feedback loop (Fig 8). Although we have not performed quantitative analyses of the contributions of each cis-acting element to the different aspects of krox20 hindbrain expression, in the case of initiation in r5, this cooperation appears to occur through an additive mode: deletion of each element leads to reduced expression (with B>A), and a drastic decrease requires combination of both deletions. The third element, C, appears only as a minor contributor to this activity. More generally, the transcriptional activity of each of these r5 initiating elements shows specificities (Fig 1B) that may reflect differences in which transcription factors act on them [20,22,23,31].
Considering autoregulation, three elements (A, D and E) can directly bind Krox20 protein (S5 Fig). Elimination of each one alone leads only to a mild phenotype (with E>A>D, Fig 3). However, combined knockdown of any two elements results in a major decrease in late expression (Fig 5B), suggesting the existence of a strong synergistic component in this co-operation. Therefore, in this case, synergy and redundancy are not exclusive, as an almost full activity is already reached with two elements.
Redundancy in the cis-acting elements controlling zebrafish krox20 autoregulation differs remarkably from the situation in the mouse, in which element A is absolutely required for late Krox20 expression [21]. While we were not able to detect sequences homologous to element D in the mouse Krox20 locus, there is a poorly conserved mouse orthologue of element E, although, it cannot rescue the deletion of element A in this species. Overlapping activities between regulatory elements add robustness to the expression of developmental genes [3,32]. We speculate that the difference in redundancy in the control of the krox20 feedback loop between zebrafish and mouse might reflect differences in both external and internal conditions that require additional robustness in the zebrafish. For example, the zebrafish embryo is much more sensitive to modifications in environmental conditions such as temperature or mechanical stress. Further, the process of hindbrain segmentation takes only 12 hours in the zebrafish compared to 36 hours in the mouse, giving the zebrafish much less time to ensure full establishment of krox20 autoregulation, a crucial step in building normal size r3 and r5 [21].
The additional involvement of element C in autoregulation is peculiar, as it seems to operate in a cis-acting, hierarchical manner with element A. In contrast, element C does not potentiate the autoregulatory activities of elements D and E, since in the absence of element A, elimination of element C does not affect autoregulation. It is possible that this independence of elements D and E from C might be related to the organisation of the locus itself, given that elements D and E are much closer (-15 kb, -12kb) to the promoter than element A (-74 kb), with element C being positioned in between (-41 kb). Interestingly, the potentiation by C is only required for the autoregulatory activity of element A, but not for its initiator activity in r5. Therefore, if element C is required for chromatin opening at element A as proposed in the mouse [24], the constraints on chromatin structure for activation by the Krox20 protein are likely to be different from those required by the initiation factors.
It is worth noting that element F, which is in charge of the initial activation of krox20 in r3, the earliest manifestation of krox20 expression in the embryo, is the only hindbrain regulatory element to be located downstream of the gene, whereas the elements responsible for initiation in r5 and autoregulation are all located upstream. We speculate that this spatial organisation might reflect the existence of two mutually exclusive DNA loops, as observed for the regulation of the HoxD cluster during vertebrate limb development [33]. Early in r3, a DNA loop might form, including the krox20 promoter and the downstream region containing element F. Later, the promoter might engage into an alternative loop including all upstream elements, allowing initiation of krox20 expression in r5, as well as establishment of autoregulation in both rhombomeres. This dynamic spatial organisation is consistent with the very early activation of element F (Fig 1) and premature downregulation of krox20 in r3 in the ∆(A-E) mutant (Fig 3). Later, this organisation would allow parallel activation of elements involved in initiation in r5 (A, B and C) and autoregulation (A, C, D, and E), with two elements (A and C) participating in both processes.
It is surprising that some cis-acting regulatory elements have the capacity for different activities that might have been expected to be carried out by distinct elements. We found two examples of such versatility. The first is zebrafish element A, which possesses two distinct types of enhancer activities: a Krox20-independent initiator activity in r5 and a Krox20-dependent autoregulatory activity in both r3 and r5. The second case, element C, is even more striking. This sequence appears to carry enhancer activity in r3-r5 when assayed in the transgenic reporter system and it contributes in vivo to the initiation of krox20 expression in r5, presumably via this classical enhancer activity. In addition, element C appears to also function through cooperation in cis with element A, to potentiate its autoregulatory activity in r3 and r5. We have previously proposed, in the case of mouse element C, that such a potentiating activity, required for the function of a positive feedback loop, may constitute an efficient safety lock against inappropriate activation of autoregulatory elements [24].
At this stage, in the absence of analyses of the precise DNA sequences required for the activities carried by A or C, it is not known whether the dual functions are borne by distinct sequences or involve some common sequence motifs and interacting factors. In the former case, we would expect adjacent or intermingled cis-acting elements. The latter possibility is more interesting, as the sharing of some binding sites might result in common properties, like temporal and/or regional domains of activity. Hence, element C enhancer and potentiator activities overlap in r3 and r5 between 3s and 10s. Preliminary efforts designed to separate initiating and autoregulatory activities of element A by external deletions have failed, both activities decreasing in parallel. Finally, composite organisation underlying different activities might facilitate the appearance and modifications of one activity by mutations, leading to increased potential for evolution.
Comparison of the krox20 cis-regulatory landscape between zebrafish and mouse revealed major differences in the number of elements, in their nucleotide sequences and in their functional activities (Figs 2 and 8). This is particularly surprising in view of the strong conservation of hindbrain segmentation and the krox20 expression pattern during vertebrate evolution, and given that modification of cis-regulatory sequences is considered a major driver of evolution in higher organisms [7]. Among the 6 cis-regulatory elements identified in zebrafish, only two are relatively strongly conserved among vertebrates–elements B and F (S1 Fig), which are the only ones involved exclusively in the initiation of krox20 expression (Fig 2). In contrast, among direct autoregulatory elements, elements A and E are poorly conserved between zebrafish and mouse (S1 Fig), and element D is not present in tetrapods, but instead detected cavefish. This correlation between initiation versus autoregulation and evolutionary conservation might be explained by the need for initiating elements to act as platforms integrating numerous signals mediated by a variety of transcription factors, to precisely define spatial and temporal domain of activity. This platform function might seriously constrain the evolution of enhancer sequences. In contrast, direct autoregulatory elements mainly need to bind the gene product, probably as well as factors that more loosely restrict the domain of autoregulation. This is likely to offer additional evolutionary plasticity. Multiplication of partially redundant elements, like in the case of autoregulation, also offers space for increased evolutionary flexibility. In contrast, elements F and B play unique or major roles in r3 and r5 initiation, respectively, and are therefore likely to be more constrained.
The search for putative binding sites for transcription factors likely to control the regulation of krox20 expression supports this interpretation. Hence, vHnf1 and MafB binding sites, and Hox/Pbx, Meis and Sp binding sites are well conserved between mouse and zebrafish elements B and C, respectively (S1 Fig). These different factors and their binding sites have been shown to play essential roles for the enhancer activities of the corresponding chick and mouse elements (Chomette et al., 2006; Wassef et al., 2008 and Labalette et al., 2015), suggesting similar functions in the zebrafish. Furthermore, like element C, element F contains conserved binding sites for Hox/Pbx, Meis and Sp factors, suggesting that the two elements bind overlapping subsets of transcription factors, and that these common factors might be essential for element F activity in r3. In this respect, it is worth noting that elimination of the Meis sites in chicken element C have been shown to affect its enhancer activity specifically in this rhombomere in transgenic mice (Wassef et al., 2008). Meis factors might therefore play a particularly important role for element F.
Another interesting evolutionary issue is the case of enhancers possessing dual activities. The zebrafish autoregulatory element A carries an additional, Krox20-independent activity in r5, in contrast to its chicken orthologue. The appearance of this additional activity might have been favoured by the redundancy in the elements governing autoregulation in the zebrafish. In any case, we explored the presence of dual activities in element A in several vertebrate species to determine whether this would correlate with the phylogenetic tree. However, the activity patterns were unexpectedly variable, with no correlation with evolution (Fig 7): the Krox20-independent activity, as tested in the zebrafish, can be restricted to r5 (zebrafish and spotted gar), to r3 (koi carp), present in r3 and r5 (mouse) or absent from the hindbrain (Xenopus, chicken). To determine whether the pattern of Krox20-independent activity of the elements A might correlate with the presence of binding sites for specific transcription factors, we searched for putative binding sites known to be involved in the r3- or r5-specific activities of elements B and C. Zebrafish element A contains several MafB sites and a single vHNF1 site. Whereas MafB putative sites were observed in the carp element, no vHNF1 site was found. As this binding site is essential for the r5-specific activity of element B, this might explain the absence of initiator activity of the koi carp element in r5. In contrast, concerning the r3-specific expression of the carp element, the distribution of putative Hox/Pbx and Meis sites does not provide any clue susceptible to explain the different behaviour of the two elements. In any case, our analysis suggests that element A shows a high potential and plasticity for developing initiation functions, possibly favouring adaptation to various embryonic environments. It will be interesting to determine whether this plasticity is linked to the dual nature of the element and whether this feature has a broad significance. Indeed, it has been proposed that evolution of novel patterns of gene expression relies on the introduction of mutations in pre-existing enhancers rather than on the invention of new ones [34,35]. In this respect, element A might have been caught red-handed.
All animal experiments were performed in accordance with the guidelines of the Council of European Union Directive n°2010/63/UE and were approved by the “Comité d’éthique pour l’expérimentation animale Charles Darwin” (Project Number: APAFIS#848-2015061510065446v3).
The constructs used to generate transgenic zebrafish lines were based on the ZED plasmid [27] digested by BspM1+BspEI, in which each of the tested regulatory elements were cloned upstream of a GFP reporter gene driven by the gata2 minimal promoter, using Clontech’s “In-Fusion HD Cloning Kits” and following their protocol. The DNA primers used to amplify the regulatory elements were designed with the forward (5’-TGAATGCTCATCCGGA…-3’) and reverse (5’-GACCTGCAGACTGG…-3’) prefixes complementary to the ends of the linearized vector, which were followed by the specific sequences of the primers (S1 Table). Primers were synthesized and purified by Eurofins Genomics. Transgenic lines were obtained from embryos injected at the one-cell stage with 50 pg of Tol2 transposase mRNA [36] together with 75 ng of the ZED construct. At least two independent transgenic lines have been generated and analyzed for elements A, B, C, D and E and the elements A from various species. Single and double whole-mount in situ hybridizations were performed as described [37], with the previously published digoxigenin-labelled riboprobes for krox20 [38] and GFP [31].
Zebrafish lines with regulatory elements deleted, embryos harbouring somatic deletions of one or several enhancers and embryos with somatic deletions of the zinc finger domain of Egr2b were generated using the CRISPR/Cas9 editing system. The sequence-specific parts of the RNA guides (S1 Table) were designed with the help of the CRISPOR design tool (http://crispor.tefor.net/) to minimise off-targeting, maximise efficiency and specificity of targeting and, in the case of somatic deletions, to target putative binding sites or particularly conserved regions within enhancers (S1 and S2 Figs). Four guides were designed to target element A on two Krox20 binding sites and two conserved regions. Four guides were designed to target element C on two putative binding sites for Hox/Pbx factors, one for Meis and one for Sp. Four guides were designed to target element D on four binding sites for Krox20. Three guides were designed to target element E on three binding sites for Krox20. Three guides were designed to target element F on one putative binding site for Hox/Pbx factors, one for putative Meis binding site and one conserved regions. The targeting parts of the RNA guides were synthesized (Integrated DNA Technologies) 5’ to a 15-nucleotide sequence complementary to the tracrRNA 5’-GUUUUAGAGCUAUGCU-3’. This “crRNA” was then hybridized with the 67-nucleotide long “tracrRNA” to form the complete RNA guide. These complete guides (50 μM) were then incubated with the Cas9 protein (45 μM) (synthesized and generously provided by Anne de Cian, Muséum National d’Histoire Naturelle) in the Cas9 buffer (20 mM Hepes pH 7.5, 150mM KCl) and injected into one-cell stage zebrafish embryos. Founder zebrafish for knockout lines and whole injected F0 embryos harbouring somatic deletions were genotyped after the in situ hybridization analysis using PCR genotyping primers (S1 Table). Sanger sequencing was performed (Eurofins Genomics) to characterise the deletions obtained through non-homologous end joining.
ATAC experiments were performed according to Buenrostro and colleagues [26], using a homemade transposome [39]. All embryos were dissected in cold PBS to remove the vitellus (50 embryos at 95% epiboly) or to isolate the hindbrain and the posterior part of the embryo (80 embryos at 5s and 50 embryos at 15s) as shown in Fig 2. Biological duplicates were performed for each ATAC experiment. Briefly, cells were lysed before transposition using 1 μl of transposome and purified using a Qiagen MinElute Kit with 10 μl of Elution Buffer. Transposed DNA was amplified by PCR [39] and quantified by qPCR as previously described [24]. Sequencing was performed on multiplexed samples using 42 bp paired-end reads on an Illumina NextSeq according to the manufacturer's specifications. For computational analysis, paired-end reads were mapped onto the zebrafish genome assembly zv9, using STAR as previously detailed [24].
The mouse Krox20 protein was expressed in bacteria using the pet3a system. Extracts were prepared from Krox20-expressing and control bacteria as previously described [40]. Double-stranded biotinylated oligonucleotides with the following sequences were used as probes:
Gel shift experiments were performed with the light shift chemoluminescent EMSA kit (PIERCE), except for the composition of the binding buffer [40].
The data have been deposited in the Gene Expression Omnibus (GEO) under accession number GSE113471 and are available at the following address: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE113471
|
10.1371/journal.pntd.0003772 | Modeling the Geographic Spread of Rabies in China | In order to investigate how the movement of dogs affects the geographically inter-provincial spread of rabies in Mainland China, we propose a multi-patch model to describe the transmission dynamics of rabies between dogs and humans, in which each province is regarded as a patch. In each patch the submodel consists of susceptible, exposed, infectious, and vaccinated subpopulations of both dogs and humans and describes the spread of rabies among dogs and from infectious dogs to humans. The existence of the disease-free equilibrium is discussed, the basic reproduction number is calculated, and the effect of moving rates of dogs between patches on the basic reproduction number is studied. To investigate the rabies virus clades lineages, the two-patch submodel is used to simulate the human rabies data from Guizhou and Guangxi, Hebei and Fujian, and Sichuan and Shaanxi, respectively. It is found that the basic reproduction number of the two-patch model could be larger than one even if the isolated basic reproduction number of each patch is less than one. This indicates that the immigration of dogs may make the disease endemic even if the disease dies out in each isolated patch when there is no immigration. In order to reduce and prevent geographical spread of rabies in China, our results suggest that the management of dog markets and trades needs to be regulated, and transportation of dogs has to be better monitored and under constant surveillance.
| In 1999, human rabies cases were reported in about 120 counties in Mainland China, mainly in the southern provinces. Now outbreaks of human rabies have been reported in about 1000 counties and the disease has spread geographically from the south to the north. Phylogeographic analyses of rabies virus strains indicate that prevalent strains in northern provinces are indeed related to the remote southern provinces. It is believed that the geographical spread of rabies virus is caused by the transportation of dogs. In this paper, a multi-patch model is proposed to describe the spatial transmission dynamics of rabies in China and to investigate how the immigration of dogs affects the geographical spread of rabies. The expression and sensitivity analysis of the basic reproduction number indicates that the movement of dogs plays an essential role in the spatial transmission dynamics of rabies. Numerical simulations on the effect of the immigration rate in three pairs of provinces, Guizhou and Guangxi, Hebei and Fujian, Sichuan and Shaanxi, are also performed. It is shown that the immigration of dogs is the main factor for the long-distance inter-provincial spread of rabies and it is necessary to manage such inter-provincial transportation of dogs.
| Rabies, as an acute and fatal zoonotic disease, is most often transmitted through the bite or scratch of a rabid animal. The rabies virus infects the central nervous system, ultimately causing disease in the brain and death. Once the symptoms of rabies have developed, its mortality rate is almost 100%. Rabies causes tens of thousands of deaths worldwide per year ([1]), more than 95% of which occur in Asia and Africa. More human deaths from rabies occur in Asia than anywhere else in the world ([2]). It was first recorded in ancient China in about 556 BC ([3]) and nowadays it is still a very serious public-health problem in China. It has been classified as a class II infectious disease in the National Stationary Notifiable Communicable Diseases and the annual data of human rabies have been archived by the Chinese Center for Disease Control and Prevention since 1950. From 1950 to 2013, 128,769 human rabies cases were reported in China ([4–7]), an average of 2,012 cases per year. It is estimated that 85%–95% of human rabies cases are due to dog bites in mainland China ([5]).
Recently, there are some studies on modeling the transmission dynamics of rabies in mainland China. Zhang et al. [8] proposed a deterministic model to study the transmission dynamics of rabies in China. The model consists of susceptible, exposed, infectious, and vaccinated subpopulations of both dogs and humans and describes the spread of rabies among dogs and from infectious dogs to humans. The model simulations agree with the human rabies data reported by the Chinese Ministry of Health from 1996 to 2010. It was shown that reducing dog birth rate and increasing dog immunization coverage rate are the most effective methods for controlling rabies in China and large scale culling of susceptible dogs can be replaced by immunization of them. Based on the model of Zhang et al. [8], Hou et al. [9] considered a deterministic model for the dog-human transmission of rabies, taking into account both domestic and stray dogs, and used the model to simulate the reported human cases in Guangdong Province, China. It was shown that the quantity of stray dogs also plays an important role in the transmission of rabies. Based on the fact that the monthly rabies data in China exhibit periodic patterns, Zhang et al. [10] constructed a susceptible, exposed, infectious, and vaccinated (SEIVS) model with periodic transmission rates to investigate the seasonal rabies epidemics. They evaluated the basic reproduction number, analyzed the dynamical behavior of the model, used the model to simulate the monthly data of human rabies cases reported by the Chinese Ministry of Health from January 2004 to December 2010, and explored some effective control measures for the rabies epidemics in China.
In the last 20 years or so, rural communities and areas in Mainland China are invaded by rabies gradually. The range of infected hosts has expanded and the number of counties with reported human rabies increased significantly (See Fig 1). Moreover, human rabies has been expanded geographically from the south provinces to the central and north provinces (see [10]). Some provinces such as Shaanxi and Shanxi in the north, used to be rabies free, have reported more and more rabies cases in the past few years ([11]). Since the trade and transportation of dogs are regarded as the main cause for the spatial spread of rabies, Zhang et al. [10] extended their early ODE model to a reaction-diffusion model to study how the movement of dogs impacts the spatial spread of rabies. Their analysis indicates that the movement of dogs leads to the traveling wave of dog and human rabies and has a large influence on the minimal wave speed.
Although dogs remain the major infection source, contributing 85%–95% of human cases in China ([5]), there are very little scientific studies and very few data on the population dynamics of dogs, let alone diseases of dogs. In order to improve rabies control and prevention, in 2005 the Chinese government implemented a trial surveillance program to monitor rabies at the national level in an attempt to obtain a more comprehensive epidemiological dataset. In addition to recording statistics on human cases, the Institute for Viral Disease Control and Prevention of China CDC cooperated with the provincial CDC laboratories and began collecting samples from dog populations in regions where human rabies cases had been reported. The positive samples were then submitted for DNA sequencing and combined with a second subset of selected sequences from publicly available sequences. Yu et al. [12] selected a subset of samples for sequencing and investigated the history and origin of the virus in China and examined the variation from a geographical perspective. Guo et al. [13] used comprehensive spatial analysis methodology to describe the spatiotemporal variation of human rabies infections in China from 2005 to 2011, detected spatiotemporal clusters of human rabies, modeled the transmission trend of rabies, and provided a scientific basis for improved targeted human rabies control interventions in China. Guo et al. [14] collected rabies virus nucleoprotein gene sequences from different provinces and investigated their phylogenetic and phylogeographic relationship. More specifically, their phylogeographical analyses of two rabies virus clades (China I and China II) lineages identified several provinces that appear to be epidemiologically linked and China I lineage plays the dominant role in the spread of rabies in China. Moreover, their analysis indicates that east China appears to be not only epidemiologically related to adjoining provinces but also to distant provinces, and seems to act as an epidemic hub for transmission of rabies virus to other regions, which is consistent with previous results by Yu et al. [12]. Other long distance translocations of rabies virus can also be identified as well as translocation events between neighboring provinces. Their analysis demonstrates a strong epidemiological linkage between Shaanxi to Sichuan and between Sichuan to Yunnan. This is consistent with surveillance data for human rabies cases which show dissemination of the virus from southwest China to neighboring provinces and into regions such as Shaanxi in the northern part of the county that have previously been incident free for several years (Yin et al. [11]). For both clades there appears to be a general trend of longitudinal transmission (Guangdong-Shandong, Fujian-Hebei, Zhejiang-Shandong) and latitudinal transmission (Yunnan-Shanghai, Guizhou-Shanghai, Hunan-Shanghai). That is also consistent with human rabies surveillance data which highlights a flow of cases from high incidence regions in the south of the country to medium and low incidence regions (Yin et al. [11]). For example, discrete phylogeographic analysis for China I strain ([12, 14]) indicates the linkage of rabies virus between Sichuan and Shaanxi, Guangxi and Guizhou, and Fujian and Hebei (Fig 2).
Zhang et al. [15] used a reaction-diffusion model to study the spatial spread of rabies in China. However, reaction-diffusion equations are based on the mathematical assumptions that the spatial domain is connected and the movement of dogs is a continuous process in the domain. While the phylogeographical analyses of rabies virus indicate that there are long distance inter-provincial spread of rabies in China, in order to investigate how the movement of dogs affects the geographic spread, we propose a multi-patch model to study the spatial transmission of rabies between dogs and from dogs to humans. We will describe the model in details, discuss the existence of the disease-free equilibrium, calculate the basic reproduction number, and study how the moving rates between patches affect the basic reproduction number. To investigate the epidemiological linkage (such as Guizhou and Guangxi, Hebei and Fujian, and Sichuan and Shaanxi) observed in Guo et al. [14], we will use the two-patch submodel to simulate the human rabies data to understand the inter-provincial spread of rabies in China.
Since the data on human rabies in mainland China are reported to the China CDC by provinces, we regard each provinces as a single patch and, in each patch, the submodel structure follows the SEIR model proposed by Zhang et al. [8] (see Fig 3). We use superscripts H and D to represent human and dog, respectively, and a subscript i to denote the ith-patch. We assume there are n patches where n ≥ 2 ([16]). For patch i, the dog population is divided into four subclasses: S i D ( t ), E i D ( t ), I i D ( t ), and V i D ( t ) , which denote the populations of susceptible, exposed infectious and vaccinated dogs at time t, respectively. Similarly, the human population in patch i is classified into S i H ( t ), E i H ( t ), I i H ( t ), and V i H ( t ), which denote the populations of susceptible, exposed, infectious and vaccinated humans at time t, respectively. Our assumptions on the dynamical transmission of rabies between dogs and from dogs to humans are presented in the flowchart (Fig 3). The model in patch i is described by the following differential equations:
d S i D d t = A i + λ i D V i D + σ i D ( 1 - γ i D ) E i D - β i D S i D I i D - ( m i D + k i D ) S i D + ∑ j = 1 n ϕ i j S S j D , d E i D d t = β i D S i D I i D - ( m i D + σ i D + k i D ) E i D + ∑ j = 1 n ϕ i j E E j D , d I i D d t = σ i D γ i D E i D - ( m i D + μ i D ) I i D + ∑ j = 1 n ϕ i j I I j D , d V i D d t = k i D ( S i D + E i D ) - ( m i D + λ i D ) V i D + ∑ j = 1 n ϕ i j V V j D , d S i H d t = B i + λ i H V i H + σ i H ( 1 - γ i H ) E i H - m i H S i H - β i H S i H I i D + ∑ j = 1 n ψ i j S S j H , d E i H d t = β i H S i H I i D - ( m i H + σ i H + k i H ) E i H + ∑ j = 1 n ψ i j E E j H , d I i H d t = σ i H γ i H E i H - ( m i H + μ i H ) I i H + ∑ j = 1 n ψ i j I I j H , d V i H d t = k i H E i H - ( m i H + λ i H ) V i H + ∑ j = 1 n ψ i j V V j H . (1)
All parameters and their interpretations are listed in Table 1.
Ai describes the annual birth rate of the dog population in patch i; β i D denotes the transmission coefficient between dogs in patch i and β i D S i D I i D describes the transmission of rabies from infectious dogs to susceptible dogs in this patch; 1 / σ i D represents the incubation period of infected dogs in patch i; γ i D is the risk factor of clinical outcome of exposed dogs in patch i. Therefore, σ i D γ i D E i D denotes dogs that develop clinical rabies and enter the susceptible class and the rest σ i D ( 1 − γ i D ) E i D denotes the exposed dogs that do not develop clinical rabies; m i D is the non-disease related death rate for dogs in patch i; k i D is the vaccination rate of dogs and λ i D denotes the loss rate of vaccination immunity for dogs in patch i; μ i D is the disease-related death rate for dogs in patch i.
For the human population, similarly Bi describes the annual birth rate of the human population in patch i; β i H denotes the transmission coefficient from dogs to humans in patch i and β i H S i H I i D describes the transmission of rabies from infectious dogs to susceptible dogs in this patch; 1 / σ i H represents the incubation period of infected humans in patch i; σ i H γ i H E i H describes exposed people that become infectious and σ i H ( 1 − γ i H ) E i H describes the exposed people that return to be susceptible; m i H is the non-disease related death rate for humans in patch i; k i H is the vaccination rate of dogs and λ i H denotes the loss rate of vaccination immunity for huamns in patch i; μ i H is the disease-related death rate for humans in patch i.
ϕ
i
j
K
≥
0
(K = S, E, I, V) is the immigration rate from patch j to patch i for i ≠ j of susceptible, exposed, infectious, and vaccinated dogs, respectively; ψ i j K ≥ 0 (K = S, E, I, V) is the immigration rate from patch j to patch i for i ≠ j of susceptible, exposed, infectious, and vaccinated humans, respectively. Then ∑ j ≠ i ϕ i j K K i D (K = S, E, I, V) describes the corresponding subclass of the dog population that enter into patch i from other patches and ∑ j ≠ i ϕ j i K K i D denotes the corresponding subclass dog population that leave patch i. Meanwhile, the immigrations of humans are described in the same way by ψ i j K (K = S, E, I, V).
Data used to simulate our model are from the Data-Center of China Public Health Science reported by China CDC. After the 2003 SARS outbreak, the Chinese government strengthened its public health disease surveillance system. From 2004, the digital monthly reporting system has been replaced by a web-based, real-time reporting system which covers 39 diseases across all regions of the country. Each case is reported with the detailed information including sex, age, date of infection, diagnosis and death, the address of reporting hospital, and the reporting district administrative code. This well-established surveillance system provides valuable data for mathematical modelers in studying these infectious diseases.
We used a two-patch submodel to simulate the data of human rabies from 2004 to 2012 in three pairs of provinces: Guangxi and Guizhou, Fujian and Hebei, and Sichuan and Shaanxi (see Fig 2). Each province is regarded as a patch in the model (n = 2). The parameters about humans inculding the annual birth rate and natural death rate of humans in each province are adopted from the “China Health Statistical Yearbook 2012” ([17]). The incubation period for rabies is typically 1–3 months ([2]), we assume that it is 2 months on average, thus σ i H = 6 / y e a r. Similarly, we also have σ i H = 10 / y e a r ([18]). The disease induced death rates of humans and dogs are assumed to be 1 ([5]). According to [5], the vaccination rate k i H of humans in China is about 0.5 and the risk factor of clinical outcome of exposed dogs γ i D is 0.4. Based on studies the minimum duration of immunity for canine is 3 years ([19]), we assume that the loss rate of vaccination immunity for dogs in patch i is λ i D = 1 / 3 / y a e r ≈ 0 . 33 / y e a r. Rabies mortality after untreated bites by rabid dogs varies from 38% to 57% ([20]), thus we take the average 47.5% as the risk factor of clinical outcome of exposed humans.
The difficulty in parameter estimations is that there is no scientifically or officially reported data on dogs in China. So the values of Ai used in simulations are estimated based on the dog density from the household survey ([21]), the total areas of provinces, the density of human population and other research results ([9, 10, 15]). Now we assume that the immigration rates of susceptible, exposed, infectious and vaccinated dogs are same. Additionally, susceptible, exposed and vaccinated humans also move in the same rate but infectious humans do not move inter-provincially which is set as ψ i j H = 0. All other parameters are left to be unknown and estimated through simulating the model by the data.
The basic reproduction number ℛ0 is defined as the expected number of secondary cases produced by a typical infection in a completely susceptible population ([22]). Here, the basic reproduction number of rabies which reflects the expected number of dogs infected by a single infected dog, is derived from the mathematical model that describes the transmission dynamics of rabies following the method in van den Driessche and Watmough [23]. Mathematically, R0 is defined as the dominant eigenvalue of a linear operator. In S1 Text, the overall basic reproduction number ℛ0 for the whole system is calculated. The isolated basic reproduction number,
𝓡 0 i = β i D σ i D γ i D A i ( m i D + λ i D ) m i D ( m i D + μ i D ) ( m i D + σ i D + k i D ) ( m i D + λ i D + k i D ) , (2)
is the basic reproduction number in one single patch (patch i here) when all the immigration rates are zero. That is the basic reproduction number in an isolated patch under the assumption that there is no immigration at all. For the two-patch submodel, R0 can be expressed as
R0=(β1DS1D*σ1Dγ1D(m2D+σ2D+k2D+ϕ12E)(m2D+μ2D+ϕ12I)+β2D*S2D*σ2Dγ2D(m1D+σ1D+k1D+k1D+ϕ21E)(m1D+μ1D+ϕ21I)+β2DS2D*σ1Dγ1Dϕ12Eϕ21I+β1DS1D*σ2Dγ2Dϕ21Eϕ12I+((β1DS1D*σ1Dγ1D(m2D+σ2D+k2D+ϕ21E)(m2D+μ2D+ϕ12I)+β2DS2D*σ2Dγ2D(m1D+σ1D+k1D+ϕ21E)(m1D+μ1D+ϕ21I)+ϕ12Eϕ21Iβ2DS2D*σ1Dγ1D+ϕ21Eϕ12Iβ1DS1D*σ2Dγ2D)2−4((m1D+σ1D+k1D+ϕ21E)(m2D+σ2D+k2D+ϕ12E)−ϕ21Eϕ12E)((m1D+μ1D+ϕ21I)(m2D+μ2D+ϕ12I)−ϕ21Iϕ12I(β1DS1D*β2DS2D*σ1Dγ1Dσ2Dγ2D))12/{2((m1D+σ1D+k1D+ϕ21E)(m2D+σ2D+k2D+ϕ12E)−ϕ21Eϕ12E)((m1D+μ1D+ϕ21I)(m2D+μ2D+ϕ21I)(m2D+μ2D+ϕ12I)−ϕ21Iϕ12I}. (3)
The value of R0 gives an important threshold that determines if the disease will die out or not eventually. Roughly speaking, if ℛ0 > 1 each primary infected dog averagely will produce more than one secondary infected dog. Therefore the disease will persist. Conversely, if ℛ0 < 1 the expected number of secondary case produces by the primary case is less than one. Thus the disease will die out. The purpose is to reduce R0 by possible disease control strategies. However, the formula is very complicated and impossible to analyze the relationship between the parameters and ℛ0 even for a two-patch model. Sensitivity analysis can aid in discovering how each parameter quantitatively affects R0. Furthermore, we will study how the immigration rate affect the basic reproduction numbers of the whole system and the isolated patchs by performing some sensitivity analyses.
In this section, we first use the two-patch submodel to simulate the reported human rabies data from Guangxi and Guizhou, Sichuan and Shaanxi, and Fujian and Hebei, respectively. Then we carry out some sensitivity analyses of the basic reproduction number in terms of some parameters of dogs, especially the immigration rates between provinces.
Fig 4 presents the reported human rabies cases in different provinces in Mainland China in the years 2004, 2008, and 2012. Although the numbers of cases decrease in some of the endemic provinces such as Guangxi and Hunan, some other provinces such as Shanxi and Shaanxi keep increasing. Some non-endemic provinces are becoming endemic in recent years. For example, Hebei, Shanxi and Shaanxi. Discrete phylogeographic analysis for China I strain ([12, 14]) indicates the linkage of rabies virus between Sichuan and Shaanxi, Guangxi, and Guizhou, and Fujian and Hebei (Fig 2).
(a) Hebei and Fujian. From Guo et al. [14], we know that Hebei and Fujian are epidemiologically linked. In Hebei, there was only one human rabies case reported in 2000 ([5]), while it is now one of the 15 provinces having more than 1,000 cumulative cases and is included in “Mid-to-long-term Animal Disease Eradication Plan for 2012–2020” project. We take Hebei and Fujian as two patches in model Eq (1) (when n = 2) and simulate the numbers of human cases from 2004 to 2012 by the model. In Fig 5, the solid blue curves represent simulation results and the dashed red curves are reported numbers of human rabies cases from 2004 to 2012, which show a reasonable match between the simulation results and reported data from China CDC. Based on the values of parameters in the simulations and the formula of the basic reproduction number in the two-patch model, we calculated that ℛ0 = 1.0319. That means the disease will not die out in this two-patch system.
Interestingly, now we assume there is no immigration of both dogs and humans in this system and calculated the isolated basic reproduction number in each province. The isolated basic reproduction numbers for Hebei and Fujian are ℛ 0 Hebei = 0 . 5477 and ℛ 0 Fujian = 0 . 8197, respectively. Under this assumption the disease would die out in both provinces since their isolated basic reproduction number is less than one. This example theoretically shows the possibility that the immigration of dogs can lead the disease to a worse scenario even it could be eliminated in each isolated patch. It is remarkable that we only mentioned the dog immigration here because a simple observation to the formula of the basic reproduction number in the S1 Text shows that only the immigration rates of dogs (ϕ i j K for K = S, E, I, V) can affect it. In fact, only dogs can carry the rabies virus and then spread it to humans and other dogs. This transmission feature supports our mathematical analysis.
(b) Guizhou and Guangxi. A statistically significant translocation event is also predicted between Guizhou and Guangxi in Yu et al. [12]. Fig 4 shows that Guizhou and Guangxi have large numbers of human rabies cases (both are in top 5 endemic provinces in China) in recent years. Particularly, the number of human deaths caused by rabies virus in Guangxi is ranked the highest in China. Similar simulations were carried out here to these two provinces and results are shown in Fig 6. The isolated basic reproduction numbers for Guizhou and Guangxi are calculated as ℛ 0 Guizhou = 1 . 5998 and ℛ 0 Guangxi = 6 . 1905, respectively, while the basic reproduction number for the whole system is estimated to be ℛ0 = 4.9211. To eliminate rabies we need some effective control strategies that can reduce ℛ0 significantly. Thus it is even more challenging to control and prevent the disease in Guangxi and Guizhou from a numerical perspective.
(c) Sichuan and Shaanxi. Shaanxi, which is now an alarming province for rabies in China, had only 15 cumulative human cases from 2000 to 2006 (only 2 to 3 cases every year on average). However, 26 human cases were reported in 2009 and the number keeps increasing after that. Rabies was found to spread along the road network [13]. With the parameters in Fig 7, the isolated basic reproduction numbers for Sichuan and Shaanxi are ℛ 0 Sichuan = 1 . 3414 and ℛ 0 Shaanxi = 1 . 0061, respectively, while the basic reproduction number for the two provinces with immigration is ℛ0 = 1.5085 which is greater than both of these two isolated ones. Numerically, that means more efforts may be needed to eliminate the virus in humans if the immigration is involved.
Additionally, we show some direct comparisons of numerical simulations on the number of human cases from the model with immigration and without immigration. The additional green curves represent simulations of the human cases without any immigration in Hebei, Guizhou and Shaanxi, respectively. In Hebei, Fig 8(a) indicates the human infectious population size goes to zero faster without immigration which is consistent with the fact that the isolated basic reproduction number (0.5477) in Hebei is less than one. Similarly result can be observed in Fig 8(b) for Guizhou. Furthermore, Fig 8(c) shows that if there is no dog immigration in Shaanxi, the human rabies cases would decrease fast while it increased fast in reality.
We now study how the basic reproduction number ℛ0 depends on parameters of dogs, especially the immigration rates ϕ i j K, where K = S, E, I, V. For the sake of implicity, we consider the two-patch submodel and the corresponding basic reproduction number given in Eq (3). We consider the following three cases.
(i) Immigration of dogs between patches with different transmission rates. Suppose β 1 D = 3 × 10 − 7 > β 2 D = 1 × 10 − 7, ϕ 12 K = ϕ 12 and ϕ 21 K = ϕ 21, where K = S, E, I, R. A1 = 2 × 66, λ 1 D = 0 . 42, σ 1 D = 0 . 42, γ 1 D = 0 . 4, m 1 D = 0 . 08, k 1 D = 0 . 09, μ 1 D = 1, the remaining parameters of dogs in patch 2 are the same as the corresponding parameters of dogs in patch 1. Here the only difference between the two patches in that the transmission coefficients of infectious dogs to susceptible dogs are different. Then the isolated basic production numbers satisfy the inequality: ℛ 0 1 = 2 . 3246 > ℛ 0 2 = 0 . 7749. So rabies is endemic in patch 1 and will die out in patch 2. First, let (the immigration rate of dogs from patch 1 to patch 2) ϕ12 = 0.02. It is shown in Fig 9 that ℛ0 decreases as ϕ21 (the immigration rate of dogs from patch 2 to patch 1) increases. Then, let ϕ21 = 0.5, ℛ0 increases as ϕ12 increases. Furthermore, if ϕ21 is small and ϕ12 is large, ℛ0 is greater than both ℛ 0 1 and ℛ 0 2. To reduce ℛ0, we need to control ϕ12 small enough. For example, let ϕ21 = 0.5, ϕ12 = 0.01, then we obtain that ℛ 0 < min { ℛ 0 1 , ℛ 0 2 }. If ϕ21 = 0.4, ϕ12 = 0.3, then ℛ0 = 1.6274, which is smaller than ℛ 0 1 but greater than ℛ 0 2. Thus, if we can control the immigration rates of dogs in an appropriate range, the endemic level will be lower.
(ii) Immigration of dogs between patches with different vaccination rates. We assume that dogs move at the same rate regardless of their subclasses (ϕ 12 K = ϕ 12 and ϕ 21 K = ϕ 21 for K = S, E, I, V). Then let dogs in patch 1 have a higher vaccination rate than those in patch 2: k 1 D = 0 . 5 > k 2 D = 0 . 09. All the remaining parameters of dogs in patch 2 are the same as the corresponding parameters of dogs in patch 1. Fig 10 presents the basic reproduction number ℛ0 in terms of the immigration rates. Firstly, ℛ0 increases as the immigration rates increase at most of the time. This is consistent with our previous simulation results: the dog movements bring difficulties to rabies control. Secondly, a detailed observation in the range of ℛ0 indicates that it is more sensitive in ϕ12. Therefore we conclude that immigration of dogs from the patch with lower vaccination rate to a patch with higher vaccination rate is more dangerous.
It is notable that ℛ0 might be greater than both isolated basic reproduction numbers. For example, let ϕ21 = 0.95 and ϕ12 = 0.4, and all other parameters be the same as in Case (ii). Then ℛ 0 = 1 . 2974 > max { ℛ 0 1 , ℛ 0 2 }. That is, the immigration of dogs might lead to a more serious situation.
(iii) Immigration of infective dogs between patches. Now we fix all immigration rates of dogs to 0.2 except ϕ 21 I (the immigration rate of infective dogs from patch 1 to patch 2), then ℛ0 increases quickly as ϕ 21 I increases, as it is shown in Fig 11(a). On the other hand we fix all immigration rates of dogs to 0.2 except ϕ 12 I (the immigration rate of infective dogs from patch 2 to patch 1), then ℛ0 decreases as ϕ 21 I increases, as it is shown in Fig 11(b). Interestingly, compare with Case ii, we found that immigration of infectious dogs from the patch with a high vaccination rate to a patch with a low vaccination rate is more dangerous. The patch with a low vaccination rate actually has a week protection from the virus, thus infectious dogs from another patch may spread the disease faster.
In 1999, human rabies cases were reported in about 120 counties in mainland China, mainly in the southern provinces. Now outbreaks of human rabies have been reported in about 1000 counties and the disease has spread geographically from the south to the north. Phylogeographic analyses for rabies virus strains ([12, 14]) indicate that prevalent strains in northern provinces are indeed related to the remote southern provinces. It is believed that the geographical spread of rabies virus are caused by the transportation of dogs.
In this paper, a multi-patch model is proposed to describe the spatial transmission dynamics of rabies in China and to investigate how the immigration of dogs affects the geographical spread of rabies. The expression and sensitivity analysis of the basic reproduction number indicates that the movement of dogs plays an essential role in the spatial transmission of rabies. As mentioned in [8], reducing dog birth rate and increasing dog immunization coverage rate are the most effective methods in controlling human rabies infections in China. They also play important roles in controlling the spatial spread of rabies based on the multi-patch model. WHO (World Health Organization) recommends that 70% of dogs in a population should be immunized to eliminate the rabies. Unfortunately, this rate is still lower than 10% in most regions in China. Therefore, efforts to bring the awareness of the importance of treatments and enhance the vaccination coverage in dogs are important to control the disease in China.
We also performed some numerical simulations to study the effects of the immigration rate in three pairs of provinces in China: Guizhou and Guangxi, Hebei and Fujian, Sichuan and Shaanxi, as shown in Fig 2. First of all, the immigration may lead a basic reproduction number to be larger than one even if the isolated basic reproduction numbers are all less than one. Therefore, the immigration of dogs is the main factor for the long-distance inter-provincial spread of rabies. We note that the transportation of dogs even between non-endemic provinces, such as Fujian and Hebei, can cause human rabies in Hebei to increase greatly. Additionally, the movement of dogs from regions with a low vaccination rate also makes the situation worse. Attention should be paid not only to the provinces with more reported cases but also to the provinces with low vaccination rates. In those extremely poor areas, where dogs have a low vaccination coverage, the dog trade business and transportation to other areas will contribute to the geographical spread of rabies significantly. To control the disease at a national level, more efforts are needed in these regions.
The primary purpose of the transportation of dogs in China is believed to be related to food business. In some areas, such as the endemic provinces Guizhou and Guangxi, people eat dogs due to minority culture or harsh climate. There is no open market for selling and buying dogs for business purpose, however the black market always exists. It is frequently reported that trucks sometimes full of dogs are intercepted by animal lovers in the inter-provincial highway. Sometimes more than one thousand dogs were crammed into many tiny cages in one truck. The efficiency of such dog transportation has been enhanced by the fast development and expansion of the highway system in China in the past ten years. Chinese law requires that the transported animals must be certified as vaccinated for rabies and other diseases. However, dog traders are found to falsify the paperwork for most of the dogs in the truck to reduce their cost. Thus it would be important to regulate the market and implement certain policies on dogs (such as vaccine records) and the dog traders (such as licenses). During our research, we found that it was very difficult to find the information on dog population in China due to the lack of dog registration management. Since a large number of dogs are transported from provinces to provinces, it is necessary to register and manage such transportation properly. In particular, dogs carrying rabies viruses can easily spread the virus to other dogs when they are crowded into a small space during the trip. The last case of our sensitivity analysis shows the oblivious dangers resulted from the transportation of infectious dogs that has a destination with a low vaccination rate. We suggest creating strict and uniform procedures to test the dogs that will be transported.
We used a deterministic system to study the geographical spread of rabies in China and simulated the annual data in some provinces. Stochasticity is not considered in our model, and we also think seasonality plays an important role in the transmission of rabies. Therefore a mathematcal model which includes certain randomness and seasonality may help us to understand this problem better. Meanwhile, we only applied two-patch model to simulate the data in two provices. A more general case which can discuss the complex transmission among three or more provinces is interesting to study.
Chinese government has devoted a large amount of financial resource to the control of rabies, particularly in vaccinations. According to the statistics reported in “Chinese Rabies Prevention and Control Status” ([17]), about 12–15 million doses of human rabies vaccines are administered in China each year, accounting for 80 percent of the total global consumption. The production and administration of human rabies vaccines cost the country more than RMB 10 billion ($1.56 billion) each year. However, most of these efforts focused on humans and the vaccination rate of dogs in China still remains low. Under this high-risk environment for rabies, the only way to reduce deaths caused by rabies is to provide treatment immediately to exposures (contacts with category II and III). Then the total cost could be about RMB 24.5 billion annually if all of these exposures receive PEP treatments. Remarkably, the vaccines for dogs are less expensive than that for humans, but the dog vaccination implementation requires a continuously huge human, material and financial resources. It will be interesting to investigate how to optimize the resources and efforts and how to take the socioeconomic factors into consideration in order to pursue the control and elimination of rabies virus in humans.
|
10.1371/journal.pgen.1006954 | PKA activity is essential for relieving the suppression of hyphal growth and appressorium formation by MoSfl1 in Magnaporthe oryzae | In the rice blast fungus Magnaporthe oryzae, the cAMP-PKA pathway regulates surface recognition, appressorium turgor generation, and invasive growth. However, deletion of CPKA failed to block appressorium formation and responses to exogenous cAMP. In this study, we generated and characterized the cpk2 and cpkA cpk2 mutants and spontaneous suppressors of cpkA cpk2 in M. oryzae. Our results demonstrate that CPKA and CPK2 have specific and overlapping functions, and PKA activity is essential for appressorium formation and plant infection. Unlike the single mutants, the cpkA cpk2 mutant was significantly reduced in growth and rarely produced conidia. It failed to form appressoria although the intracellular cAMP level and phosphorylation of Pmk1 MAP kinase were increased. The double mutant also was defective in plant penetration and Mps1 activation. Interestingly, it often produced fast-growing spontaneous suppressors that formed appressoria but were still non-pathogenic. Two suppressor strains of cpkA cpk2 had deletion and insertion mutations in the MoSFL1 transcription factor gene. Deletion of MoSFL1 or its C-terminal 93-aa (MoSFL1ΔCT) was confirmed to suppress the defects of cpkA cpk2 in hyphal growth but not appressorium formation or pathogenesis. We also isolated 30 spontaneous suppressors of the cpkA cpk2 mutant in Fusarium graminearum and identified mutations in 29 of them in FgSFL1. Affinity purification and co-IP assays showed that this C-terminal region of MoSfl1 was essential for its interaction with the conserved Cyc8-Tup1 transcriptional co-repressor, which was reduced by cAMP treatment. Furthermore, the S211D mutation at the conserved PKA-phosphorylation site in MoSFL1 partially suppressed the defects of cpkA cpk2. Overall, our results indicate that PKA activity is essential for appressorium formation and proper activation of Pmk1 or Mps1 in M. oryzae, and phosphorylation of MoSfl1 by PKA relieves its interaction with the Cyc8-Tup1 co-repressor and suppression of genes important for hyphal growth.
| The cAMP-PKA signaling pathway plays a critical role in regulating various cellular processes in eukaryotic cells in response to extracellular cues. In the rice blast fungus, this important pathway is involved in surface recognition, appressorium morphogenesis, and infection. However, the exact role of PKA is not clear due to the functional redundancy of two PKA catalytic subunits CPKA and CPK2. To further characterize their functions in growth and pathogenesis, in this study we generated and characterized the cpkA cpk2 double mutant and its suppressor strains. Unlike the single mutants, cpkA cpk2 mutant had severe defects in growth and conidiation and was defective in appressorium formation and plant infection. Interestingly, the double mutant was unstable and produced fast-growing suppressors. In two suppressor strains, mutations were identified in a transcription factor gene orthologous to SFL1, a downstream target of PKA in yeast. Deletion of the entire or C-terminal 93 residues of MoSFL1 could suppress the growth defect of cpkA cpk2. Furthermore, the terminal region of MoSfl1 was found to be essential for its interaction with the MoCyc8 co-repressor, which may be negatively regulated by PKA. Therefore, loss-of-function mutations in MoSFL1 can bypass PKA activity to suppress the growth defect of cpkA cpk2.
| Magnaporthe oryzae is the causal agent of rice blast, which is one of the most important rice diseases worldwide. In the past two decades, M. oryzae has been developed as a model organism to study fungal-plant interactions because of its economic importance and the experimental tractability [1–3]. For plant infection, the fungus forms a highly specialized infection cell called an appressorium to penetrate plant cuticle and cell wall [4]. After penetration, the narrow penetration peg differentiates into bulbous invasive hyphae [5] that grow biotrophically inside penetrated plant cells [6]. Various apoplastic and cytoplasmic effectors are known to play critical roles in suppressing plant defense responses during different stages of invasive growth [7]. At late infection stages, lesions are formed and the pathogen produces conidiophores and conidia on diseased plant tissues under favorable conditions.
Appressorium formation is initiated when conidia land and germinate on plant surfaces. On artificial hydrophobic surfaces that mimic the rice leaf surface, M. oryzae also forms melanized appressoria. On hydrophilic surfaces, appressorium formation can be induced by cAMP, IBMX, or cutin monomers [8]. Although late stages of appressorium formation is regulated by the Pmk1 MAP kinase, the cAMP-PKA (protein kinase A) pathway is involved in recognizing surface hydrophobicity to initiate appressorium formation, appressorium turgor generation, and invasive growth [9–11]. Deletion of the MAC1 adenylate cyclase (AC) gene results in mutants that are defective in appressorium formation [12]. In addition to Cap1 AC-interacting protein [13], heterotrimeric G-proteins and Rgs1 have been shown to function upstream from the cAMP-PKA pathway [2, 14]. The PdeH high-affinity cAMP phosphodiesterase is also important for successful establishment and spread of the blast disease [15].
The PKA holoenzyme consists of two regulatory subunits and two catalytic subunits. Binding of cAMP with the regulatory subunit results in the detachment and activation of the catalytic subunits [16]. In M. oryzae, the CPKA gene encoding a catalytic subunit of PKA is dispensable for hyphal growth but the cpkA mutant was delayed in appressorium formation and defective in appressorium turgor generation and plant penetration. In addition, the cpkA mutant still responds to exogenous cAMP for appressorium formation on hydrophilic surfaces [10, 11], suggesting that another PKA catalytic subunit gene must exist and play a role in surface recognition and infection-related morphogenesis in M. oryzae. In the budding yeast Saccharomyces cerevisiae, three genes, TPK1, TPK2, and TPK3, encode PKA catalytic subunits and the triple mutant is inviable [17]. The fission yeast Schizosaccharomyces pombe has only one PKA catalytic subunit gene, PKA1, that is important but not essential for normal growth [18]. In the human pathogen Aspergillus fumigatus, the pkaC1 pkaC2 double mutant is delayed in conidium germination in response to environmental nutrients and is significantly reduced in virulence [19]. In the wheat scab fungus Fusarium graminearum, deletion of both CPK1 and CPK2 caused severe defects in growth and conidiation. The cpk1 cpk2 double mutant was sterile in sexual reproduction and nonpathogenic [20]. In the basidiomycete Ustilago maydis, the phenotype of the adr1 uka1 double mutant has similar phenotype with the adr1 mutant and is defective in yeast growth, mating, and plant infection [21].
In S. cerevisiae, Sfl1 is one of the downstream transcription factors of the cAMP-PKA pathway. When functioning as a repressor, it is involved in the repression of flocculation-related genes, including FLO11 and SUC2 [22, 23]. As an activator, SFL1 is involved in the activation of stress-responsive genes such as HSP30 [24]. The major PKA catalytic subunit Tpk2 negatively regulates its repressor function [25]. In M. oryzae, deletion of MoSFL1 has no obvious effect on vegetative growth but results in reduced virulence and heat tolerance [26]. Several Sfl1-interacting proteins have been identified in the budding yeast, including Cyc8, Tup1, and various mediator components [23, 27]. Although it lacks intrinsic DNA-binding activities, the Cyc8-Tup1 (also known as Ssn6-Tup1) co-repressor complex interacts with various transcription factors with sequence-specific DNA binding motifs, including Sfl1, Mig1, Crt1, and α2, to negatively regulate different subsets of genes [27, 28]. In S. cerevisiae, Cyc8 functions as an adaptor protein required for the interaction between Tup1 tetramers and DNA-binding transcription factors [29].
To further characterize the roles of PKA in growth and infection, in this study we generated and characterized the cpk2 and cpkA cpk2 double mutants and spontaneous suppressors of cpkA cpk2 in M. oryzae. Our results demonstrate that CPKA and CPK2 have specific and overlapping functions. The cpkA cpk2 double mutant had severe defects in growth and conidiation and failed to form appressoria or infect plant through wounds. Spontaneous mutations or deletion and truncation mutations in MoSFL1 suppressed the defects of cpkA cpk2 in hyphal growth and appressorium formation but not invasive growth and lesion development. In affinity purification and co-IP assays, MoCyc8 interacted with the full-length but not truncated MoSfl1. Treatment with exogenous cAMP also reduced the interaction of MoSfl1 with MoCyc8 and MoTup1. Furthermore, the S211D mutation in MoSFL1 suppressed the growth defect of cpkA cpk2. In F. graminearum, 29 of 30 suppressor strains of cpk1 cpk2 mutant had mutations in FgSFL1, with 15 of them truncated of its C-terminal region. Taken together, our results indicate that PKA activity is essential for appressorium formation in M. oryzae, and phosphorylation of MoSfl1 by PKA likely relieves its interaction with the Cyc8-Tup1 co-repressor and suppression of genes important for hyphal growth and appressorium development. The inhibitory function of Sfl1 orthologs on hyphal growth is likely conserved in filamentous fungi because similar suppressor mutations in FgSFL1 were identified in spontaneous suppressor strains of the cpk1 cpk2 mutant in F. graminearum.
In M. oryzae, the PKA regulatory subunit is encoded by SUM1, a suppressor of the mac1 deletion mutant [30]. In an effort to identify proteins interacting with known virulence factors, including Sum1, we generated the SUM1-S-tag fusion and transformed it into the wild-type strain 70–15. Total proteins were isolated from the resulting transformant B22 (Table 1) and subjected to affinity purification and MS analysis after trypsin digestion as described [13, 31]. CpkA and Cpk2 (MGG_02832) were among the Sum1-interacting proteins identified in all three independent biological replicates (S1 Table). Cpk2 shares 48% amino acid identity with CpkA but it has a shorter N-terminal region (S1 Fig).
To confirm their interaction by co-immunoprecipitation (co-IP) assays, the SUM1-S, CPKA-3×FLAG, and CPK2-3×FLAG constructs were generated and transformed into the wild-type strain Guy11 in pairs. Western blot analysis with the resulting transformants showed that both PKA catalytic subunits strongly interact with Sum1 (S2 Fig).
To determine the function of PKA catalytic subunits, we generated the cpk2 and cpkA cpk2 deletion mutants (Table 1; S1 Fig). On complete medium (CM) plates, the cpkA mutant had no obvious growth defects but the cpk2 mutant was slightly reduced in growth rate. The cpkA cpk2 double mutant was viable but it was significantly reduced in growth rate (Table 2; Fig 1A). Unlike cpkA and cpk2, the cpkA cpk2 double mutant rarely produced conidia. In cultures induced for conidiation, the double mutant produced only a few conidia per plate. Under the same conditions, over 1×107 conidia/plate were produced by the wild type (Table 2).
Unlike cpkA that was delayed in appressorium formation, cpk2 had no obvious defects in appressorium formation (Table 2). However, the cpkA cpk2 mutant failed to form appressoria on hydrophobic plastic coverslips and GelBond membranes although conidium germination was normal (Fig 1B; Table 2). Even after prolonged incubation up to 72 h, no appressorium formation was observed in the double mutant. The cpkA cpk2 mutant also failed to form appressoria on barley and rice leaves (S3 Fig).
In spray infection assays with two-week-old seedlings of rice cultivar CO-39, numerous blast lesions were observed on leaves sprayed with Guy11 or the cpk2 mutant but no lesions were caused by the cpkA and cpkA cpk2 mutants (Fig 1C). Because the cpkA cpk2 mutant failed to form appressoria, we conducted injection infection assays. Whereas the cpk2 mutant was as virulent as the wild type, the cpkA mutant failed to cause lesions on intact leaves but caused limited necrosis at the wounding sites. No lesions or necrosis at the wounding sites were observed on leaves inoculated with the cpkA cpk2 mutant (Fig 1C). These results indicate that PKA activities are essential for appressorium formation and invasive growth after penetration in M. oryzae.
Although the cpkA cpk2 mutant was blocked in appressorium formation, we noticed that majority (over 83%) of its germ tubes were curved one direction after incubation on hydrophobic side of GelBond membranes for 24 h (Fig 1B). Interestingly, when assayed for appressorium formation on the hydrophilic side of GelBond membranes, the majority of the cpkA cpk2 conidia failed to germinate (Fig 1B; Table 3). For the ones (<25%) germinated, germ tubes of cpkA cpk2 mutant failed to form appressoria (Fig 1B). Conidia of the wild-type, cpkA, and cpk2 strains germinated but failed to form appressoria under the same conditions. In the presence of 5 mM cAMP, over 76% of the wild-type germ tubes formed appressoria on hydrophilic surfaces. However, exogenous cAMP had no stimulatory effects on either conidium germination or appressorium formation in the cpkA cpk2 mutant (Fig 1B; Table 3). Over 75% of the double mutant conidia failed to germinate and the ones germinated failed to form appressoria or display germ tube curling defects in the presence of 5 mM cAMP. These results indicate that the cpkA cpk2 mutant still recognizes surface hydrophobicity for germination and germ tube growth but not for appressorium formation.
We then assayed the intracellular cAMP level in vegetative hyphae harvested from liquid CM cultures. In the cpk2 mutant, the intracellular cAMP level was similar to that of the wild type. However, the cpkA and cpkA cpk2 mutants had higher intracellular cAMP levels than the wild type (Fig 2A). In comparison with the wild type, the double mutant was increased approximately 3-fold in intracellular cAMP. These results suggest that reduced or lack of PKA activities results in an increase in intracellular cAMP in M. oryzae.
Because the Pmk1 MAP kinase is essential for appressorium formation [34], we assayed its activation with an anti-TpEY phosphorylation specific antibody. To our surprise, although its expression was not affected, Pmk1 phosphorylation was increased in the cpkA cpk2 mutant (Fig 2B). However, the double mutant was reduced in the expression and phosphorylation levels of Mps1 MAP kinase (Fig 2B) that is required for appressorium penetration and conidiation [33]. These results suggest that over-activation of the Pmk1 MAP kinase pathway is not sufficient to stimulate appressorium formation in the absence of PKA activities, and reduced Mps1 activities may be related to conidiation defects of cpkA cpk2.
Interestingly, the cpkA cpk2 mutant was unstable when cultured on the oatmeal agar (OTA) plates and fast-growing sectors caused by spontaneous suppressor mutations often became visible in cultures older than 10 days (Fig 3A). Twenty suppressor strains with faster growth rate than the original mutant were isolated. All of them had similar colony morphology and produced more aerial hyphae than the cpkA cpk2 mutant. On average, the growth rate of suppressor strains recovered to approximately 83% of that of the wild type (S4 Fig). Conidiation also was partially rescued in these suppressor strains although to a much lesser degree than the recovery in growth rate (S4 Fig).
Although they varied slightly in growth rate and conidiation (S4 Fig), all the 20 suppressor strains had similar defects as strain CCS1 that was described and presented in figures below. Besides having similar colony morphology, suppressor strains produced melanized hyphal tips in aerial hyphae of 10-day-old OTA cultures (Fig 3B). In infection assays with rice seedlings, none of the suppressor strains caused lesions on intact or wounded leaves (Fig 3C). Therefore, mutations occurred in these suppressor strains only suppressed the defects of the cpkA cpk2 mutant in hyphal growth but not plant infection.
On artificial hydrophobic surfaces, over 95% of the conidia from the suppressor strains formed appressoria after incubation for 24 h. Interestingly, they also developed appressoria on the hydrophilic surface of GelBond membranes (Fig 3D). However, approximately 40% of appressoria formed by suppressor strains were abnormal in morphology on hydrophobic or hydrophilic surfaces (Fig 3E). Unlike normal dome-shaped appressoria, the majority of appressoria formed by suppressor strains had irregular shapes (Fig 3E). Although they were still melanized, many appressoria formed by the suppressor strains had projections at one side (Fig 3E). These results indicate that suppressor mutations in these strains also only partially rescued the defect of cpkA cpk2 in appressorium morphogenesis.
To identify suppressor mutations, we selected eight genes (S2 Table) [27, 35–40] that are orthologous to downstream targets of PKA in the budding yeast, including SOM1 and CDTF1 [40] for PCR and sequencing analysis in the selected suppressor strains CCS1, CCS4, CCS7 and CCS14. Whereas suppressor stains CCS4 and CCS14 had no mutations in these candidate genes, both CCS1 and CCS7 had mutations in the MoSFL1 [26] gene that encodes a transcription factor with a conserved HSF (heat shock factor) DNA-binding domain in the N-terminal region (residues 124–225) [26]. In suppressor strain CCS1, 10 extra nucleotides CCCCCGCCGC were inserted in the coding region of MoSFL1 (between 1556 and 1557), resulting in a frameshift change at residue P414. In suppressor strain CCS7, a 1241-bp deletion occurred in the coding region of MoSFL1 (Δ405–1645), resulting in the truncation of 78% of its amino acids.
In M. oryzae, deletion of MoSFL1 had no effects on hyphal growth although it was reduced in virulence [26]. To confirm whether insertion or truncation mutation in MoSFL1 has suppressive effects, the MoSFL1 gene replacement construct was transformed into the cpkA cpk2 mutant. Bleomycin-resistant transformants were screened by PCR for deletion of MoSFL1 and confirmed by Southern blot (S5 Fig). The resulting cpkA cpk2 Mosfl1 mutant had similar phenotypes as spontaneous suppressor strains (Fig 4; Table 4), including recovered growth rate and increased conidiation in comparison with cpkA cpk2. Melanized appressoria were efficiently formed by the triple mutant but it failed to cause lesions on rice leaves, further confirming that loss-of-function mutations in MoSFL1 rescue the growth defect of cpkA cpk2. Therefore, MoSFL1 must function as a negative regulator of vegetative hyphal growth and phosphorylation of MoSfl1 by PKA relieves its suppressive effects.
In suppressor strain CCS1, the 10-bp insertion in MoSFL1 causes frameshift and results in the truncation of its C-terminal 414–588 aa. Sequence alignment showed that this C-terminal region of MoSfl1 is well conserved among its orthologs from other filamentous fungi, including Fusarium graminearum and Neurospora crassa (S6 Fig). To verify its importance, we generated a MoSFL1ΔCT gene-replacement construct (S5 Fig) to delete residues 496–588 of MoSFL1 in the cpkA cpk2 mutant. The resulting cpkA cpk2 MoSFL1ΔCT triple mutants (S5 Fig) had the same phenotypes with the cpkA cpk2 Mosfl1 mutant and spontaneous suppressor strains (Fig 4; Table 4). These results suggested that the C-terminal region of MoSfl1 is essential for its negative regulator function although it has no known protein motifs. Whereas the N-terminal region of MoSfl1 is involved in DNA binding, the C-terminal region may be responsible for protein-protein interactions to suppress the expression of its target genes important for hyphal growth.
Because deletion of residues 496–588 is suppressive to cpkA cpk2, this C-terminal region of MoSfl1 is likely responsible for interacting with other proteins as a negative regulator. To identify proteins differentially interacting with MoSfl1 and MoSfl1ΔCT, the 3×FLAG-MoSFL1 and 3×FLAG-MoSFL1ΔCT constructs were generated and transformed into the cpkA cpk2 mutant. Total proteins were isolated from the resulting 3×FLAG-MoSFL1 and 3×FLAG-MoSFL1ΔCT transformants and used for affinity purification with anti-FLAG M2 beads. Proteins co-purified with MoSfl1 or MoSfl1ΔCT were identified by mass spectrometry (MS) analysis after trypsin digestion as described in previous studies [13, 31].
Based on MS results from three biological replicates, MGG_03196 was the only protein that co-purified with MoSfl1 but not MoSfl1ΔCT (Table 5). Its ortholog in yeast, Cyc8 (Ssn6), forms a transcriptional co-repressor complex with Tup1 to regulate genes involved in a wide variety of physiological processes [28, 41, 42]. Interestingly, the Tup1 ortholog, MGG_08829, was one of the proteins that were commonly co-purified with MoSfl1 and MoSfl1ΔCT (Table 5). However, the number of MoTup1 peptides identified by MS analysis was significantly lower in the MoSFL1ΔCT transformant than in the MoSFL1 transformant (Table 5), suggesting a weaker interaction of MoSfl1 with MoTup1 when its C-terminal region is deleted. Based on the conserved nature of Tup1, Cyc8, and other components, it is likely that the Cyc8 and Tup1 orthologs also form a transcriptional co-repressor complex with MoSfl1 in M. oryzae, which is consistent with the interaction of Sfl1 with the Cyc8-Tup1 complex in yeast [27].
To confirm the importance of the C-terminal region of MoSfl1 in its interaction with Cyc8, the CYC8-S-tag construct was generated and co-transformed into the cpkA cpk2 mutant with 3×FLAG-MoSFL1 or -MoSFL1ΔCT. The resulting transformants CYS15 and CNC19 (Table 1) were confirmed by western blot analyses for the expression of transforming constructs. In co-IP assays, the MoSfl1 band was detected in both total proteins and elution from anti-S-tag agarose beads in the transformant expressing the CYC8-S and 3×FLAG-MoSFL1 constructs (Fig 5). However, the MoSfl1ΔCT band was detected only in total proteins isolated from the transformant expressing CYC8-S and 3×FLAG-MoSFL1ΔCT (Fig 5). These results confirmed that Cyc8 interacts with the full-length MoSfl1 but not MoSfl1ΔCT in M. oryzae. Interestingly, additional bands smaller than MoSfl1 or MoSfl1ΔCT were detected by the anti-FLAG antibody in transformants CYS15 and CNC19 but not in Guy11, suggesting that MoSfl1 proteins may be cleaved in vegetative hyphae.
Because deletion of the C-terminal region of MoSFL1 suppressed the growth defect of cpkA cpk2, phosphorylation of MoSfl1 by PKA may affect its interaction with the Cyc8-Tup1 complex. To test this hypothesis, the MoCYC8-S and 3×FLAG-MoSFL1 constructs were co-transformed into the wild-type strain Guy11. In the resulting transformant, treatment with 5 mM cAMP significantly reduced the MoCyc8-MoSfl1 interaction compared to treatment with 10 μM PKA inhibitor (PKI) H-89 (Fig 6A). These results indicate that stimulation of PKA activities by exogenous cAMP reduces the interaction of MoSfl1 with MoCyc8. Therefore, phosphorylation of MoSfl1 by PKA likely reduced the interaction of MoSfl1 with the co-repressor MoCyc8 to negative regulation of hyphal growth-related genes.
We also generated transformants of Guy11 expressing the MoTUP1-S and 3×FLAG-MoSFL1 constructs. MoTup1 was found to interact with MoSfl1 in co-IP assays with the resulting transformant (Fig 6B). Treatment with exogenous cAMP also significantly reduced the interaction of MoSfl1 with MoTup1 (Fig 6B). These results indicate that stimulation of PKA activities by exogenous cAMP reduces the interaction of MoSfl1 with MoCyc8 and MoTup1.
To determine the function of MoTUP1, we generated the Motup1 deletion mutant in the wild-type strain Guy11. The resulting deletion mutant MTU35 (Table 1) was significantly reduced in growth rate and it rarely produced aerial hyphae and conidia (Fig 6C). Unlike the cpkA cpk2 mutant, the Motup1 mutant was normal in appressorium formation. One distinct phenotype of the Motup1 mutant was the production of swollen bodies in the subapical regions in hyphae grown on CM (Fig 6C), suggesting cell wall integrity defects. The phenotype differences between the Motup1 mutant and Mosfl1 or cpkA cpk2 mutant indicate that the MoCyc8-MoTup1 co-repressor is involved in regulating different sets of genes by interacting with transcription factors other than MoSfl1.
In S. cerevisiae, Sfl1 has two predicted consensus PKA phosphorylation sites S207 and S733 [43] that are conserved in MoSfl1 (S211 and S554) and its orthologs from other fungi (Fig 7A). PSITE analysis identified T441 as the only other consensus PKA phosphorylation site in MoSfl1. To determine their functions, the MoSFL1S211D, MoSFL1T441D, and MoSFL1S554D alleles were generated and transformed into the cpkA cpk2 mutant. Whereas MoSFL1S211D transformants grew faster, the MoSFL1T441D and MoSFL1S554D transformants had similar growth defects with the original cpkA cpk2 mutant (Fig 7B). Similar approaches were used to generate transformants of cpkA cpk2 expressing the MoSFL1S211A, MoSFL1T441A, and MoSFL1S554A alleles (Table 1). None of these S/T to A mutations had suppressive effects on the growth defect of cpkA cpk2 (Fig 7B). These results indicate that phosphorylation of MoSfl1 at S211 may play a critical role to release its inhibitory functions. However, the MoSFL1S211D transformant failed to form appressoria on hydrophobic surfaces (Fig 7C). Therefore, the S211D mutation in MoSFL1 could suppress the growth but not appressorium formation defect of cpkA cpk2.
To identify genes affected by deletion of both CPKA and CPK2, we conducted RNA-seq analysis with RNA isolated from hyphae collected from 2-day-old CM cultures. Considering the significant reduction in growth rate, it was surprising that only 451 genes were down-regulated in the cpkA cpk2 double mutant in comparison with the wild type (S3 Table). However, many of them are functionally important for growth, including several genes encoding ribosomal proteins (MGG_00546, MGG_03372, MGG_09927, MGG_01113, MGG_06571) and enzymes important for cell wall synthesis (MGG_00592, MGG_03208, MGG_07331, MGG_01575, and MGG_03883). When the promoter regions (1000-bp upstream of the start codon) of genes down-regulated in the double mutant were analyzed, 111 of them contain the putative HSF-binding element AGAA-n-TTCT (n≤20) [27]. Among them, 29 genes have more than one HSF-binding elements in their promoter regions. These results indicate that the putative MoSfl1-binding element is enriched among the genes significantly down-regulated in the double mutant.
The cpk1 cpk2 double mutant of F. graminearum also had severe growth defects [20]. Similar to the cpkA cpk2 mutant of M. oryzae, fast-growing sectors were often observed in V8 cultures of cpk1 cpk2 mutant that were older than 10 days. We isolated 30 suppressor strains that had similar growth rate with the wild type (Fig 8A). In infection assays with corn silks, suppressor strains of cpk1 cpk2 were still defective in plant infection (Fig 8B). When the FgSFL1 gene was amplified and sequenced, 29 of them had mutations in the open reading frame (ORF) (Fig 8C; S4 Table). Suppressor strain HS29 lacked mutations in the ORF of FgSFL1 although its phenotype was similar to that of other suppressor strains with mutations in FgSFL1, suggesting possible mutations at its interacting site on FgCYC8. The most common mutation is the C1717 to T (Q501 to stop codon) mutation that resulted in the truncation of C-terminal 91 amino acids. A total of 15 suppressor strains had this C1717T mutation. Therefore, truncation of the C-terminal region also suppressed the cpk1 cpk2 mutant in F. graminearum. These results indicate that the function of SFL1 orthologs in hyphal growth is well conserved in M. oryzae, F. graminearum, and possibly other filamentous ascomycetes.
Like many other filamentous ascomycetes, the rice blast fungus has two genes encoding catalytic subunits of PKA. Whereas the cpkA mutant is defective in appressorium formation and pathogenesis, deletion of CPK2 had no effects on plant infection and appressorium morphogenesis. Interestingly, unlike cpkA, the cpk2 mutant was slightly reduced in growth rate. Therefore, although CpkA plays a more critical role than Cpk2 for pathogenesis, CPK2 may be more important during vegetative growth. In F. graminearum and A. fumigatus, deletion of the CPK2 ortholog had no detectable phenotype [19, 20], which differs slightly from cpk2 in M. oryzae. However, the cpkA cpk2 mutant had more severe defects than the single mutants in growth and conidiation, which is similar to F. graminearum and A. fumigatus [19, 20]. Therefore, the overlapping functions of CpkA and Cpk2 orthologs during vegetative growth and asexual reproduction may be evolutionally conserved in filamentous ascomycetes.
Although cAMP signaling is known to be important for appressorium formation in a number of fungal pathogens, including Colletotrichum species [44, 45], no mutants deleted of both catalytic subunits have been reported in plant pathogenic fungi except in U. maydis. In U. maydis, the mutant deleted of both catalytic subunits was defective in plant infection but its defects in appressorium formation was not examined [21]. Our results showed that PKA activities are essential for appressorium formation in M. oryzae, which has not been previous reported in plant pathogens.
The cAMP-PKA pathway is responsible for surface recognition in M. oryzae. To our surprise, although no tip swelling or appressorium formation was observed, the majority of cpkA cpk2 germ tubes curled or rotated on hydrophobic surfaces. Therefore, germ tubes of the double mutant still responded to surface hydrophobicity although they were blocked in appressorium formation. Sensing of surface hardness and hydrophobicity likely involves mechanosensor proteins, which may trigger polarity disturbance and germ tube curling independent of cAMP signaling. Several putative mechanosensor genes have been shown to be up-regulated during appressorium formation [46]. It is puzzling that most cpkA cpk2 conidia failed to germinate on hydrophilic surfaces, and curling germ tubes were not observed in a few of them that germinated. In filamentous ascomycetes such as A. nidulans and C. trifolii, cAMP signaling is known to regulate conidium germination [44, 47]. However, the cpkA cpk2 mutant was normal in conidium germination on hydrophobic surfaces. Because surface attachment is a cue for stimulating conidium germination in M. oryzae, one likely explanation is that deletion of both CPKA and CPK2 may affect the attachment of conidia to hydrophilic surfaces.
The Pmk1 MAP kinase pathway is essential for appressorium formation in M. oryzae and other plant pathogens [9, 34]. The cpkA cpk2 mutant had an increased phosphorylation level of Pmk1 but was defective in appressorium formation. It is possible that PKA activity is required to release the suppressive effect of MoSfl1 on genes important for germ tube tip swelling and appressorium formation. Spontaneous suppressors of cpkA cpk2 produced melanized tips in aerial hyphae and appressoria on hydrophilic surfaces, which is similar to transformants expressing the dominant active RAS2 [48]. Therefore, releasing the repressor role of MoSfl1 or MoCyc8-MoTup1 in the cpkA cpk2 mutant in which Pmk1 is over-activated is sufficient to activate appressorium formation under non-conducive conditions. One likely explanation is that some genes required for tip deformation or appressorium formation are only expressed when MoSfl1 is phosphorylated by PKA although the essential role of Pmk1 in appressorium formation involve other downstream targets. Because deletion of MoSFL1 had no effect on appressorium formation [26], it will be interesting to determine the effects of expressing the dominant active MST7 allele in the Mosfl1 deletion mutant.
In M. oryzae, the mac1 mutant is known to produce spontaneous suppressors and some of them had suppressor mutations in the SUM1 gene [30]. In fact, instability of adenylate cyclase mutant and suppressor mutations in regulatory subunit genes are well characterized in S. cerevisiae, U. maydis, and other fungi [30, 49, 50]. However, to our knowledge, spontaneous suppressors of PKA mutants have not been reported in other fungi. In S. cerevisiae, the tpk1 tpk2 tpk3 triple mutant is not viable. It will be important to assay whether the mutants deleted of both catalytic subunits also produce spontaneous suppressors in U. maydis and A. fumigatus [19, 21]. Because the cpk1 cpk2 mutant of F. graminearum [20] was also found to be unstable and had mutations in FgSFL1 in 29 of the 30 suppressor strains sequenced, it is likely that SFL1 orthologs have a conserved role in the repression of genes important for growth and conidiation, at least in filamentous ascomycetes. For the one F. graminearum and two M. oryzae suppressor strains without mutations in the SFL1 ortholog, identification of the suppressor mutations by whole genome sequencing [51] and characterization of the corresponding genes in these mutants in the future will be helpful to better understand the cAMP-PKA pathway in filamentous fungi.
In yeast, SFL1 can function as either a transcriptional activator or repressor [23, 24]. In M. oryzae, deletion of MoSFL1 by itself did not affect vegetative growth but resulted in a reduction in virulence [26]. Phenotype characterization of the Mosfl1 mutant is suitable for characterizing its activator but not repressor functions. In this study, we showed that deletion or truncation of MoSFL1 could suppress the defects of cpkA cpk2 mutant in vegetative growth and appressorium formation but not plant infection and conidiation. Interestingly, the suppressor mutants with mutations in FgSFL1 were also recovered in vegetative growth but not pathogenicity in F. graminearum. Whereas M. oryzae forms melanized appressoria for plant penetration, F. graminearum produces infection cushions and hyphopodia [52, 53]. However, after plant penetration, both of them form invasive hyphae inside plant cells that are different from vegetative hyphae in hyphal morphology and possibly cell cycle regulation [52, 54, 55]. The cAMP-PKA pathway is important for plant infection in M. oryzae [34, 56] and F. graminearum [20, 53], possibly by regulating the growth of invasive hyphae after penetration in these two plant pathogenic fungi with different infection mechanisms. It is possible that MoSFL1 plays an activator role in regulating genes important for invasive growth but negatively regulates genes important for vegetative growth in M. oryzae. However, it is more likely that the cAMP-PKA pathway regulates genes important for plant penetration and invasive growth via other transcription factors.
In yeast, Sfl1 inhibits the transcription of its target genes by interacting with the Cyc8-Tup1 co-repressor [27]. However, it is not clear which region of Sfl1 interacts with Cyc8 or Tup1. Our data clearly showed that the C-terminal 93 amino acids of MoSfl1 is essential for its interaction with Cyc8 (Fig 9). This C-terminal region of MoSfl1 is well conserved in its orthologs from filamentous ascomycetes. In F. graminearum, 15 of the 29 suppressor strains of cpk1 cpk2 mutant had the nonsense mutation at Q501 (S4 Table) in FgSFL1 resulting in the truncation of its C-terminal region. Other 12 suppressor strains had either nonsense or frameshift mutations upstream from Q501. These results indicate that the C-terminal region of Sfl1 orthologs likely plays a conserved role in regulating the expression of genes important for hyphal growth via its association with the Cyc8-Tup1 co-repressor complex. The difference between MoSfl1 and yeast Sfl1 in the C-terminal region may be directly related to the importance of PKA activities in hyphal growth in M. oryzae, F. graminearum, and possibly other filamentous fungi.
Besides the Cyc1-Tup1 co-repressor, Sfl1 also interacts with the mediator proteins Ssn2, Ssn8, Sin4, and Rox3 in S. cerevisiae [23]. Although their orthologs are conserved in M. oryzae, none of them was identified by affinity purification. One possibility is that phosphorylation by PKA is necessary for MoSfl1 to interact with these mediator components but the cpkA cpk2 mutant was used to identify proteins that differentially interacted with MoSfl1 and MoSfl1ΔCT and responsible the suppression of PKA deficiency. Nevertheless, it is also possible that their interactions with MoSfl1 is mediated by the Cyc8-Tup1 complex, which may be too dynamic or transient in M. oryzae. Among the MoSfl1-interacting proteins identified by affinity purification, MGG_01588 and MGG_06958 are orthologous to BMH1 and SSA1, respectively that interact with Sfl1 in S. cerevisiae. However, several MoSfl1-interacting proteins in M. oryzae (Table 5) such as the putative pathway-specific nitrogen regulator MGG_03286 are unique to filamentous fungi. It will be important to determine their functions in hyphal growth and pathogenesis in M. oryzae.
In S. cerevisiae, Sfl1 is a substrate of Tpk2 [25, 57]. However, it is not clear whether S207D mutation (= S211 of MoSfl1) is sufficient to suppress the growth defect of inviable tpk1 tpk2 tpk3 PKA-deficient mutant [58]. In M. oryzae, expression of the MoSFL1S211D allele in the cpkA cpk2 mutant rescued its growth but not appressorium formation defects. It is possible that phosphorylation of MoSfl1 by PKA disrupts the interaction of MoSfl1 with the Cyc8-Tup1 co-repressor, which in turn activates the expression of genes important for hyphal growth (Fig 9). Residue S211 of MoSfl1 is well conserved in its orthologs from other filamentous ascomycetes, suggesting that its phosphorylation by PKA likely has a conserved role in the regulation of hyphal growth by the cAMP-PKA pathway. Considering the fact that CpkA and Cpk2 are highly similar in the kinase domain and they have overlapping functions in hyphal growth, it is possible that both of them can phosphorylate MoSfl1 at S211. Nevertheless, MoSfl1 may be phosphorylated by CpkA and Cpk2 at different amino acid residues. Therefore, it will be important to characterize the phosphorylation sites of MoSfl1 in the cpkA and/or cpk2 deletion mutants.
All the M. oryzae strains used in this study (Table 1) were cultured on oatmeal agar (OTA) or complete medium (CM) plates at 25°C and stored on desiccated Whatman #1 filter paper at -20°C [59]. Protoplast preparation and PEG-mediated transformation were performed as described [60]. Transformants were selected with 250 μg/ml hygromycin B (CalBiochem), 250 μg/ml geneticin G418 (Sigma), or 200 μg/ml zeocin (Invitrogen) in the top agar. Growth rate and conidiation were assayed with OTA cultures as described [60, 61].
To generate the CPK2 gene replacement construct by double-joint PCR, its 1.2-kb upstream and downstream flanking sequences of CPK2 were amplified with primer pairs 1F/ 2R and 3F/4R (S5 Table), respectively (S1 Fig). The hph cassette was amplified with primers Hyg/F and Hyg/R from pCX63 [62]. The resulting products of double-joint PCR were transformed into protoplasts of the wild-type strain Guy11. Putative cpk2 mutants were screened by PCR with primers 5F and 6R and further confirmed by Southern blot analyses with its downstream flanking sequence as the probe. Vegetative hyphae harvested from two-day-old CM cultures were used for DNA and protein isolation as described [63].
The same strategy was used to generate the CPKA gene replacement construct. The 1.2-kb upstream and downstream flanking sequences of CPKA were amplified with primer pairs A1F/A2R and A3F/A4R, respectively (S1 Fig). The G418 cassette was amplified with primers G418/F and G418/R from pFl7. The products of double-joint PCR were transformed into protoplasts of the cpk2 mutant to generate the cpkA cpk2 mutant.
Conidia were harvested from 10-day-old OTA cultures and resuspended to 5×104 conidia/ml in sterile water. For appressorium formation assays, 50 μl droplets of conidium suspensions were placed on glass cover slips (Fisher Scientific) or GelBond membranes (Cambrex) and incubated at 25°C for 24 h as described [13, 64]. To assay its stimulatory effects, cAMP was added to the final concentration of 10 μM to conidium suspensions [65]. For infection assays, conidia were resuspended to 5×105 conidia/ml in 0.25% gelatin. Two-week-old seedlings of rice cultivar CO-39 were used for spray or injection infection assays as described [66, 67]. Lesion formation was examined 7 days post inoculation (dpi).
Vegetative hyphae were harvested from 2-day-old CM cultures and used for protein extraction as described [68]. Total proteins (approximately 20 mg) were separated on a 12.5% SDS-PAGE gel and transferred to nitrocellulose membranes for western blot analysis [62]. Expression and phosphorylation of Pmk1 and Mps1 were detected with the PhophoPlus p44/42 MAP kinase antibody kits (Cell Signaling Technology) following the manufacturer’s instructions.
Intracellular cAMP was extracted from vegetative hyphae harvested from two-day-old CM cultures as described [69] and detected with the cAMP enzyme immunoassay (EIA) system (Amersham Pharmacia Biotech) following the manufacturer’s instructions.
The yeast gap repair approach was used to generate the S-tag and 3×FLAG fusion constructs [68]. To generate the SUM1-, TUP1-, and CYC8-S-tag constructs, each gene was amplified and cloned into vector pXY203 [66, 70]. MoSFL1 was cloned into vector pFL6 to generate the 3×FLAG-MoSFL1 construct. CPKA and CPK2 were cloned into vector pFL7 [70] to generate the CPKA-3xFLAG and CPK2-3×FLAG fusion constructs. All of the resulting S-tag and 3×FLAG fusion constructs were confirmed by sequencing analysis and transformed into Guy11 or the cpkA cpk2 mutant in pairs. Total proteins were isolated from the resulting transformants and incubated with the anti-S-Tag Antibody Agarose beads (Bethyl Laboratories). Proteins bound to anti-S-tag agarose were eluted and used for western blot analysis [13, 31]. The presence of related fusion proteins was detected with the anti-FLAG (Sigma-Aldrich) or anti-S (Abcam) antibody as described [31].
Fast-growing sectors of the cpkA cpk2 mutant were transferred with sterile toothpicks to fresh oatmeal agar plates. After single-spore isolation, each subculture of spontaneous suppressors was assayed for defects in growth, conidiation, and plant infection [71, 72]. To identify suppressor mutations, all the candidate downstream target genes of PKA were amplified with primers listed in S5 Table and sequenced. Mutation sites were identified by sequence alignment with sequences of target genes in the reference genome [1] and their PCR products.
To generate the cpkA cpk2 Mosfl1 mutant, the upstream and downstream flanking sequences of MoSFL1 were amplified with primer pairs Sfl1ko1F/Sfl1ko2R and Sfl1ko3F/ Sfl1ko4R (S5 Table), respectively, and fused with the ble cassette amplified from pFL6 [73] by double-joint PCR [74]. To generate the cpkA cpk2 MoSFL1ΔCT mutant, the flanking sequences of MoSFL1 were amplified with primer pairs CCSko1F/CCSko2R (S5 Table) and Sfl1ko3F/ Sfl1ko4R (S5 Table). The resulting MoSFL1 and MoSFL1CT gene replacement PCR products were transformed into protoplasts of the cpkA cpk2 mutant. Putative Mosfl1 or MoSFL1CT mutants were screened by PCR analysis and verified for the deletion of MoSFL1 or its C-terminal region (496–588 aa).
The full length MoSFL1 and MoSFL1ΔCT fragments were amplified with primers listed in S5 Table and cloned into vector pFL6 by yeast gap repair [70]. The 3×FLAG-MoSFL1 and 3×FLAG-MoSFL1ΔCT fusion constructs were rescued from the resulting yeast transformants and transformed into the cpkA cpk2 mutant. Hyphae of the 3×FLAG-MoSFL1 and 3×FLAG-MoSFL1ΔCT transformants were homogenized with a glass beater at 4°C for protein extraction [31, 68]. Proteins eluted from anti-FLAG resins (Sigma-Aldrich) were digested with trypsin and the resulting tryptic peptides were analyzed with nanoflow liquid chromatography tandem mass spectrometry (MS) as described [31, 75–77]. Proteins were identified by searching MS data against NCBI non-redundant F. graminearum protein database with the SEQUEST™ algorithm [78]. At least three independent biological replicates were analyzed to identify proteins that interact with MoSfl1WT and MoSfl1ΔCT.
To generated the MoSFL1S211D allele, PCR fragments amplified with primer pairs MoSfl1-FL5F/MoSfl1-S211D1R and MoSfl1-S211D2F/MoSfl1-FL5R (S5 Table) were connected by overlapping PCR [79] and cloned into vector pFL5 [70] by yeast gap repair. The MoSFL1S211D construct was rescued from Trp+ yeast transformants and verified for the S211D mutation by sequencing analysis. Similar approaches were used to generate the MoSFL1S211A, MoSFL1T441D, MoSFL1T441A, MoSFL1S554D, and MoSFL1S554A constructs. All the MoSFL1 mutation alleles were transformed into the cpkA cpk2 deletion mutant. The resulting transformants were characterized for defects in growth, conidiation, appressorium formation, and plant infection as described [71].
|
10.1371/journal.pntd.0005288 | Emergence of Rare Species of Nontuberculous Mycobacteria as Potential Pathogens in Saudi Arabian Clinical Setting | Clinical relevance of nontuberculous mycobacteria (NTM) is increasing worldwide including in Saudi Arabia. A high species diversity of NTM’s has been noticed in a recent study. However, the identification in diagnostic laboratories is mostly limited to common species. The impact of NTM species diversity on clinical outcome is so far neglected in most of the clinical settings.
During April 2014 to September 2015, a nationwide collection of suspected NTM clinical isolates with clinical and demographical data were carried out. Primary identification was performed by commercial line probe assays. Isolates identified up to Mycobacterium species level by line probe assays only were included and subjected to sequencing of 16S rRNA, rpoB, hsp65 and 16S-23S ITS region genes. The sequence data were subjected to BLAST analysis in GenBank and Ez-Taxon databases. Male Saudi nationals were dominated in the study population and falling majorly into the 46–59 years age group. Pulmonary cases were 59.3% with a surprising clinical relevance of 75% based on American Thoracic Society guidelines. Among the 40.7% extra-pulmonary cases, 50% of them were skin infections. The identification revealed 16 species and all of them are reporting for the first time in Saudi Arabia. The major species obtained were Mycobacterium monascence (18.5%), M. cosmeticum (11.1%), M. kubicae (11.1%), M. duvalli (7.4%), M.terrae (7.4%) and M. triplex (7.4%). This is the first report on clinical relevance of M. kubicae, M. tusciae, M.yongonense, M. arupense and M.iranicum causing pulmonary disease and M. monascence, M. duvalli, M. perigrinum, M. insubricum, M. holsaticum and M. kyorinense causing various extra-pulmonary diseases in Saudi Arabia. Ascites caused by M. monascence and cecum infection by M. holsaticum were the rarest incidents.
To the first time in the country, clinical significance of various rare NTM’s are well explored and the finding warrants a new threat to the Saudi Arabian clinical settings.
| Nontuberculous mycobacteria (NTM) are ubiquitous in nature and they are opportunistic pathogens. In the last decade, infections caused by NTM’s increased—around the world in immune-suppressed and immune-competent individuals and Saudi Arabia is not an exception. Developments in diagnostic technologies increased the identification of several new or rare species of NTM’s. Indeed, the species diversity of NTM has a direct impact on clinical outcome and therapies. Saudi Arabian clinics so far report only the common species of NTM’s and rare species are mostly neglected due to the lack of proper infrastructure or ignorance. To the first time in the country, an exploration on the existence of clinically relevant rare NTM species was conducted on a nationwide level. The findings showed a huge diversity of rare NTM species causing both pulmonary and extra-pulmonary diseases. Clinical relevance of pulmonary infection based on American Thoracic Society guidelines was confirmed as an aggressive 75%, which is really alarming. Interestingly, 16 species of NTM’s were isolated in the study, and all of them are reporting for the first time in country. Overall finding shows Saudi Arabia faces serious threat from rare NTM species with high clinical significance and warrants the immediate need for more advanced infrastructure.
| In the last decade, the prevalence of pulmonary and extra-pulmonary diseases caused by nontuberculous mycobacteria (NTM) has been increased [1–7]. This elevation in case rates, whether it is a real emergence or due to the development of advanced diagnostic tools is still unclear. On the other hand, the elevation in immunosuppressive conditions including infectious or non-infectious diseases and therapies contribute considerably in this phenomenon. To date, more than 140 species of NTM’s have been described from different sources with varying pathogenicity and almost 50 species were identified in the last 8 years alone [8]. However, in literature only a small number of reports are available about the new species as their role in clinical microbiology is largely undetermined. Mostly, the species defined as “rare” will remain unrecognized or misidentified due to the lack of proper resources, lack of knowledge or ignorance [8]. The clinical characteristics of diseases caused by the rare or new NTM’s are still not fully established. The advancement in technologies such as genome sequencing, line probe assays, high performance liquid chromatography (HPLC) and matrix assisted laser desorption ionization time-of-flight (MALDI-TOF) to identify the NTM species increased the detection of rare and new NTM species. However, accessibility to such tools in resource poor settings is a major concern for timely identification. Thus, the species level identification is mostly neglected regardless of its importance in clinical outcome.
Following the global trend of NTM prevalence, Saudi Arabia also reports with an increasing numbers of NTM diseases [7, 9]. In 2013, Varghese et al. reported in the first national level study a highly diverse population of clinically relevant NTM’s with the potential of causing pulmonary and extra-pulmonary diseases [9]. Interestingly, a new species of pathogenic mycobacteria also has been identified from the country named M. riyadhense [10]. However, the diagnosis of NTM’s is mostly limited to the common species only in majority of the laboratories, because of the lack of infrastructure. Therefore, there is no data available on the existence of rare NTM species in the country so far.
To explore the diversity of rare NTM species with clinical relevance in the Saudi Arabian clinical setting, a prospective analysis on a nationwide isolate collection has been designed. Sequencing of 16S rRNA, rpoB, hsp65 and 16S-23S ITS region genes were carried out to identify the species. Clinical significance of pulmonary isolates in the study was determined by applying the criteria based on American Thoracic Society (ATS) guidelines for NTM pulmonary diseases [11]. The species diversity and clinical significance of each isolates have been evaluated.
This study has been conducted as part of the first national NTM surveillance survey of Saudi Arabia. The duration of collection was 18 months, from April 2014 to September 2015. All the suspected NTM isolates from different mycobacterial diagnostic laboratories were collected and transferred to the Mycobacteriology Research Section of King Faisal Specialist Hospital and Research Centre, Riyadh. The demographical and clinical data were collected by using standard data collection forms without keeping any patients identifiers. Pulmonary cases were defined as clinically relevant based on ATS guidelines [11]. Briefly, a minimum of two positive cultures from separate sputum samples or at least one positive culture from bronchial wash, lavage or one positive culture from trans-bronchial or other lung biopsy showing mycobacterial- histopathological features were considered as clinically relevant to define NTM pulmonary disease.
Isolates were maintained on Lowenstein Jensen slants and modified Middle Brook 7H9 medium (Becton Dickinson, USA). The genomic DNA was extracted by using the QIAamp DNA Mini kit (Qiagen, Germany). The primary screening to identify the NTM’s was carried out by commercially available line probe assay kit- Genotype MTBC (Hain Life science, Germany). The non-MTBC isolates were initially identified with Genotype Mycobacterium CM kit (Hain Life science, Nehren, Germany) and unidentified isolates were further tested with Genotype Mycobacterium AS kit (Hain Lifescience, Nehren Germany).
Isolates which were detected by the Genotype Mycobacterium AS assay up to genus level (Mycobacterium species) only were included in the study as “unidentified” species. This study has been reviewed and approved by the Office of Research Affairs in King Faisal Specialist Hospital and Research Centre, Riyadh, Saudi Arabia.
Sequencing assay was carried out by using the BigDye Terminator cycle sequencing chemistry kits (Applied Biosystems, CA, USA) in DNA analyzer 3730 (Applied Biosystems, CA, USA). The first attempt of identification was based on a 645-655bp hyper variable region of the 16S rRNA and a 342bp region of rpoB genes based on previously standardized protocol [12, 13]. Isolates which could not be identified by 16S rRNA and rpoB gene sequencing were subjected to further sequencing of highly conservative regions of two more genes hsp65 (439bp) and 16S-23S ITS region (480bp) according to previously validated primers [14–16].
The line probe assay test strips were scanned with Genoscan (Hain Lifescience, Germany) and the results were interpreted with the Blotrix software (Hain Lifescience, Germany). The sequence base calling and assembly were carried out by using Sequence Analysis software v5.3.1 (Applied Biosystems, USA) and Lasergene core suite 12 (DNA STAR, WI, USA) respectively. Assembled sequences were subjected to BLAST analysis in NCBI GenBank and EzTaxon (http://www.ezbiocloud.net/identify; 16S rRNA Based Database) online data bases [17]. A stringent similarity index of ≥99–100% was kept with the type strain in GenBank and EzTaxon. Statistical data analysis was carried out by using SPSS V20.0 software package (IBM, USA).
During the study period, 510 suspected NTM clinical isolates were collected and subjected to line probe assay identification. Of the total, 27 isolates met the inclusion criteria and enrolled for further analysis. Demographical summary of the study subjects showed 22 (81.5%) of the enrolled cases were Saudi nationals with a male (77.8%) gender domination. Interestingly, the age group of the study subjects showed a predominance of 46–59 years (48.2%). Seven cases out of 27 had a previous history of tuberculosis and 3 were reactive to HIV antigens. The major comorbidities noticed among the study subjects were rheumatoid arthritis (18.5%), malignancies (18.5%) and diabetes (14.8%). Considerable percentage of Chronic Obstructive Pulmonary Disease (COPD), asthma and bronchiectasis also were observed (Table 1).
Analysis of 16S rRNA, rpoB, hsp65 and 16-23S ITS region genes, showed an extreme diversity of NTM species distributed to 16 species. The majorly detected species were M. monascence (18.5%), M. cosmeticum (11.1%), M.kubicae (11.1%), M. duvalii (7.4%), M. triplex (7.4%) and M. terrae (7.4%). Rest of the 10 species were reported with one case each (3.7%) (Fig 1).
The major site of infection observed in the study was pulmonary (59.3%). Clinical relevance of pulmonary isolates based on the ATS guidelines was dominating (75%). Of the 16 pulmonary cases, four cases did not qualify the clinical relevance guidelines and thus considered as colonization. Clinically relevant diseases were caused by M. arupense, M. cosmeticum, M. iranicum, M. kubicae, M. monascence, M. novocatrense, M. tusciae and M. yongonense (Table 1).
On the other hand, extra-pulmonary involvement was found with 11 (40.7%) cases. Among extra-pulmonary cases, skin (45.5%) was the most affected site of infection followed by lymphnode (27.3%). M. insubricum, M. perigrinum and M. marinum were found exclusively causing granuloma or sepsis. Interestingly, all these five skin infections were observed among non-Saudi patients. Cecum infection by M. holsaticum and ascites caused by M. monascence were the rarest incidents in this study (Table 1).
For the first time in Saudi Arabia, the existence of rare NTM species has been explored on a nationwide collection of clinical isolates. The findings showed a strong presence of clinically relevant NTM rare species in the Saudi Arabian clinical settings. The species diversity of rare NTM’s causing both pulmonary and extra-pulmonary diseases was huge (16 species). To date, all of these 16 species with clinical relevance are reporting for the first time in the country.
Demographical analysis showed predominance of male Saudi nationals and mainly the age group 46–59 years and a similar domination had been noticed in a recent nationwide study of NTM prevalence [9]. Clinical data showed various comorbidities existed among the study group. Rheumatoid arthritis and various malignancies were the major problems followed by diabetes. There were no studies so far analyzed the reasons behind the predominance of male Saudi nationals towards the NTM disease. Perhaps, there were some speculations like higher rate of consanguinity existed in the community which leads to several genetic susceptibility diseases, increasing rate of immunosuppressive therapies and various malignancies in the geographical region [18–20]. Indeed, the confounding factors for this predominance need further detailed scientific exploration.
In this study, pulmonary diseases caused by rare NTM species were predominant with higher clinical relevance (75%). Of the 16 identified species, 11 species except M. insubricum, M. kyorinense, M. holsaticum, M. perigrinum and M. marinum were isolated from respiratory samples. Of the 11 species, except M. duvalii and M. terrae all the others reported with clinically relevant diseases. Moreover, most of these species are isolated very rarely from clinical specimens with relevance so far around the world [21]. Interestingly, to date, M. tusciae, M. yongonense, M. novocastrens and M. monascence pulmonary diseases were reported in hardly 2–3 publications. Therefore, these findings are important as it shows and confirms the growing problems with rare NTM species in Saudi Arabia as well for the first time in the Gulf Cooperation Council (GCC) states also [22] [21, 23, 24]. The higher clinical relevance established after consulting the ATS guidelines shows the increasing potential of NTM respiratory diseases in the country. Supportively, NTM respiratory diseases caused by various species have been observed in a recent nationwide study in the country [9]. In the current study majority of the isolation was from sputum samples, except six bronchial washes. The frequency of isolation from sputum was peaked up to 7 occasions for a case of M. cosmeticum. The higher frequency shows the increasing potential of NTM’s as en establishing pathogen in the clinical settings. Most of the NTM species isolated in the current study causing pulmonary diseases are not only rare in Saudi Arabia but also around the globe. In concordance to the current findings, previous global studies showed an increasing prevalence of NTM’s and particularly the domination of pulmonary isolates. The existence of numerous rare or new species was observed in most of the large level studies [5–7, 21]. The increased clinical relevance is a warranting key to be vigilant on the pathogenicity and potential of NTM’s to cause confirmed diseases rather than colonization.
In the current study, 11 cases of extra-pulmonary diseases were observed with a predominance of skin infections (50%) caused by M. cosmeticum, M. insubricum, M. perigrinum, M. marinum and M. terrae. The M. perigrinum and M. marinum were isolated from two immigrant patients from the Western coast of the country, where the major fishing harbors located with considerable number of foreign fishermen. The skin infection caused by M. marinum and M. perigrinum among people who work in fisheries or swimming pool maintenance is generally reported elsewhere [25, 26]. The rarest cases in the study were the ascites caused by M. monascence and cecum infection caused by M. holsaticum. To our knowledge, this might be the first cases of the same reporting globally. On the other hand, lymphadenitis caused by M. kyorinense and M. triplex are rare manifestations which have been reported in only two or three cases so far globally and the current study also found a case of each [27, 28]. The infections caused by M. duvalii and M. monascence in patients undergoing peritoneal dialysis are reported for the first time in Saudi Arabia and in GCC nations. Such cases are rarely reported around the world [8].
As the first report on the existence of rare NTM species in Saudi Arabia, the findings showed an alarming diversity of clinically relevant NTM’s causing both pulmonary and extra-pulmonary diseases. The diagnostic facilities in the country requires more advanced infrastructure to identify the rare species on time, as it influence the final clinical outcome and treatment adversely.
The following 16S rRNA gene sequences have been deposited in GenBank/DDBJ/EMBL data bases. M. monacense (KY287007), M. iranicum (KY287008), M. kubicae (KY287009), M. cosmeticum (KY287010), M. duvalii (KY287011), M. terrae (KY287012), M. arupense (KY287013) and M. novocastrense (KY287014).
|
10.1371/journal.pntd.0004362 | Efficacy, Safety, and Dose of Pafuramidine, a New Oral Drug for Treatment of First Stage Sleeping Sickness, in a Phase 2a Clinical Study and Phase 2b Randomized Clinical Studies | Sleeping sickness (human African trypanosomiasis [HAT]) is caused by protozoan parasites and characterized by a chronic progressive course, which may last up to several years before death. We conducted two Phase 2 studies to determine the efficacy and safety of oral pafuramidine in African patients with first stage HAT.
The Phase 2a study was an open-label, non-controlled, proof-of-concept study where 32 patients were treated with 100 mg of pafuramidine orally twice a day (BID) for 5 days at two trypanosomiasis reference centers (Angola and the Democratic Republic of the Congo [DRC]) between August 2001 and November 2004. The Phase 2b study compared pafuramidine in 41 patients versus standard pentamidine therapy in 40 patients. The Phase 2b study was open-label, parallel-group, controlled, randomized, and conducted at two sites in the DRC between April 2003 and February 2007. The Phase 2b study was then amended to add an open-label sequence (Phase 2b-2), where 30 patients received pafuramidine for 10 days. The primary efficacy endpoint was parasitologic cure at 24 hours (Phase 2a) or 3 months (Phase 2b) after treatment completion. The primary safety outcome was the rate of occurrence of World Health Organization Toxicity Scale Grade 3 or higher adverse events. All subjects provided written informed consent.
Pafuramidine for the treatment of first stage HAT was comparable in efficacy to pentamidine after 10 days of dosing. The cure rates 3 months post-treatment were 79% in the 5-day pafuramidine, 100% in the 7-day pentamidine, and 93% in the 10-day pafuramidine groups. In Phase 2b, the percentage of patients with at least 1 treatment-emergent adverse event was notably higher after pentamidine treatment (93%) than pafuramidine treatment for 5 days (25%) and 10 days (57%). These results support continuation of the development program for pafuramidine into Phase 3.
| Sleeping sickness (human African trypanosomiasis [HAT]) is caused by parasites, and has a chronic progressive course that may last from several months to several years before death occurs. The present studies were done to assess the effectiveness and safety of oral pafuramidine versus intramuscular pentamidine (the standard treatment), in patients with first stage HAT. The results indicated that, several months after treatment, pafuramidine administered for 10 days was as effective as pentamidine administered for 7 days, and it had a better safety profile than pentamidine. With further study, pafuramidine could be a promising alternative for patients with first stage HAT. In addition, the design of the studies can be used a guide for future studies for identification and delivery of treatment to affected individuals in rural Africa.
| Sleeping sickness (human African trypanosomiasis [HAT]) is a neglected tropical disease with limited treatment options that currently requires parenteral administration. It is caused by the protozoan parasites Trypanosoma brucei (T.b.) gambiense (the West African form of the disease) and Trypanosoma brucei rhodesiense (the East African form of the disease). T.b. gambiense is found in 24 countries in west and central Africa and currently accounts for over 98% of reported cases of sleeping sickness [1]. Sleeping sickness due to T.b. gambiense is characterized by a chronic progressive course, which may last from several months to several years before death occurs. The disease has two defined stages. The first stage is characterized by trypanosomes in the hemolymphatic system, which multiply in subcutaneous tissues, blood, and lymph. First stage symptoms entail bouts of fever, headaches, joint pains and itching. In the first stage, a person can be infected for months or even years without major signs or symptoms of the disease. When more evident symptoms emerge, the patient is often already in an advanced disease stage where the central nervous system is affected (second stage). The second (or neurological) stage begins once parasites penetrate the central nervous system, where their presence initiates deterioration in neurological function, including disruption of sleep/wake patterns that lend the name “sleeping sickness” to this stage of the disease [1].
After resurgence of the disease in the 1990s, the number of annual cases has subsequently dropped to less than 10,000 in recent years [2, 3]. Although the global incidence of HAT seems to be declining, several non-governmental organizations have cautioned that HAT still remains a “hidden epidemic” in regions impacted by civil war, such as the northeastern Democratic Republic of the Congo (DRC), southern Sudan, and the Central African Republic [4].
The socioeconomic burden on a household with an individual with HAT is high; between 1.5 and 10 months of household income can be lost, even when the diagnostics and antitrypanosomal drugs are provided free of charge. The extraordinary burden that HAT places on affected households and communities is often not very visible in national or regional health data because of its focal nature, and the disease often falls very heavily on a few locations [4].
The majority of current HAT research is focused on the second stage of the disease, which requires drugs that can cross the blood-brain barrier. The organo-arsenic compound melarsoprol has until very recently been the most widely used drug for treatment of second stage HAT, but because of its toxicity (encephalopathic syndrome, associated with 5–10% mortality), it is being progressively replaced by nifurtimox-eflornithine combination therapy [5]. The nifurtimox-eflornithine combination was placed on the World Health Organization (WHO) list of essential medicines and has been increasingly used [6, 7]. In addition, several new molecules have also shown promise for the treatment of second stage HAT. The most developmentally advanced of these compounds is fexinidazole, which has completed a Phase 1 clinical study and is now in Phase II/III evaluation [8].Two diamidine analogues [9, 10] and a benzoxaborole compound [11] have also exhibited promising activity (short in vitro time to kill and no cross-resistance) in animal models of second stage HAT.
An effective, safe drug for stage 1 HAT that can be easily administered in rural African settings is critically needed. There is no vaccine for T.b. gambiense HAT and only two drugs are approved for treatment of stage 1 disease: pentamidine (in the field only used for T.b. gambiense) and suramin (only for only for T.b. rhodesiense). In expatriate patients, pentamidine is currently the treatment of choice since is it generally well tolerated, whereas suramin can cause undesirable effects in the urinary tract in addition to allergic reactions [1]. Pentamidine is administered by the intramuscular route and has a reported treatment failure rate after a course of five injections of approximately 7% [12, 13, 14]. Despite this encouraging efficacy profile, treatment with pentamidine has limitations. It requires injection, which hampers its use in rural treatment facilities. Though adverse reactions are usually reversible and persistent manifestation of its most serious long-term consequence, diabetes, is rare, a high frequency of adverse events, including hypotension, nephrotoxic effects, leukopenia, and hypo- and hyperglycemia, has been noted [15, 16].
Pafuramidine (DB289) is the orally available dimethoxime prodrug of DB75 (furamidine), a novel diphenylfuran diamidine shown to be active in vitro against African trypanosomes and in animal models for trypanosomiasis [17, 18]. Pafuramidine is potentially a significant improvement over pentamidine, the drug currently used to treat first stage HAT. Its oral formulation greatly facilitates administration under challenging field conditions and makes pafuramidine readily available not only in sleeping sickness centers, but also public health facilities. Further, high doses of pafuramidine have been remarkably well tolerated in animal models of trypanosomiasis [19] and Pneumocystis jiroveci (a fungal infection of the lungs, formerly called Pneumocystis carinii, or PCP pneumonia) [20]. In addition, the prodrug pafuramidine and its active metabolite (DB75) have shown increased efficacy compared to pentamidine in animal models of T.b. rhodesiense infection [17, 19].
Prior to the initiation of the current Phase 2a study, pafuramidine had been successfully administered to healthy volunteers in both single-dose and multiple-dose studies, to evaluate the safety of pafuramidine and the pharmacokinetics (PK) of pafuramidine and DB75 [21]. Overall, multiple-dose treatment was well tolerated up to the maximum dose of 100 mg twice a day (BID), and was also well tolerated in the single-dose study up to 600 mg, although there was no substantially increased area under the plasma drug concentration versus time curve at doses above 100 mg.
The objectives of the present Phase 2 studies were to assess, for the first time, the efficacy, safety, and dosage of pafuramidine (Phase 2a), and to compare the efficacy, safety, and dosage of oral pafuramidine versus intramuscular pentamidine (Phases 2b and 2b-2) for treatment of first stage T.b. gambiense sleeping sickness. Pafuramidine may offer a potentially significant improvement over pentamidine, since it is orally administered and may be better tolerated.
All subjects provided written informed consent. This will certify that the Institutional Review Boards (IRBs) at the University of North Carolina at Chapel Hill, administered by the Office of Human Research Ethics, are organized and operate according to applicable laws and regulations governing research involving human subjects. These include, when applicable, statutes of the State of North Carolina and regulations of the Food and Drug Administration (21 CFR 50 and 56) and the Department of Health and Human Services [45 CFR 46 (the "Common Rule") and 45 CFR 164 (the Health Insurance Portability and Accountability Act, HIPPAA]. In addition, the IRBs conform, when applicable, to Good Clinical Practice (GCP) guidelines of the International Conference of Harmonization (ICH), to the extent these do not contradict DHHS and FDA regulations. The University of North Carolina at Chapel Hill holds a Federalwide Assurance, FWA 4801, approved by the federal Office for Human Research Protections (OHRP). These studies were approved by the following independent ethics committees: Ethikkommission beider Basel, EKBB, Comité de Ética Republica de Angola, and Comité Éthique République Démocratique du Congo. IRB# 01-PATH/LAB-308. International Protocol #289-C-006.
The Phase 2a study was a multi-center, multi-country, open-label, non-controlled, proof-of-concept study to assess the efficacy and safety of pafuramidine in 32 patients with first stage T.b. gambiense sleeping sickness. Patients were treated with 100 mg of pafuramidine orally BID for 5 days, and were hospitalized for a total of 12 days including a 6-day post-dose observation period. This study was conducted at two trypanosomiasis reference centers: one in Viana, Angola, and one in Maluku, Democratic Republic of the Congo (DRC) from 31 August 2001 (first patient enrolled) to 28 November 2004 (last patient follow-up completed). This study was approved by the following independent ethics committees: Ethikkommission beider Basel, EKBB, Comité de Ética Republica de Angola, and Comité Éthique République Démocratique du Congo.
The Phase 2b study was an open-label, parallel group, controlled, randomized trial to compare the efficacy and safety of pafuramidine with standard pentamidine treatment in 81 patients with first stage T.b. gambiense sleeping sickness. Patients were randomized (1:1) to either pafuramidine 100 mg BID administered orally for 5 days (n = 41) or pentamidine intramuscular injections (4 mg/kg QD) for 7 days (n = 40). This study was conducted at two sites: one trypanosomiasis reference center (Maluku, DRC) and one hospital (Vanga, DRC) from 01 April 2003 (first patient enrolled) to 08 February 2007 (last patient follow-up completed). This study also included a substudy/subset of patients enrolled at selected sites to assess the incidence and severity of laboratory anomalies (including aspartate aminotransferase [AST]/alanine aminotransferase [ALT] and glycemia) and incidence of electrocardiogram (ECG) anomalies.
The Phase 2b protocol was later amended to add an open-label study sequence (Phase 2b-2), in which an additional 30 patients were recruited at the same sites to confirm safety and efficacy of prolonged pafuramidine treatment. These patients received 100 mg of pafuramidine BID for 10 days with a total hospitalization time of 14 days.
The Phase 2b and 2b-2 studies were approved by the Ethikkommission beider Basel, EKBB, Committee on the Protection of the Rights of Human Patients, University of North Carolina Chapel Hill, and Comité Ethique Republique Democratique du Congo.
In all studies, patients were stratified by site to ensure similar numbers of enrolled patients at each site. The flowchart for all the Phase 2 studies is depicted in Fig 1.
All study amendments were approved by the ethical committees and IRBs previously noted. The Phase 2a amendment 1 included: 1) changed inclusion criterion for white blood cell (WBC) count from 20 to 5 cells/mm-3 to reduce the possibility of enrolling undetected late stage patients; 2) added a minimal weight of 40 kg to maintain consistency with Phase 1 trials; 3) added exclusion due to traumatic lumbar puncture to reduce the chance of incorrect staging of the disease at diagnosis; and 4) changed and refined the definition of treatment failures, based on the improved diagnostic tools to be used in the trial.
Phase 2b amendment 1 added stricter rules for opening non-hospital-based sites for enrolment, by adding a “delay” of treating 40 patients. Phase 2b recruitment was subsequently terminated before opening accrual to rural sites and protocol modifications were implemented after five treatment failures were observed directly after treatment in the 5-day pafuramidine group in Phase 2a. After examination of the PK properties of pafuramidine, in particular, the lack of proportional conversion of DB289 to DB75 at therapeutic doses [22, 23], we decided against using a higher dose to improve efficacy for the modified protocol and instead increased pafuramidine treatment to 10 days at the same dose (100 mg BID) (Phase 2b-2) (Fig 2). Phase 2b amendment 3 added additional testing for parasites in lymph and circulating blood to reduce the potential for false-negative tests provides a chronological representation of the conduct of each study and illustrates the time points at which decisions were made, based on preliminary data, to proceed to the next study in the development program.
Male and female subjects were eligible to participate if they had first stage T.b. gambiense sleeping sickness, documented by the presence of parasites in the blood or lymph and their absence in the cerebrospinal fluid (CSF), confirmed by <5 mm-3 WBCs detected in the CSF by microscopic examination. In the Phase 2a study, patients were 16 years or older, with a minimal weight of 45 kg, and in the Phase 2b study, patients were 15 to 50 years of age with a minimal weight of 35 kg. Women of childbearing potential were included if they were neither lactating nor pregnant, and were instructed to abstain from sexual intercourse from the day of consent until the end of the in-hospital observation period.
Key criteria for exclusion for all patients included 1) late stage T.b. gambiense infection; 2) active clinically relevant medical conditions that in the investigator’s opinion may have jeopardized patient safety or interfere with participation in the study (eg, significant liver diseases, chronic pulmonary diseases, significant cardiovascular diseases, diabetes, thyroid diseases, gout, infection including acquired immune deficiency syndrome, central nervous system trauma or seizure disorders); 3) traumatic lumbar puncture (red blood cells visible in CSF); 4) clinically significant abnormal laboratory values at screening; 5) score of less than 9 on the Glasgow Coma Scale; and 6) previous treatment for HAT.
Patients in all studies provided written informed consent. Patients in the Phase 2a study were screened either by mobile diagnostic teams or in the treatment centers. Patients with positive results at the mobile units were referred to the study sites for repeat testing. Patients in the Phase 2b and Phase 2b-2 studies were screened only in the treatment centers. Screening for T.b. gambiense was done using the card agglutination test for trypanosomes [24, 25]. All patients were tested for malaria and filaria in thick blood smears and, if indicated, malaria treatment was given before study enrolment; filariasis therapy was administered after study treatment.
All patients who had palpable lymph nodes underwent puncture at the screening or baseline visit. The aspirate was assessed microscopically for trypanosomes and if the result was negative, a blood sample was examined by hematocrit centrifugation and miniature anion exchange centrifugation technique. Lumbar puncture was performed in all positive cases detected by either method and the disease stage was determined by microscopic examination of CSF for trypanosomes and by WBC counts. Other screening and baseline evaluations included demographic and medical history, concomitant medications, vital signs, and physical examination including the Coma Scale. In all studies, baseline ECGs were performed and evaluated by a cardiologist to confirm patient inclusion.
Clinical supplies of pafuramidine were provided to the sites in bottles (50 gel caps of 100 mg) labelled to indicate study drug, strength, expiration date, protocol number, and other information according to local regulations. Pentamidine was provided locally as pentamidine isethionate for injection in single-dose vials at 200 mg per vial. Study drugs were stored at ambient temperature.
During the treatment periods in all studies, routine safety assessments included examinations for possible treatment-emergent adverse events, physical examinations, hematology tests (hemoglobin and leukocyte count), and chemistry profile (serum glucose, creatinine, AST, ALT, total bilirubin, urea, and C reactive protein).
In the Phase 2a study, patients were administered 100 mg of pafuramidine orally BID (morning and evening) for 5 days. In addition to the safety assessments previously mentioned, prothrombin time was also evaluated. For PK analyses, blood plasma levels were collected just before dosing and at several time points after the last dose. The 6-day post-dose observation period included additional ECGs after the last dose of study drug on Day 6; lymph node puncture and lumbar puncture on Day 7 (24 hours after last treatment, which is the primary endpoint); plasma samples for PK analyses; and a CSF sample for PK analysis on Day 7. Follow-up visits to evaluate treatment outcomes continued for 24 months, with lymph node punctures at Months 3, 6, 12, and 24 and lumbar punctures at Months 12 and 24.
In the Phase 2b study, patients received either pafuramidine 100 mg administered orally BID for 5 days or pentamidine intramuscular injections (4 mg/kg, once a day [QD]) for 7 days. Patients in both treatment groups stayed in the hospital for 7 days. In addition to the previously mentioned routine safety assessments, female patients in this study underwent pregnancy tests at baseline and end of treatment. A lymph node and lumbar puncture were also done at Day 7 (end of treatment). For patients in the substudy, blood samples for PK analyses were taken after the last dose on Day 6 and 7; ECGs were repeated on Days 1, 3, 7, and 9; and hematology and chemistry tests were repeated at Day 9. Follow-up visits to evaluate treatment outcomes continued for a period of 24 months, with lymph node, lumbar punctures, and a CSF sample at Month 3, 6, 12, and 24.
As part of Phase 2 amendment 2, patients received 100 mg of pafuramidine BID for 10 days with a total hospitalization time of 14 days. Study procedures in Phase 2b-2 were identical to Phase 2b, except the post-treatment procedures were done at Days 12 and 14. Follow-up visits to evaluate treatment outcomes also continued for a period of 24 months, with lymph node and lumbar punctures and a CSF sample as noted above.
For the Phase 2a study, the sample size of 30 evaluable patients was considered to be sufficient to meet the primary objective because this size was comparable to a typical Phase 1 pilot study.
Based on a non-inferiority approach, calculation of the Phase 2b sample size assumed a 2% relapse rate in the pentamidine control group and a maximally accepted relapse rate of 10% in the pafuramidine treatment group. To achieve a power of 90% to refute the null hypothesis, in the event that the relapse rate of the new drug was equal to that of the standard (also 2%), 147 evaluable patients in each group were planned with a target enrolment of 175 patients in each group. A dropout rate of 15% was expected by the first follow-up examination at 3 months post-treatment.
The decision to continue the study in Phase 2b-2 with 30 evaluable patients was made based on assumption of a set rate of treatment failures (parasite-positive analyses 24 hours after treatment) and a binomial distribution for the number of relapses, and calculation of the probability to have N or more relapses. It was appropriate to make the nominal significance level smaller than 0.05 in order to keep the overall Type 1 error rate (the probability to incorrectly discard the treatment tested as not effective) at 0.05 [28]. For example, with a sample size of 30 and an assumed relapse rate of 1%, occurrence of 2 or more relapses had a 4% probability.
No interim analysis was planned or conducted. There was no formal stopping rule for the Phase 2a and 2b trials. In case of accumulation of unexpected severe adverse events, the study director and the sponsor's medical director could have decided, after consultation with the advisory board chairperson, the investigator, and the country coordinator of the national HAT programs of Angola (ICCT), whether the trial needed to be stopped. In case of a major protocol violation by the investigator, the sponsor could have stopped the trial.
For Phase 2b-2, the decision to continue the trial was based on accepting (or assuming) a set rate of refractory cases (ie, a patient was parasite positive 24 hours after treatment) and a binomial distribution for the number of relapses. The probability to have N or more refractory cases in 30 patients can be seen in Table 1.
In all studies, patients were assigned a patient number in the order in which they enrolled, starting with Patient -001. In Phase 2a, there was no randomization since the study was open label. In Phase 2b, patients were randomized (1:1) in blocks of 10 in the order in which they were enrolled, stratified by clinical site, according to a randomization schedule prepared at Immtech International. Each study site was provided with series of individual envelopes each containing a card with the treatment assignment for 1 patient and a control number. After a patient signed the informed consent and inclusion/exclusion criteria were confirmed, the investigator opened the next envelope in the randomization list to obtain treatment assignment for that patient, then transferred the control number to the patient’s case report form.
Patients were identified on the case report forms by the patient number, initials, study center, and study identification numbers. The investigators kept a separate confidential enrolment log that matched identifying codes with the patients’ names and residences. The process was carefully monitored during the trial and envelopes were designed in a way that did not allow tampering or identification of the documentation inside.
There was no blinding in the studies due to the different routes of administration. Phase 2a and Phase 2b-2 were non-controlled studies.
For the Phase 2a study, the parasitological cure rate was calculated as 100 minus the combined relapse and treatment failure rates at the specified time point. The denominator for computation of relapse rates was the number completing treatment.
The evaluable population used in the efficacy assessments reported for Phase 2b included patients who completed treatment with the assigned regimen of study medication and who either had a treatment failure or a relapse at any time prior to the scheduled time point of interest or underwent diagnostic procedures at the time point of interest or later. The last observation carried backwards was used for patients with missing data who had an assessment at a later visit, as long as the patient was not a relapse. For example, if a patient’s last visit was prior to 3 months (ie, the patient was lost to follow-up after the end-of-treatment visit), the patient was considered to be a relapse at 3 months.
For the Phase 2b and 2b-2 studies, the number and percentage of patients with parasitological cure at each evaluation was summarized by treatment group (pafuramidine for 5 days, pafuramidine for 10 days, and pentamidine for 7 days). The denominator included patients who had parasitologically confirmed infection with T.b. gambiense prior to treatment, completed the assigned regimen of study medication, and underwent diagnostic procedures at the end of treatment visit. All efficacy data were tabulated.
All patients who received treatment with the study drug were included in the analysis of safety and tolerability. In all cases, the primary outcome measure for safety analysis was the rate of occurrence of WHO Toxicity Scale Grade 3 (severe) or Grade 4 (potentially life-threatening) adverse events during the observation period. All safety data, including vital signs and adverse events, were tabulated.
All statistical analyses were performed with commercially available software (SAS version 9.0).
The Phase 2a study recruited patients from 31 August 2001 (first patient enrolled) to 28 November 2004 (last patient follow-up completed) and the Phase 2b study recruited patients from 01 April 2003 (first patient enrolled) to 08 February 2007 (last patient follow-up completed).
First stage HAT patients rarely present at a hospital or a treatment center. Therefore, intense screening activities were necessary to identify first stage patients. A total of 107,354 patients were screened to find 869 patients affected with HAT: 360 in Phase 2a, 311 in Phase 2b, and 198 in Phase 2b-2 (Fig 1). The exclusion rate was high (726 of 869 patients, 83.5%); primary reasons were stage 2 HAT, unknown disease stage, and inclusion criteria not met. Despite the high rate of screening exclusion, 32 patients were randomized and treated in Phase 2a, 81 patients in Phase 2b (40 to pafuramidine and 41 to pentamidine), and 30 patients in Phase 2b-2.
As shown in Table 2, study completion rates were high: 29 of 32 (90.6%) patients in Phase 2a, and 100% in both Phase 2b studies. In Phase 2a, 2 patients discontinued due to adverse events and 1 patient discontinued for administrative reasons. The follow-up attendance at Month 24 in all Phase 2b treatment groups was very good through the last assessment. A total of 83% (33 of 40) of patients treated with pafuramidine for 5 days, 78% (32 of 41) of patients treated with pentamidine, and 70% (21 of 30) patients treated with pafuramidine for 10 days underwent evaluation at the 24-month follow-up visit.
Demographic variables (age, gender, weight, height, and body mass index) were summarized for all study patients (Table 3). Across studies, the treatment groups were similar with respect to age, distribution of men and women, height, weight, and body mass index.
Within the Phase 2b study, the treatment groups were also comparable in the time elapsed since symptoms were first observed, and the majority of patients in each treatment group were negative for malaria (range: 87% to 88%) and filaria (range: 83% to 100%), and did not have diarrhea at hospital entry (range: 98% to 100%).
As shown in Fig 1, 29 of 32 patients (90.6%) in Phase 2a, 39 of 40 patients (97.5%) in the Phase 2b pafuramidine 5-day group, all 40 patients in the pentamidine group, and 28 of 30 (93.3%) of patients in the pafuramidine 10-day group were included in the efficacy analysis. There were only 7 patients excluded from the efficacy analysis: 3 in Phase 2a (due to premature discontinuation), 2 in Phase 2b (1 patient in each group lost to follow-up), and 2 patients in Phase 2b-2 (lost to follow-up). There were no protocol deviations.
All enrolled patients were included in the safety analysis. As shown in Table 5, in the Phase 2a study, the majority of adverse events were mild (Grade 1) or moderate (Grade 2) and the most commonly reported adverse events were headache (44%, 14 of 32 patients) and pyrexia (9%, 3 of 32 patients). There was only 1 patient with a severe (Grade 3) adverse event in the Phase 2a study (hypertension, unspecified, lasting for 1 day), which led to premature discontinuation of the study drug. One additional patient prematurely discontinued the study drug to an adverse event of moderate (Grade 2) pyrexia (the duration of the event is not available). There were no events higher than Grade 3.
No adverse event was considered possibly or probably related to the study drug and no serious adverse events were reported. There were minor increases in mean ALT, AST, and creatinine values, but these changes were not considered to be clinically significant; no patient experienced a ≥2-fold increase in ALT or AST.
As shown in Table 6, the overall rate of patients with treatment-emergent adverse events in the Phase 2b studies was higher among patients who received pentamidine (93%, 38 of 41 patients) than in patients who received pafuramidine for 5 days (25%, 10 of 40 patients) or 10 days (57%, 17 of 30 patients). The most commonly reported adverse events were ALT increased and AST increased, which were notably more prevalent in the pentamidine group than in either of the pafuramidine groups. Specifically, in the pentamidine group, 71% (29 of 41) of patients experienced increased ALT; of these, 16 had severe (Grade 3) elevations. In addition, a total of 85% (35 of 41) of pentamidine patients experienced increased AST. In the pafuramidine 5-day and 10-day groups, increased ALT occurred in 1 and 3 patients, respectively, and increased AST occurred in 4 and 5 patients, respectively. Elevations of liver enzymes in HAT patients were concurrent with treatment, were considered mild, resolved spontaneously (within 2–4 days), and were asymptomatic.
Other severe (Grade 3) adverse events included hypertension (1 patient in the pentamidine group) and headache (2 patients who received pentamidine and 1 patient who received pafuramidine for 5 days). All events of headache were considered to be related to trypanosomiasis and its treatment. No other Grade 3 or higher treatment-emergent adverse events were reported.
There were 2 serious adverse events in the Phase 2b studies. One patient who received pafuramidine for 5 days died 85 days after the last dose of the study drug. He was considered to have probable second stage disease with rapid progression, and death due to encephalopathic syndrome occurred during melarsoprol (rescue) treatment. One patient who received pentamidine was lost to follow-up after the 3-month evaluation and was subsequently reported to have died of causes not likely to be related to trypanosomiasis.
Overall, pafuramidine was well tolerated in all Phase 2 studies, and the rate of treatment-emergent adverse events was similar between the 5- and 10-day treatment groups. In the Phase 2b study, elevated ALT and AST values were more frequent in patients who received pentamidine 4 mg/kg QD for 7 days (71% and 85%, respectively) than in patients who received either pafuramidine 100 mg BID for 5 days (3% and 10%, respectively) or 10 days (10% and 17%, respectively).
Despite protocol-directed contraceptive measures, 2 pregnancies occurred during the treatment period, (1 in Phase 2a and 1 in Phase 2b-2). The courses of the pregnancies were normal and there were no abnormalities reported at birth. These children were repeatedly checked in the treatment center until the end of the study and reported to be in overall good health with normal development.
The ECGs from patients enrolled in these studies are included in a separately published study on cardiac alterations in HAT [29]. The results of the analyses suggested that a prolonged QTc interval, repolarization changes, and low voltage were significantly more frequent in first and second stage HAT patients compared with healthy individuals, which did not change during treatment with pafuramidine or pentamidine.
These were the first clinical studies in the field of trypanosomiasis conducted in the centers in Angola and the DRC with local teams that were previously inexperienced in clinical studies. However, they were fully compliant with Good Clinical Practice and regulatory standards.
The results reported here demonstrate the efficacy of 5-day pafuramidine for treatment of T.b gambiense HAT with a 93% cure rate 24 hours post-treatment. Moreover, the parasitological cure rate 3 months after treatment was comparable in the pentamidine 7-day group (100%) and the 10-day pafuramidine treatment group (93%). The cure rate at 6, 12, and 24 months remained comparable between the pentamidine 7-day group (97%, 94%, and 97%, respectively) and the pafuramidine 10-day group (84%, 84%, and 90%, respectively). No treatment failures were observed directly after treatment in the pafuramidine 10-day group.
Pafuramidine was well-tolerated in a treatment regimen of 100 mg given orally BID for 10 days, and the toxicity appeared to be less than that observed for pentamidine 4 mg/kg QD given intramuscularly for 7 days. The efficacy (comparable to pentamidine) and safety (better tolerated than pentamidine) results obtained with the extended 10-day dosing regimen of pafuramidine support the continued clinical development of this drug.
From the perspective of study design, it is noteworthy that the 3-month surrogate endpoint for efficacy used in the Phase 2b study effectively predicted the clinical outcomes determined at the 24-month evaluation. Given that this 3-month endpoint was implemented based on the progression of clinical efficacy observed in the Phase 2a study (with a primary endpoint at 24 hours post-treatment), the overall results represent one form of adaptive design for a series of studies within a clinical development program.
The conduct of the studies also established an organization and infrastructure that greatly facilitated the implementation of the subsequent Phase 3 study, eg, capacity building, improvement of laboratory infrastructure, investment in and deployment of high quality laboratory equipment, introduction of improved diagnostic methods, and improved experience in Independent Ethics Committee set-up and support.
As noted, the conduct of these studies required large-scale screening, which was supported by national HAT programs of Angola (ICCT) and DRC (PNLTHA). The success of the screening, in addition to the robust number of individuals who underwent screening, should contribute to improvement of HAT control in Angola and the DRC. The completion of the Phase 2 studies not only established the parameters for the design of a subsequent Phase 3 study, but also provided a model for future studies of HAT in these and similar populations, and a blueprint for the identification and delivery of treatment to affected individuals in rural Africa.
Both clinical trials were registered in the International Clinical Trials Registry Platform at www.clinicaltrials.gov (Phase 2a study, NCT00802594; Phase 2b and 2b-2, NCT00803933).
|
10.1371/journal.pgen.1001122 | Differentiation of Zebrafish Melanophores Depends on Transcription Factors AP2 Alpha and AP2 Epsilon | A model of the gene-regulatory-network (GRN), governing growth, survival, and differentiation of melanocytes, has emerged from studies of mouse coat color mutants and melanoma cell lines. In this model, Transcription Factor Activator Protein 2 alpha (TFAP2A) contributes to melanocyte development by activating expression of the gene encoding the receptor tyrosine kinase Kit. Next, ligand-bound Kit stimulates a pathway activating transcription factor Microphthalmia (Mitf), which promotes differentiation and survival of melanocytes by activating expression of Tyrosinase family members, Bcl2, and other genes. The model predicts that in both Tfap2a and Kit null mutants there will be a phenotype of reduced melanocytes and that, because Tfap2a acts upstream of Kit, this phenotype will be more severe, or at least as severe as, in Tfap2a null mutants in comparison to Kit null mutants. Unexpectedly, this is not the case in zebrafish or mouse. Because many Tfap2 family members have identical DNA–binding specificity, we reasoned that another Tfap2 family member may work redundantly with Tfap2a in promoting Kit expression. We report that tfap2e is expressed in melanoblasts and melanophores in zebrafish embryos and that its orthologue, TFAP2E, is expressed in human melanocytes. We provide evidence that Tfap2e functions redundantly with Tfap2a to maintain kita expression in zebrafish embryonic melanophores. Further, we show that, in contrast to in kita mutants where embryonic melanophores appear to differentiate normally, in tfap2a/e doubly-deficient embryonic melanophores are small and under-melanized, although they retain expression of mitfa. Interestingly, forcing expression of mitfa in tfap2a/e doubly-deficient embryos partially restores melanophore differentiation. These findings reveal that Tfap2 activity, mediated redundantly by Tfap2a and Tfap2e, promotes melanophore differentiation in parallel with Mitf by an effector other than Kit. This work illustrates how analysis of single-gene mutants may fail to identify steps in a GRN that are affected by the redundant activity of related proteins.
| Neural crest-derived pigment cells, known as melanocytes, are important to an organism's survival because they protect skin cells from ultraviolet radiation, camouflage the organism from predators, and contribute to sexual selection. Networks of regulatory proteins control the steps of melanocyte development, including lineage specification, migration, survival, and differentiation. Gaps in our understanding of these networks hamper progress in effective prevention and treatment of diseases of melanocytes, including metastatic melanoma and vitiligo. Studies conducted in tissue-culture cells and mouse embryos implicate regulatory proteins including the transcription factor TFAP2A, the growth factor receptor KIT, and the transcription factor MITF as being important for multiple steps in melanocyte development. Abnormalities in TFAP2A, KIT, and MITF expression in melanoma highlight the importance of this pathway in human disease. Here we show that a gene closely related to TFAP2A, tfap2e, is expressed in zebrafish embryonic melanocytes and human melanocytes. We provide evidence that Tfap2e cooperates with Tfap2a to promote expression of zebrafish kita in embryonic melanocytes. Further we show that an effector of Tfap2a/e activity other than Kita is required for melanocyte differentiation and that this effector acts upstream or in parallel with Mitfa activity. These findings reveal unexpected complexity to the gene-regulatory network governing melanocyte differentiation.
| An important participant in the gene-regulatory-network (GRN) that governs the differentiation of melanocytes from neural crest precursors (i.e., the melanocyte GRN) is the class III receptor tyrosine kinase Kit. In mouse embryos, binding of this growth-factor receptor by its ligand, stem cell factor (SCF), promotes the growth, survival, migration, and possibly terminal differentiation of melanocytes [1]. Mouse embryos homozygous for hypomorphic alleles of Kit completely lack melanocytes (embryos homozygous for Kit null alleles die prior to pigmentation) [2]–[6]. While ligand-bound Kit stimulates many signal transduction pathways, its effects on melanocyte growth and differentiation appear to occur via the Ras/Raf/Map Kinase pathway. Activity of this pathway results in phosphorylation of Microphthalmia transcription factor (Mitf); phosphorylation of Mitf regulates its activity and stability [7], [8]. Within melanoblasts, Mitf promotes a) cell-cycle exit, by activating expression of the p21WAF1, a cyclin-dependent kinase inhibitor [9], b) cell survival, by upregulating the expression of BCL2 [10], and c) melanin synthesis, by activating expression of Tyrosinase (Tyr), Tyrosinase-related protein 1 (Tyrp1), and Tyrosinase-related protein 2 (Tyrp2, also known as Dopachrome tautomerase, Dct) [11]–[14]. Thus, Kit signaling is essential for normal melanocyte development, at least in part via its ability to stimulate Mitf activity. Of note, KIT levels are reported to be lower in metastatic melanoma cell lines than in benign nevi, and forced expression of KIT in these cells has been shown to induce apoptosis [15]. These findings highlight the importance of understanding the regulation of Kit expression within the melanocyte lineage.
While there is evidence that the KIT gene is dependent on direct stimulation by the Transcription Factor Activator Protein 2 alpha (TFAP2A) in melanoma, analyses of mutant model organisms indicate a more complex regulatory scenario within embryonic melanocytes. TFAP2A and other members of the TFAP2 family control cell fate specification, cell differentiation, cell survival and cell proliferation within neural crest, skin, breast epithelium, and other embryonic cell types and stem cells [16], [17]. Gel shift experiments showed that TFAP2A can bind an element 1.2 kb upstream of the KIT transcription start site, and expression driven by this enhancer in melanoma cells is lost when the TFAP2 binding sites are deleted [18]. Moreover, forced expression of the TFAP2A DNA binding domain, which presumably unseats endogenous TFAP2A and thus acts as a dominant negative AP2, prevents expression of KIT in these cells [18]. Mice lacking the Tfap2a gene do not live long enough to develop melanocytes, due to failure of body wall closure [19], [20]. However, in embryos with Wnt1-CRE-mediated deletion of Tfap2a specifically within the neural crest, melanocytes are absent from the belly [21]. Interestingly, this phenotype resembles that of heterozygous, not homozygous, Kit loss-of-function mutants, suggesting that loss of Tfap2a leads to a reduction rather than complete loss of Kit expression. Zebrafish have two orthologues of mammalian Kit, known as kita and kitb; only kita is expressed in the melanophore lineage [22]. In kita homozygous null mutants (i.e., kita mutants) relative to their wild-type counterparts, embryonic melanophores are reduced in number by about 40%, migrate less, and eventually undergo apoptosis [23]. In zebrafish tfap2a homozygous null mutants (i.e., tfap2a mutants), kita expression is reduced and embryonic melanophores exhibit reduced migration [24], [25]. However, in contrast to the melanophores in kita mutants, those in tfap2a mutants do not appear to die, at least as long these animals survive [23], [26]. The simplest explanation for this difference is that kita expression in melanophores is initially dependent on tfap2a but later becomes independent of it. How can the dominant negative AP2 block Kit expression while loss of Tfap2a only diminishes or delays it? Because many Tfap2 family members have the same DNA binding affinity, it is possible that another such family member cooperates with Tfap2a to activate Kit expression.
Here we show that Tfap2e, a homolog of Tfap2a with the equivalent DNA binding specificity, is expressed in zebrafish melanoblasts and in cultures of primary human melanocytes. With single and double knockdown studies, we show that while Tfap2e is not required for the development of embryonic melanophores, it functions redundantly with Tfap2a in maintaining kita expression in embryonic melanophores. Interestingly, in contrast to the situation in kita mutants, the melanophores in embryos doubly deficient for tfap2a/e fail to differentiate. These results imply that Tfap2 activity has targets other than kita that are important for melanophore development. We find that forced expression of mitfa partially restores melanophores in embryos lacking tfap2a and tfap2e, implying that the targets of Tfap2a/e function to stimulate Mitfa activity or act in parallel with it. These findings reveal unexpected roles for Tfap2 activity in the melanocyte GRN.
To determine if a second Tfap2 family member is expressed in the melanoblast lineage, we identified orthologues of Tfap2b, Tfap2c, Tfap2d, and Tfap2e in a database of expressed sequence tags (www.ensembl.org), amplified partial clones of at least 1 kb from each to make a probe for in situ hybridization, and examined the expression of each in embryos that ranged in stage from 0.5 hours post fertilization (hpf), revealing maternal expression, to 48 hpf. Expression patterns of tfap2b and tfap2c have previously been reported [27], [28]. We did not detect expression of tfap2b, tfap2c, or tfap2d in melanoblasts or melanophores (Figure S1), so we did not pursue these orthologues in the context of melanophore development.
In 8-cell zebrafish embryos, maternal tfap2e transcripts were detected by both in situ hybridization and semi-quantitative RT-PCR (not shown). At 24 hpf, tfap2e expression was detected in several regions of the brain, including presumed olfactory bulb, as in mouse embryos [29], [30] (Figure 1A), and also within dispersed cells in the trunk that we assumed to be a subset of migrating neural crest cells (Figure 1B and 1D). At this stage, tfap2e expression was detectable in early-differentiating melanophores close to the ear (Figure 1C), suggesting that the dispersed, non-melanized cells expressing tfap2e were melanoblasts. To test this possibility, we probed homozygous mitfa null mutant embryos (i.e., mitfab692), which are devoid of melanoblasts [31], and found that tfap2e expression was absent from the dispersed cells in the trunk (Figure 1E-1G). This result was consistent with expression of tfap2e in melanoblasts. However, because mitfa is co-expressed with xdh and fms, two markers of xanthophore precursors [32], it was conceivable that tfap2e was expressed in the xanthophore lineage, in an Mitfa-dependent fashion. To test whether tfap2e is expressed in xanthophores, we processed embryos to simultaneously reveal expression of tfap2e mRNA and Pax7 protein, a marker of the xanthophore lineage [33]. We did not detect overlap of the two signals, which implies that tfap2e is not expressed in xanthophores (Figure 1H). In wild-type embryos at 36 hpf, tfap2e expression was present in the forebrain and presumed optic tectum, and expanded in the hindbrain relative to earlier stages (Figure 1I and 1J). However, at this stage expression was not detected in melanophores (Figure 1K). At 48 hpf, high-level tfap2e expression was also observed in the retina (Figure 1L).
To assess if melanocyte-specific expression of TFAP2E is conserved in humans, we performed quantitative RT-PCR on cDNA generated from various human cell lines. We detected higher levels of TFAP2E message in three independent isolates of primary melanocytes, consistent with microarray data indicating expression of TFAP2E in melanocytes and melanoma cell lines [34]. Expression in melanocytes was 2–10 fold higher than in a keratinocyte cell line, and approximately 50–100 fold higher than in a lymphocyte cell line (Figure 1M). In summary, tfap2e is expressed in zebrafish melanoblasts and in human melanocytes.
As discussed in the Introduction, KITA has been reported to be a direct target of TFAP2A, and a dominant negative AP2 variant was found to block KIT expression in cultured cells [18]; however the status of kita expression in tfap2a mutants has not been fully investigated. In zebrafish tfap2a mutants or tfap2a MO-injected embryos at 28 hpf, kita expression in the melanophore lineage is reduced to undetectable levels as assessed by in situ hybridization [24], [25]. However because melanophores undergo cell death in kita mutants but do not do so in tfap2a mutants, it has been proposed that kita is expressed in the melanophore lineage of tfap2a mutants at a later stage [24]. To test this prediction, we crossed heterozygote tfap2a null mutants (i.e., lockjaw, tfap2ats213) and identified homozygous mutant offspring (hereafter, tfap2a mutants) at 28 hpf by virtue of their pigmentation phenotype. We fixed a fraction of these embryos at 28 hpf, and incubated the remainder in water containing phenylthiourea (PTU) to prevent melanin synthesis, until 36 hpf. We then processed all embryos to reveal kita expression. In tfap2a mutants at 28 hpf, kita expression in melanophores was undetectable by in situ hybridization (Figure 2C), as previously reported. However, at 36 hpf, kita expression was clearly visible in cells present in the dorsum of these embryos (Figure 2E). Thus normal kita expression in melanoblasts at 28 hpf is dependent on tfap2a, but later becomes independent of it. To explain these observations we hypothesized that Tfap2e compensates for the loss of Tfap2a and activates kita expression by 36 hpf.
To test whether Tfap2e maintains kita expression in tfap2a mutants, we first assessed tfap2e expression in tfap2a mutants, and found that it was expressed on schedule in migrating neural crest, as in wild-type embryos (Figure S2). Next we injected embryos with a morpholino (MO) targeting the tfap2e exon 3 splice donor site (i.e., tfap2e e3i3 MO) (Figure 2A). To confirm the efficacy of this MO towards its intended target, we harvested RNA from embryos injected with the tfap2e e3i3 MO, generated first-strand cDNA, and performed PCR using primers in exon 1 and exon 4. Sequencing of the major aberrant splice product revealed that the e3i3 MO causes deletion of exon 3 in its entirety, resulting in a frame shift and a severe truncation of the predicted protein that eliminates the DNA binding domain (Figure 2A). By semi-quantitative PCR, this MO appears to inhibit normal splicing of the majority of tfap2e transcripts at 36 hpf, but to act with greatly reduced efficiency at 3 days post fertilization (dpf) (Figure 2A). By 24 hpf, wild-type zebrafish embryos injected with tfap2e e3i3 MO showed evidence of cell death in the central nervous system (CNS), i.e., patches of opacity in the brain and spinal cord, but no other gross morphological defects; possibly this was due to non-specific toxicity of the MO to the embryo. Despite this cell death, the melanophores that developed in such embryos looked normal and were normally distributed (see below and Figure S3). tfap2e e3i3 MO-induced CNS cell death was reduced by co-injection of p53 MO, implying that Tfap2e has a role in cell survival in the CNS, or that the tfap2e e3i3 MO has non-specific toxicity towards the nervous system, which is true of many MOs (Figure S3) [35]. To preserve the morphology of embryos, in all experiments discussed hereafter we have included p53 MO with tfap2e e3i3 MO. Interestingly, in tfap2a mutants injected with the tfap2e e3i3 MO (hereafter, tfap2a/e doubly-deficient embryos), kita was absent from the dorsum at 36 hpf, although kita expression was readily detected in the cloaca and pharyngeal pouches (Figure 2G and not shown). These findings imply that in absence of Tfap2a, Tfap2e promotes kita expression in the melanophore lineage.
Because of the sustained loss of kita expression in tfap2a/e doubly-deficient embryos, we expected that the phenotype in these embryos would be similar to that of kita homozygous null mutants, although perhaps not as severe because MO-mediated inhibition of gene expression is transient and partial; instead, however, we detected a much more severe phenotype. At 36 hpf, compared to the embryonic melanophores in their non-mutant siblings (Figure 3A), those in kita null mutants (i.e., kitab5) (Figure 3B) appeared normally melanized, but were reduced to about 60% of their normal numbers (because of a presumed defect in cell division) and did not migrate as extensively as their wild-type counterparts [23], [36]. In control MO-injected tfap2a mutants (Figure 3C), embryonic melanophores exhibited these same phenotypes. In tfap2e MO-injected sibling embryos (Figure 3D) there was no apparent melanophore phenotype. However, in tfap2a/e doubly-deficient embryos there were far fewer melanophores than present in control MO-injected tfap2a mutant embryos. Compared with control MO-injected tfap2a mutants, tfap2a/e doubly-deficient embryos had fewer pigmented melanophores in the dorsum and almost no visible melanophores on the lateral sides of the trunk or on the yolk sac (Figure 3E); this difference was still apparent at 84 hpf (not shown). In summary, whereas wild-type embryos injected with the tfap2e MO developed normally until at least 4 dpf, tfap2a/e doubly-deficient embryos displayed melanophore defects more severe than those of tfap2a or kita mutants. These findings suggest that Tfap2a and Tfap2e have partially redundant function in zebrafish melanophore development, and that this function exceeds the simple maintenance of kita expression.
To confirm the specificity of the tfap2e e3i3 MO-induced melanophore phenotypes, we co-injected mRNA encoding a glucocorticoid-fused version of Tfpa2a (tfap2aGR), whose nuclear transport is dexamethasone-inducible, or lacZ as a control, into embryos injected with MOs targeting tfap2a, tfap2e, and p53 (hereafter also termed tfap2a/e doubly-deficient embryos). Dexamethasone was added to both groups at 70% epiboly to avoid gastrulation defects caused by tfap2a over-expression [28]. Embryos were then scored for the rescue of under-melanized melanophores, seen in tfap2a/e doubly-deficient embryos, at 36 hpf. We found that tfap2aGR mRNA effectively rescued melanophores in tfap2a/e doubly-deficient embryos, whereas lacZ did not (Figure S4G and S4H). As an alternative approach for testing specificity, we purchased two additional independent tfap2e MOs—one targeting the exon 2 splice donor site (i.e., e2i2 MO) and the other the translation start site of the tfap2e gene (i.e., AUG MO) (Figure 2A). Injection of either the tfap2e e2i2 MO or the tfap2e AUG MO into wild-type embryos had no effect on melanophore development, although both induced some degree of nervous-system cell death. Upon injection of either the tfap2e e2i2 MO or tfap2e AUG MO into embryos derived from tfap2a mutant heterozygous parents, about one fourth of embryos exhibited the melanophore phenotype seen with the tfap2e e3i3 MO (Figure S4A-S4F); co-injection of p53 MO did not alter the melanophore phenotypes although it reduced nervous system cell death (not shown). These multiple tests of specificity strongly argue that the melanophore phenotypes we observe in tfap2e MO-injected embryos result from inhibition of tfap2e expression and not from off target effects.
The reduced number of melanophores in tfap2a/e doubly-deficient embryos relative to tfap2a mutants could reflect a role for Tfap2a/e activity in the specification of melanoblasts or, alternatively, in either survival or differentiation of melanophores. To distinguish among these possibilities, we examined the expression of mitfa, an early marker of the melanoblast and xanthoblast lineages [31], [32]. At 29 hpf, mitfa-expressing cells are visible in the head and trunk of wild-type embryos injected with a control MO (Figure 4A). The number of mitfa-expressing cells is reduced by about half in tfap2a mutant embryos injected with a control MO (Figure 4B); this reduction results at least in part from the absence of kita in such mutants at this stage, because melanophores are reduced by this amount in kita mutants [23], as are mitfa-expressing cells (our unpublished observations). In tfap2e MO-injected, wild-type embryos, the number of mitfa cells is not grossly different from that in control MO-injected, wild-type embryos (Figure 4C). Interestingly, in tfap2a/e doubly-deficient embryos, the number of mitfa-expressing cells did not appear to be further decreased relative to that in control MO-injected tfap2a mutants (Figure 4D). To confirm these impressions, we counted mitfa-expressing cells over the hind yolk (see Materials and Methods) at 24 hpf, and compared the results for tfap2a mutants injected with control MO versus those injected with tfap2e MO; we found no significant difference (See Figure 4 legend for numbers). In addition, we used fluorescence-activated cell sorting (FACS) to count GFP-positive cells in dissociated mitfa:egfp transgenic embryos injected with MOs, and this analysis supported our findings from histology [37]. Thus, GFP-expressing cells were similarly reduced in tfap2a MO-injected and tfap2a/e doubly-deficient mitfa:egfp embryos (i.e., to about 40% of the number in controls), although the number of differentiated melanophores in tfap2a/e doubly-deficient embryos was clearly reduced relative to that in tfap2a MO injected embryos (Figure 4E, histogram). These findings imply that Tfap2 activity, provided by the redundant actions of Tfap2a and Tfap2e, is involved in a step of melanophore development that occurs subsequent to specification of the mitfa-positive lineage.
To determine which step in melanophore development depends on Tfap2 activity, we analyzed the expression of genes involved in melanophore differentiation: tyr, tyrp1b and dct [12]. In tfap2a mutant embryos at 29 hpf, the number of cells expressing each of these melanophore markers was reduced by about half relative to that in siblings, consistent with the previously described decrease in melanophores in tfap2a mutants (Figure 5A, 5E, 5I and 5C, 5G, 5K) [24], [25]. In tfap2e MO-injected embryos, the number of cells expressing each of these genes appeared to be normal (Figure 5B, 5F, and 5J), while in tfap2a/e doubly-deficient embryos their numbers were further reduced relative to that in tfap2a mutant embryos (Figure 5D, 5H, and 5L). To quantify this effect, we counted cells in embryos processed for in situ hybridization. We discovered that the reduction in gene expression was not equal in all cases. The number of cells expressing dct was most clearly and most consistently reduced in tfap2a/e doubly-deficient embryos, i.e., by approximately 47% relative to the number in tfap2a mutant embryos (Figure 5A-5D, and 5M). The reduction in tyrp1b and tyr expressing cells was more variable, with an average reduction of approximately 30% and 23%, respectively (Figure 5E-5L, and 5M).
The results described above indicate that when the expression of tfap2a and tfap2e is reduced, melanoblasts express mitfa but fail to progress to a stage at which they express normal levels of melanophore differentiation genes, such as dct, tyrp1b, and tyr. To test this model more quantitatively, we injected mitfa:egfp transgenic embryos [37] with either tfap2a MO or both tfap2a MO and tfap2e MO, dissociated them at 29 hpf, sorted and collected GFP-expressing cells, and measured the levels of various transcripts by quantitative RT-PCR (Figure 5N). Using this method, we saw a trend similar to that observed in the histology analysis: in GFP-positive cells sorted from tfap2a/e MO-injected embryos relative to those sorted from tfap2a MO-injected embryos, dct expression was reduced by approximately 45%, tyrp1b expression was reduced by 17%, and unexpectedly, tyr expression was not reduced. Taken together with the cell counts, these results reveal that Tfap2 activity, redundantly provided by Tfap2a and Tfap2e, promotes the differentiation of embryonic melanophores.
We tested the possibility that the loss of differentiated melanophores in tfap2a/e doubly-deficient embryos results from a fate switch of melanophores to xanthophores, because mitfa is co-expressed with c-fms, a marker of xanthophore precursors [32]. We injected embryos with a control MO, tfap2a MO, tfap2e MO, or tfap2a/e MOs, and at 36 hpf processed them to reveal expression of anti Pax7 IR, a marker of xanthophores [33] (Figure 6A-6C and not shown). While the numbers of xanthophores in these groups did not differ significantly (Figure 6D), melanophore differentiation was clearly affected in tfap2a/e doubly-deficient embryos. These findings suggest that loss of Tfap2 activity in the melanophore lineage does not result in a cell fate switch.
We also assessed whether melanophores in tfap2a/e doubly-deficient embryos undergo cell death, i.e. despite the presence of p53 MO. First, we co-injected embryos with MOs targeting tfap2a and tfap2e and with an mRNA encoding Bcl2, an inhibitor of apoptosis [38]. Injection of bcl2 mRNA reduced the number of cells expressing a marker of programmed cell death in control embryos at 25 hpf (Figure 6E and 6F), but had no effect on the melanophore phenotype in tfap2a/e doubly-deficient embryos (Figure 6G and 6H). Secondly, embryos were incubated in acridine orange (AO), which is taken up by dying cells, from 16 hpf to 30 hpf and assessed for the presence of AO-containing cells in the dorsal neural tube and migratory neural crest. Relative to control MO-injected wild-type embryos, control MO-injected tfap2a mutants had an elevated number of such cells, but these numbers were not detectably increased in tfap2e MO-injected tfap2a mutants (data not shown). These findings suggest that loss of Tfap2 activity in melanophores does not result in either a switch in cell fate specification or promotion of cell death, but more likely in inhibition of normal melanophore differentiation.
In tfap2a mutants and MO-injected embryos, embryonic melanophores initially appear somewhat under-melanized [24], [25]. The tfap2a gene is expressed both in skin and neural crest, and we have reported evidence based on transplant studies that Tfap2a has both cell-autonomous and cell non-autonomous effects on melanophore differentiation [25]. Because tfap2e is expressed in melanoblasts but not skin, we assumed that the even poorer differentiation of melanophores in tfap2a/e doubly-deficient embryos is primarily a consequence of a cell autonomous role for Tfap2 activity. To confirm this prediction, we created genetic chimeras by carrying out transplantations at the blastula stage. Specifically, we transplanted cells from 4 hpf wild-type donors, which had been injected with a biotin-dextran as a lineage tracer, into 4 hpf hosts injected with tfap2a/e MO. We then reared the transplanted hosts to 48 hpf, and processed them for biotin staining to reveal the donor-derived cells. Melanophores lacking lineage tracer were indistinguishable from those seen in the untransplanted tfap2a/e MO-injected controls (Figure 7C-7F, arrows), whereas those positive for the lineage tracer were clearly darker, similar to wild-type controls (Figure 7A and 7B), indicating an increase in the level of melanin. In addition, they displayed a more normal morphology (Figure 7E and 7F, arrowheads). These findings indicate that normal melanophores can develop from wild-type cells that are flanked by tfap2a/e-deficient epidermis. This supports a cell-autonomous requirement for Tfap2a/e activity in melanophore differentiation.
Several signals are known to modulate Mitf transactivation activity [39], [40]. If Tfap2a/e is required for the expression of a component of such a signaling pathway, Mitfa activity might be reduced in tfap2a/e doubly-deficient embryos despite levels of mitfa mRNA being similar to those in tfap2a mutants. Alternatively, the Tfap2a/e effector required for melanocyte differentiation might be co-activated by Mitf. In either of these scenarios, forced mitfa expression might rescue melanophore differentiation in tfap2a/e doubly-deficient embryos. We injected tfap2a/e doubly-deficient embryos with a plasmid in which the sox10 promoter drives mitfa expression (sox10:mitfa) [41], and found sox10:mitfa-injection increased the number of tfap2a/e doubly-deficient embryos with differentiated melanophores (compare Figure 8B to Figure 8C, 8D). We observed an increase in the number of darkly-pigmented melanophores in tfap2a/e doubly-deficient embryos injected with sox10:mitfa compared to in tfap2a/e doubly-deficient embryos alone (Figure 8E). We also quantified the mean gray value of single melanophores in these embryos (as a measure of pigment density), within a defined region, using ImageJ software. We found that there was a significant reduction in the pigment density of tfap2a/e doubly-deficient embryo melanophores, compared to control MO-injected embryo melanophores, and that this density was restored in doubly-deficient embryos co-injected with sox10:mitfa (Figure 8F).
Since sox10 is expressed throughout the neural crest, we considered the possibility that sox10:mitfa might induce a conversion of neural crest to the melanoblast lineage, and that if this were to occur in neural crest that expressed another Tfap2 family member, normally differentiated melanophores might emerge in tfap2a/e doubly deficient embryos. However, arguing against this alternative model, we did not detect an increase in the number of melanophores in control-MO injected embryos co-injected with the sox10:mitfa plasmid (Figure 8E). Moreover, in this alternative model, tfap2b is the best candidate Tfap2 family member, as it is expressed in Rohon Beard sensory neurons [27], which are closely related to trunk neural crest [42], [43]. However, we found that even in embryos triply depleted of tfap2a/b/e using MOs, co-injection of sox10:mitfa plasmid elevated the number of normal-looking melanophores (our unpublished observation). Together these observations support the model that over-expression of mitfa can compensate for the role in melanophore differentiation normally played by Tfap2a/e, implying that the effector of Tfap2a/e-type activity necessary for melanophore differentiation acts upstream or in parallel with Mitfa.
Here we have presented two new findings relevant to the gene-regulatory-network (GRN) that governs the differentiation of zebrafish embryonic melanophores. First, kita expression in embryonic melanophores is positively regulated by Tfap2e, at least when Tfap2a levels have been reduced. Expression of tfap2a is present throughout the neural crest starting at the neurula stage, while the expression of tfap2e starts at approximately the time of neural crest delamination and appears to be restricted to melanoblasts [24], [25]. The relative timing of tfap2a and tfap2e expression explains why kita expression (in melanophores) in tfap2a mutants is reduced at 28 hpf, but present at later stages; Tfap2e compensates for the absence of Tfap2a but only after 28 hpf. The presence of TFAP2E expression in human melanocytes suggests that TFAP2A and TFAP2E have redundant or partially redundant function in mammalian melanocytes, as in fish melanophores. If so it would explain the observation, mentioned in the Introduction, that the coat color phenotype in mice with neural crest-specific deletion of Tfap2a is less severe than that of Kit homozygous null mutants [21].
The second unexpected finding is that Tfap2 activity (provided by Tfap2a and Tfap2e) promotes the differentiation of embryonic melanophores. This was revealed by reduced expression of the dct and tyrp1b mRNAs, as well as of melanin—changes that are evident in tfap2a mutants and more pronounced in tfap2a/e doubly-deficient embryos. Does Tfap2 activity also direct neural crest cells to join the melanophore sublineage? There is precedent for such a possibility, because Tfap2 activity provided by Tfap2a and Tfap2c appears to direct ectodermal precursors to join the neural crest lineage [28], [44]. In tfap2a single mutants, neural crest induction appears to occur normally, but mitfa-expressing cells, which are primarily melanoblasts, are reduced in number. This reduction may reflect a role for Tfap2 in melanophore specification or alternatively a reduction of Kita-mediated proliferation of melanoblasts. Whatever the explanation for reduced melanoblasts in tfap2a mutants, simultaneous reduction of tfap2a and tfap2e leads to a further reduction of melanophore numbers without a further reduction of mitfa-expressing cells, arguing Tfap2 promotes differentiation of melanoblasts to melanophores. While a reduction of melanophores without a reduction in mitfa-expressing cells might have been consistent with a cell fate change of melanophores to xanthophores (because markers of melanoblasts and xanthoblasts are briefly co-expressed [32]), xanthophore numbers are equivalent in tfap2a deficient and tfap2a/e doubly-deficient embryos, arguing against such a fate transformation. Does Tfap2 also promote survival of melanophores? We did not detect evidence of cell death of melanophores shortly after their differentiation in tfap2a/e doubly-deficient embryos. We predict that in embryos permanently deprived of both Tfap2a and Tfap2e melanophores would die as a consequence of the absence of Kita. However, because melanophores persist for several days in kita mutants, and this is longer than MOs are effective (see Figure 2A), it will be necessary to isolate a tfap2e mutant to test this prediction. Together these observations reveal that Tfap2 activity has multiple roles in melanophore development, including promoting melanophore differentiation.
Another result that will be important to revisit when a tfap2e mutant is available is the apparent heightened Tfap2-dependence of dct expression relative to tyr expression. Consistent with differential regulation of these related genes, in mice, Dct expression appears prior to Tyr expression, and this has also been suggested to be the case in zebrafish [45], [46]. However, because we knock-down tfap2e expression with an MO, the stronger effect on dct expression relative to on tyr expression may simply reflect loss of MO effectiveness over time. There may be a similar explanation for the inconsistent findings regarding tyr expression between the RNA in situ hybridization and the quantitative RT-PCR analyses. The cell dissociation protocol required for quantitative RT-PCR introduces a delay in the analysis of gene expression relative to that obtained using the RNA in situ hybridization protocol, giving further time for the MO to lose efficacy. Nevertheless, these results reveal that Tfap2 activity, redundantly provided by Tfap2a and Tfap2e, promotes the differentiation of embryonic melanophores.
How does Tfap2 activity, mediated by Tfap2a and Tfap2e, effect melanophore differentiation? In tfap2a/e doubly-deficient embryos, melanophore differentiation fails but can be rescued by forced expression of mitfa. One model to explain these findings is that Mitfa and Tfap2 normally co-activate genes important for melanophore differentiation, but in the absence of Tfap2, elevated levels of Mitfa can suffice to do so (Figure 9A). Thus, Tfap2 family members may directly activate genes involved in melanin synthesis, such as dct, tyrp1b, and possibly tyr, all of which are known to be Mitfa targets [47]–[49]. Consistent with this possibility, recent studies have identified conserved DNA elements adjacent to the dct and tyrp1b genes that have melanocyte enhancer activity [13], and some of these contain putative Tfap2 binding sites. Simultaneous inhibition of tyrp1a and tyrp1b blocks melanization of zebrafish melanophores, suggesting that tyrp1a/b may partially mediate Tfap2a/e activity within these cells [50]. A variation of this model is that, rather than Tfap2 itself functioning as a co-activator with Mitfa, the protein product of a gene stimulated by Tfap2 does so. For instance, Tfap2 activates expression of estrogen receptor alpha (ERα) [51], [52]. ERα, together with p300, interacts with Mitf to strongly activate the Dct promoter [53].
It is also possible that the effector of Tfap2 activity is an enzyme that alters the activity, translation, or longevity of the Mitfa protein (Figure 9B). Thus, perhaps mitfa RNA levels are the same in tfap2a deficient vs. tfap2a/e deficient embryos, but Mitfa activity is reduced in the latter. For instance, the Tfap2-effector may be a receptor tyrosine kinase (RTK) whose activity results in posttranslational activation of Mitfa, i.e. similar to a proposed role of Kit [7], [8]. Supporting such a possibility, Kita itself is necessary for differentiation of embryonic melanophores in zebrafish in certain experimental conditions [54] [23]. A variety of RTKs are candidates for the Tfap2 effector in melanophore differentiation, including Erbb3 [55], [56], IGF1R [57], FGF receptor [58], c-Ret [59], and c-MET [60]. Two G-protein coupled receptors, which like RTKs can stimulate the MAP Kinase pathway, are also candidates. First, Endothelin receptor b (Ednrb signaling) promotes melanocyte differentiation in mammals, in part by activating MAP Kinase signaling and Mitfa [61]–[64]. While embryonic melanophores differentiate normally in zebrafish ednrb1 mutants [65], uncharacterized ednrb homologues are present in the zebrafish genome (e.g., on chromosome 9) and may function in embryonic melanophores. Second, Melanocortin 1 receptor (Mc1r) is necessary for normal levels of pigmentation in zebrafish [66] and in mammals [67], and MC1R expression may be directly regulated by TFAP2A, because it has been shown that TFAP2A binds DNA adjacent to the MC1R gene in HeLa cells (chromatin immunoprecipitation results) [68]. Finally, Tfap2 could normally repress expression of an Mitfa phosphatase, alter processing of the mitfa transcript, change Mitfa translation or change Mitfa protein stability. All these scenarios would result in similar mitfa mRNA levels in situ but weaker Mitfa activity when Tfap2 levels are reduced, and would potentially be by-passed by over-expression of mitfa mRNA. The direct targets of Tfap2 in melanocytes are currently under investigation.
Zebrafish embryos and adults were reared as described previously [69], in the University of Iowa Zebrafish Facility. Embryos were staged by hours or days post fertilization at 28.5°C (hpf or dpf) [70]. Homozygous mutant embryos were generated from heterozygous adults harboring a presumed null allele of tfap2a (lockjaw, tfap2ats213) [26], mitfa (mitfab692) [31], or kita (kitab5) [23], as indicated.
First-strand cDNA was synthesized from total RNA harvested from embryos at 4 hpf and 24 hpf as described [25]. A 1.4 kb full-length zebrafish tfap2e cDNA was amplified from the wild-type cDNA using the following primers: forward, 5′-GGA TTC ATG TTA GTC CAC TCC TAC TC-3′, reverse, 5′-TTA TTT GCG GTG CTT GAG CT-3′. This cDNA includes the entire open reading frame and was inserted into the pCR4-TOPO vector (Invitrogen, Carlsbad, CA). A 1.3 kb fragment of zebrafish tyrp1b cDNA was amplified from the wild-type 24 hpf cDNA using the following primers: forward, 5′-GAG AGC GGA TGA TAT AAG GAT GTG G-3′, reverse, 5′-GCC CAA TAG GAG CGT TTT CC-3′. This cDNA was inserted into pSC-A vector (Stratagene, La Jolla, CA).
In designing a tfap2e construct in which expression is disrupted, the exon 2 splice donor site and the exon 3 splice donor sites had to be inferred from comparison of the cDNA to the corresponding genomic sequence (http://uswest.ensembl.org/Danio_rerio/Info/Index). MOs complementary to these sites were ordered: tfap2e e2i2 MO, 5′-ATA CAA GAG TGA TTG AAC TCA CCT G-3′; tfap2e e3i3 MO, 5′-CAC ATG CAG ACT CTC ACC TTT CTT G-3′ (Gene Tools, Philomath, OR). In addition, a MO targeting the tfap2e translation start site (AUG MO) was designed, 5′-GCT GGA GTA GGA GTG GAC TAA CAT C-3′. MOs were reconstituted to 5 mg/ml in water and stored at room temperature (25°C). Immediately before use, they were diluted to 0.5 mg/ml in 0.2 M KCl. MOs (4–8 nl of diluted stock) were injected into the yolk underlying the blastomeres of embryos at the 1–4 cell stage. Upon injection of 3 ng or more of either MO, we saw evidence of non-specific toxicity, i.e., patches of opacity in the brain and spinal cord that did not develop when 5 ng of a p53 MO (5′-GCG CCA TTG CTT TGC AAG AAT TG-3′) was injected [71]. To assure strong penetrance while preventing non-specific toxicity, we used 3 ng/embryo of tfap2e e3i3 MO plus 5 ng/embryo of p53 MO to generate tfap2a−/eMO embryos. For double MO experiments (tfap2aMO/tfap2eMO), 3 ng of tfap2e e3i3 MO, 5 ng tfap2a e2i2 MO (5′-GAA ATT GCT TAC CTT TTT TGA TTA C-3′) and 5 ng of p53 MO were injected together. To test the efficacy of the tfap2e MOs, we used a pair of primers flanking a 305 bp fragment between exon 2 and exon 4 of tfap2e for RT-PCR (forward, 5′-CAC CAC GGC CTG GAT GAT ATT-3′; reverse, 5′-AGG ACT CCT CCA AGC AGC GA-3′). Additionally, where noted, a control MO (controlMO) was used for comparison (5′-CCT CTT ACC TCA GTT ACA ATT TAT A-3′).
To create genetic chimeras, we injected donor embryos with 5 nl of 1% lysine-fixable biotinylated-dextran, 10,000 MW (Sigma, St. Louis, MO). At the sphere stage (4 hpf), about 100 cells were withdrawn from each donor embryo using a manual-drive syringe fitted with an oil-filled needle (Fine Science Tools, Vancouver, BC), and about 20 cells were inserted into each of several host embryos at the same stage. In placing these cells, we aimed for a position near the animal pole, to target clones to the ectoderm [72]. Host embryos were allowed to develop to 48 hpf, fixed, images taken, and then processed using an ABC kit (Vector Labs, Burlingame, CA) and DAB to reveal biotin as previously described [73], and subsequently photographed.
The following restriction fragments were used to generate DIG-labeled antisense RNA probes (Roche Diagnostics, Mannheim, Germany) for whole mount in situ hybridization: tfap2e, NotI/T3; tyrp1b, BamHI/T3; dct, EcoRI/T7 [74]; mitfa, EcoRI/T7 [31]. Standard procedures were followed as previously described [75]. For total cell counts, 10–20 embryos were analyzed per group (see figure legend).
For immunohistochemistry, a monoclonal anti-Pax7 antibody [33] was used at a 1∶25 dilution (supernatant obtained from the Developmental Studies Hybridoma Bank at the University of Iowa, USA). The primary antibody and an anti-DIG antibody were added during routine whole mount in situ hybridization. Following development of whole mount in situ hybridization with NBT/BCIP, the embryos were blocked and then incubated with an Alexa-488 conjugated goat-anti-rabbit secondary antibody, as previously described [76]. After several washes, the embryos were mounted in 50% glycerol/PBST, and photographed. Cell counts were performed on ten embryos per group, along the entire length of the hind yolk.
Live embryos were reared to an appropriate stage, homogenized with a pestle, and dissociated with PBS containing trypsin and EDTA for 30 minutes at 33°C. After dissociation, cells were resuspended in PBS plus 3% fetal bovine serum (FBS). EGFP-positive cells were counted using a Becton Dickinson FACScan. For cell sorting, cells were dissociated as previously described, and subsequently sorted, on a Becton Dickinson FACS DiVa, directly into buffer RLT and β-mercaptoethanol for subsequent RNA isolation (RNeasy Plus Mini Kit, QIAGEN, Valencia, CA). FACScan cell counting, FACS DiVa cell sorting, and data analyses were conducted at the University of Iowa Flow Cytometry Facility.
The isolation and culture of normal melanocytes and keratinocytes was performed as described previously, Mel 1 and Ker [77], [78], Mel 2,3 [79] (see Figure 1M). Total messenger RNA was isolated using an RNeasy Plus Mini Kit (QIAGEN, Valencia, CA), along with on-column DNase digestion according to the manufacturer's instructions. Lymphocytes (Jurkat cells, clone E6-1) were obtained (ATCC, Manassas, VA), and total RNA was isolated using the PerfectPure RNA Kit (following manufacturer's instructions, 5 PRIME Inc., Gaithersburg, MD). RNA concentrations were determined using a NanoDrop spectrophotometer (Thermo Scientific) and diluted to equal concentrations. For complementary DNA (cDNA) reactions, approximately 200 ng of total RNA was added to 0.5 µg random hexamers, plus 2.5 µl of 10 mM dNTPs (Invitrogen; Carlsbad, CA), and brought to 30 µl with nuclease-free water. Reactions were heated to 65°C for 5 minutes, and cooled to 4.0°C for 5 minutes in a PTC-200 Peltier Thermo Cycler (MJ Research; Ramsey, MN). We then added 19 µl of a master mix containing 10 µl of 5x First-Strand buffer (Invitrogen), 5 µl of 0.1 M dithiothreitol, 20 units of RNasin (Promega, Madison, WI), and nuclease-free water to a volume of 19 µl. Reactions were incubated at 25°C for 10 minutes, and then at 37°C for 2 minutes. Then 1 µl of Moloney-murine leukemia virus Reverse Transcriptase (New England Biolabs, Ipswich, MA) or 1 µl nuclease-free water was added to each reaction. Reactions were carried out at 37°C for 2 hours, followed by incubation at 75°C for 15 minutes. PCR reactions (25 µl) were prepared with approximately 10 ng of cDNA, using the SYBR Green kit (Applied Biosystems, Foster City, CA) following the manufacturer's instructions. The following primers were used at a final concentration of 200 nM in separate PCR reactions: human TFAP2E (forward: 5′-AAT GTG ACG CTG CTG ACT TC-3′; reverse: 5′-GGT CCT GAG CCA TCA AGT CT-3′); or human GAPDH (forward: 5′-AGG TCG GAG TCA ACG GAT TTG-3′; reverse: 5′-GTG ATG GCA TGG ACT GTG GT-3′). Quantitative real-time PCR in Low 96-well plates (Bio-Rad, Hercules, CA) was conducted using a Bio-Rad thermal cycler (CFX96 Real-Time PCR Detection System) and following the default protocol. Primers were designed to flank large exon-intron boundaries to avoid the potential amplification of contaminating genomic DNA. Also, RNA samples not reverse-transcribed (-RT) were used as a negative control. The 2ΔΔCT method was used to determine relative levels of gene expression between samples (normalized to GAPDH) [80]. Experiments were performed in triplicate and mean and standard error were calculated. Following real-time PCR, melt-curve analysis was performed to determine reaction specificity. Similar methods were used for qRT-PCR of sorted cells, with the exception that approximately 20 ng of RNA was used for cDNA synthesis. The following primers were used at a final concentration of 200 nM in separate PCR reactions: tyr (forward: 5′-GGA TAC TTC ATG GTG CCC TT-3′; reverse: 5′-TCA GGA ACT CCT GCA CAA AC-3′); tyrp1b (forward: 5′-TAT GAG ACA CTG GGC ACC AT-3′; reverse: 5′-CAC CTG TGC CAT TGA GAA AC-3′); dct (forward: 5′-CCT CGA AGA ACT GGA CAA CA-3′; reverse: 5′- CAA CAC CAA CAC GAT CAA CA-3′); and β-actin (forward: 5′-CGC GCA GGA GAT GGG AAC C-3′; reverse: 5′-CAA CGG AAA CGC TCA TTG C-3′). Again, the 2ΔΔCT method was used to determine relative levels of gene expression between samples, first normalizing both samples to β-actin, and then comparing relative gene expression levels in tfap2a/e doubly-deficient cells to those in tfap2a deficient cells.
Apoptotic cell death was revealed in whole embryos by terminal transferase dUTP nick-end labeling (TUNEL) as described [81]. The terminal transferase reaction was terminated by incubation at 70°C for 30 min, and embryos were processed with anti-FITC-alkaline phosphatase antibody and developed with NBT/BCIP, as for an RNA in situ hybridization.
For tfap2aGR mRNA rescue experiments, approximately 5 nL of 0.075 mg/mL tfap2aGR or lacZ encoding mRNA, transcribed in vitro (mMessage mMachine kit, Ambion, Austin, TX) was injected into one of four cells of embryos previously injected with tfap2a/e/p53 MOs (similar concentration as indicated before). Embryos were raised until they reached approximately 75% epiboly, at which point dexamethasone (dissolved in EtOH) was added to the fish water at a final concentration of 40 µM. For DNA rescue experiments, 5 nL of a 0.025 mg/ml plasmid encoding 4.9 Kb of the sox10 promoter driving full length mitfa [41] was injected at the one cell stage, followed by co-injection of various MO combinations (control MO and p53 MO or tfap2aMO,tfapeMO). Embryos were then raised until approximately 36 hpf and fixed in 4% paraformaldehyde overnight. Finally, embryos were rinsed in PBST, mounted in 3% methylcellulose, and photographed.
To analyze the mean gray value of melanophores, embryos were first fixed at the appropriate stage in 4% paraformaldehyde overnight. Embryos were then rinsed in PBST and mounted in 3% methylcellulose, and images of single melanophores were taken near the otic vesicle at 40x. All lighting conditions remained constant throughout image capturing. 6–10 melanophores were imaged per embryo, and 10 embryos were analyzed per group (roughly 70–80 melanophores per group). Images were converted to a 32 bit gray image and then processed using the auto threshold function in ImageJ software (Version 1.40 g, National Institutes of Health, Bethesda, MD), creating an outline of the melanophore being analyzed. After application of the auto threshold function, a selection was created of the pixels highlighted, and a measurement reporting mean gray value for the given area was taken. An inverse of the selection was then created, highlighting the background (area not occupied by the melanophore), and a similar measurement was taken, reporting the mean gray value of the surrounding background. The difference was then calculated between the mean gray value of the melanophore and the surrounding background, resulting in the normalized mean gray value of the melanophore. Averages were then calculated for all melanophores measured per group, and standard deviation was calculated.
For mitfa-positive and TUNEL-positive cell counts, the entire region overlying the hind yolk was counted. For melanophore cell counts in sox10:mitfa rescue experiments, the total number of melanophores in the embryo body (excluding yolk and hind yolk) were counted. Embryos were fixed in 4% paraformaldehyde overnight, washed in PBST, and mounted in 3% methylcellulose for counts. Embryos were mounted and then counted blindly by an independent observer.
|
10.1371/journal.pntd.0007349 | Modelling the impact of a Schistosoma mansoni vaccine and mass drug administration to achieve morbidity control and transmission elimination | Mass drug administration (MDA) is, and has been, the principal method for the control of the schistosome helminths. Using MDA only is unlikely to eliminate the infection in areas of high transmission and the implementation of other measures such as reduced water contact improved hygiene and sanitation are required. Ideally a vaccine is needed to ensure long term benefits and eliminate the need for repeated drug treatment since infection does not seem to induce lasting protective immunity. Currently, a candidate vaccine is under trial in a baboon animal model, and very encouraging results have been reported. In this paper, we develop an individual-based stochastic model to evaluate the effect of a vaccine with similar properties in humans to those recorded in baboons in achieving the World Health Organization (WHO) goals of morbidity control and elimination as a public health problem in populations living in a variety of transmission settings. MDA and vaccination assuming different durations of protection and coverage levels, alone or in combination, are examined as treatment strategies to reach the WHO goals of the elimination of morbidity and mortality in the coming decade. We find that the efficacy of a vaccine as an adjunct or main control tool will depend critically on a number of factors including the average duration of protection it provides, vaccine efficacy and the baseline prevalence prior to immunization. In low prevalence settings, simulations suggest that the WHO goals can be achieved for all treatment strategies. In moderate prevalence settings, a vaccine that provides 5 years of protection, can achieve both goals within 15 years of treatment. In high prevalence settings, by vaccinating at age 1, 6 and 11 we can achieve the morbidity control with a probability of nearly 0.89 but we cannot achieve elimination as a public health problem goal. A combined vaccination and MDA treatment plan has the greatest chance of achieving the WHO goals in the shorter term.
| Nearly 258 million people are infected worldwide by schistosome parasites. The World Health Organization (WHO) has set control guidelines to combat the morbidity and mortality induced by infection, defined by reaching ≤5% and ≤1% prevalence of heavy-intensity infections in school-aged children (SAC), respectively. Mass drug administration (MDA) is the major route for morbidity control and elimination. However, MDA does not provide long-term protection against schistosome parasites and frequent drug administration is therefore required to control morbidity. Infection does not induce lasting acquired immunity to reinfection. Drug resistance is another issue with MDA which, if it arises, could possibly make drug treatment ineffective over time as drug-resistant genes in the parasite population increase in frequency. A vaccine is ideally needed to both reduce the possibility of reinfection and to achieve transmission elimination within a feasible time frame. Based on the recent results obtained for a new candidate vaccine in the baboon animal model, we employ an individual-based stochastic model to assess the impact of a vaccine with an efficacy of 100% when applied in endemic regions with different intensities of transmission. Simulations suggest that the probability of achieving morbidity control and elimination as a public health problem depends on the duration of protection provided by vaccination, the age categories of the human host population vaccinated, and the coverage levels achieved. In order to achieve elimination as a public health problem, model simulations suggest that combining vaccination (with 5 years of protection) with MDA (treating 75% of school-aged children, 5–14 years of age) is the best option, particularly in high transmission settings.
| Schistosomiasis inflicts significant levels of human morbidity and mortality in regions of the world with endemic infection. It is estimated that nearly 258 million people are infected worldwide with up to 700 million at risk of being infected, leading to an estimated 280000 deaths annually [1–3]. Schistosomiasis is an intestinal or urogenital disease caused predominantly by infection with Schistosoma mansoni, S. japonicum or S. haematobium, and is one of the diseases included within the World Health Organization (WHO) 2020 goals for neglected tropical diseases (NTD) control. Individuals become infected when cercariae (larval forms of the parasitic worm), released by an intermediate host (various freshwater snail species), penetrate the skin during contact with contaminated water [4]. Control programmes are at present based on mass drug administration (MDA) using the drug praziquantel, and behaviour modification directed at reducing water contact and improvements in sanitation. MDA has to be repeatedly used, since clearing infection does not result in acquired immunity and treated individuals can be re-infected. Age-related water contact behaviour results in most infection residing in school-aged children (SAC; 5–14 years of age), since age intensity of infection profiles are convex in shape. Treatment is therefore specifically focused on this age group. At present, pre-school aged children (pre-SAC) are not eligible for treatment with praziquantel [5] due to the absence of clinical data on the drug effects and safety in the very young. In the coming years a new formulation of praziquantel may be approved for very young children [6]. In areas of high transmission, WHO guidelines also recommend treatment of adults at risk [1], [7]. By 2020, WHO aims to increase coverage in areas of endemic infection such that 75% of SAC at risk will be regularly treated [2], but progress to date in reaching this target has been poor in many regions.
Currently WHO recommends using prevalence of infection in SAC to determine how often to treat in a given endemic area [1]. The recommended treatment strategy for schistosome infection is dependent upon whether the community has a low (< 10%), moderate (10–50%) or high (≥ 50%) prevalence at baseline before the implementation of MDA. The strategy for low-risk communities is to treat all SAC twice during their primary schooling age, generally once every three years, and supply praziquantel in local health centres to treat suspected cases. For moderate-risk communities, the recommendation is to treat all SAC and at-risk adults once every two years. For high-risk communities, the recommended approach is to treat all SAC and at-risk adults once a year. At present in national NTD control programmes, schistosomiasis has one of the lowest levels of MDA coverage of all helminth diseases [8], [9].
Given that MDA needs to be administered to individuals frequently, and that it does not provide long-term protection against the infection in the absence of a strong acquired immunological response to infection, a vaccine is ideally needed for control in the longer term. At present, there is no vaccine for use in humans that can protect against the schistosome infection. However, recent experimental studies by Afzal Siddiqui and colleagues on a candidate vaccine against Schistosoma mansoni infection in a baboon animal model have produced some encouraging results. In four independent, double-blinded studies, a Sm-p80-based vaccine exhibited potent prophylactic, anti-egg induced pathology and transmission-blocking efficacy against S. mansoni in the baboon (Papio ursinus) animal model [10]. The vaccine reduced female worm establishment by 93.45% and significantly resolved the major clinical manifestations of hepatic/intestinal schistosomiasis by reducing the tissue-egg load by 91.35%. A 40-fold decrease in faecal egg excretion by those few female parasites that established in the vaccinated animals, combined with a 79.21% reduction in hatching ability of eggs (the release of viable miracidia), suggests the vaccine may have a high transmission blocking potential. The study showed comprehensive evidence for the effectiveness of a Sm-p80-based vaccine for schistosomiasis and provided support for the need to move beyond animal models to human studies.
Based on the baboon experiments by Siddiqui and colleagues, and assuming efficacy would be similar in humans, published epidemiological analyses based on mathematical models have predicted that the Sm-p80-based vaccine could potentially block infection in areas of low and moderate transmission provided the duration of protection provided by the vaccine is 5 years or more [11], [12]. These models were simple in structure and built on a deterministic framework. This study extends these analyses using an individual based stochastic model to look at the impact of a vaccine, with varying durations of protection, employed in different community-based vaccination programmes involving either vaccinating young children in a cohort-based approach or vaccinating the whole community across all age classes). Analyses are also presented of the impact on transmission and the prevailing levels of infection using either vaccination alone, MDA alone (the current most commonly used intervention to control morbidity) and or using both in different combinations. A description of the impact of MDA, alone on the prevalence and intensity of S. mansoni infection in various transmission settings, is covered in a series of recent publications, as is model structure, model assumptions and data sources for the key transmission and biological parameters [3], [4], [7], [8].
The focus in the present analyses is on the relative merits of vaccination versus MDA, alone or in combination, as a tool for the community control of the morbidity induced by S. mansoni and the likelihood of transmission elimination.
Past work on the impact of MDA on Schistosoma mansoni has employed a hybrid deterministic model (with deterministic and stochastic components) based on sets of partial-differential equations to describe changes in the mean worm burden M(t, a), for host a over time t [13–15].
Stylianou et al, developed an age independent deterministic model to explore the effect of community vaccination programmes [11]. We extend this deterministic model and develop an individual-based stochastic model (an earlier version is described in [4]), where an individual of age a can be in one of the two categories; (i) unvaccinated group or (ii) vaccinated group, denoted by Nu(a, t) and Nv(a, t) respectively. We assume that the number of births is the same as the number of deaths (constant size for the human host), hence the total population of age a, at time t is N(a, t) = Nu(a, t) +Nv(a, t). The unvaccinated and vaccinated host dynamics can be described by the following system of partial differential equations (PDEs):
∂Nu(a,t)∂t+∂Nu(a,t)∂a=−q(a,t)Nu(a,t)+ωNv(a,t)−μ(a)Nu(a,t)
(1)
∂Nv(a,t)∂t+∂Nv(a,t)∂a=q(a,t)Nu(a,t)−ωNv(a,t)−μ(a)Nv(a,t)
(2)
Here q(a,t) is the fraction of the population of age a vaccinated at time t, ω=1durationofvaccineprotection is the vaccine decay rate and μ(a) is the host mortality rate.
The vaccine candidate is assumed to act on the following variables [cf. Eqs (3) and (4)]; (i) parasite establishment within the human host by reducing the rate of infection, β, (ii) parasite survival and growth within the human host, by reducing adult worm life expectancy, σ and (iii) reducing the rate of egg production, λ, due to a reduced growth rate in humans. We assume that the vaccine’s impact on worm death rate, eggs per gram (EPG) and age-specific contact rates are v1, v2 and v3 respectively, where the values range from 0 to 1. The total worm burden in the unvaccinated and vaccinated hosts are denoted by Mu and Mv and the changes in Mu and Mv, over time for host a are described by the following equations:
∂Mu(a,t)∂t+∂Mu(a,t)∂a=Lβ(a)Nu(a,t)−q(a,t)Mu(a,t)+ωMv(a,t)−(μ(a)+σ)Mu(a,t)
(3)
∂Mv(a,t)∂t+∂Mv(a,t)∂a=Lv3β(a)Nv(a,t)+q(a,t)Mu(a,t)−ωMv(a,t)−(μ(a)+v1σ)Mv(a,t)
(4)
Here L represents the concentration of the infectious material in the environment, namely, how each individual of age a, contributes to the pool of released eggs. This is discussed in detail in [14] and [16]. It is assumed that the rates of turn over for the miracidia, snail intermediate host and cercaria are much faster (life expectancies days to weeks) than the adult worm in the human host (life expectancy 4–6 years), so the dynamics of these life cycle stages are collapsed into the equations for the adult worms in humans as detailed in Anderson & May [15].
The total worm burden in the population is given by the sum of the total worm burden in the unvaccinated and vaccinated hosts.
If we denote the total worm burden in the population as the sum of the total worm burden in the unvaccinated and vaccinated hosts by M(a,t)¯=Mu(a,t)+Mv(a,t) and add Eqs (3) and (4) together we obtain the following,
∂M(a,t)¯∂t+∂M(a,t)¯∂a=Lβ(a)Nu(a,t)+Lv3β(a)Nv(a,t)−σM(a,t)¯−μ(a)M(a,t)¯
(5)
In Eq (5) we have assumed v1 = 1. We can express M(a,t)¯ in terms of the mean worm burden, M(a,t), as M(a,t)¯=N(a,t)M(a,t). Then we obtain;
∂M(t,a)∂t+∂M(t,a)∂a=Lv3β(a)Nv(a,t)+Lβ(a)Nu(a,t)N(a,t)−σM(a,t)
(6)
The egg output (from the vaccinated and unvaccinated populations) is given by
E=ψL¯∫a=0∞{Nu(a,t)F(Mu(a,t)Nu(a,t);λ)+Nv(a,t)F(Mv(a,t)Nv(a,t);v2λ)}ρ(a)da
(7)
given
dLdt=E−μ2L
(8)
where the death rate is that of infected snails.
In the above equation ψ describes the flow of the infectious material into the reservoir while the function F(M(a,t); λ) generates the egg output as a function of mean worm burden and ρ(a) represents the age-specific relative contribution of infectious stages to the environmental reservoir. In our simulations we assume the host contribution to the reservoir to be the same as the age-specific contact rates, β(a).
This model has a full age structure for the human host where the outputs are grouped into three age categories, pre-SAC (0–4 years of age), SAC (5–14 years of age) and adults (15+ years of age). We use these age groupings based on WHO definitions of treatment groups [1–3] to calculate the necessary coverage levels (MDA or vaccination) for each category in order to interrupt transmission. This is typically defined as the overall R0 <1 in infectious disease epidemiology, but as shown by Anderson and May [14], the system of equations defined above has three possible equilibria; namely, a stable endemic state, an unstable boundary (transmission breakpoint) and a stable state of parasite extinction. This model is hybrid in the sense that assumes a negative binomial form for the distribution of parasite numbers per host with a fixed aggregation parameter k, density dependent fecundity, and assumed monogamous sexual reproduction among worms.
The mean expected behavior of the individual based stochastic model is identical to the predictions of a deterministic version of the model. However, an individual-based stochastic model permits the examination of the probability distribution of a given event occurring, such as transmission elimination, in a defined period of time during which control measures are applied.
Autopsy data show that worms tend to aggregate more in some individuals than in others, due to poorly understood factors such as environmental, social, host genetic or immunological effects [17]. Epidemiological studies also show that those heavily infected are predisposed to this state [18]. To take account of such effects in our model, individuals in each age category are assigned a contact rate drawn from a gamma distribution with shape parameter α, which, via compounding across individual distributions, leads to a negative binomial distribution of worms within the total host population. It is important to note that the aggregation parameter, k, within the stochastic model, fluctuates in value over time, as a result of changes in the mean worm burden. In the deterministic model k is held fixed in value. The stochastic model more accurately mirrors observed patterns where k tends to decrease in value as prevalence declines under the impact of control measures [19]. The egg contribution to the infectious reservoir depends on the age-specific contact rate for each individual and is governed by a deterministic formulation. Treatment events are predetermined, they occur at time tj and the time step to the next treatment event is randomly drawn from an exponential distribution. The rate parameter for this distribution is given by the overall rate that any event happens. Which event occurs is drawn at random, on the basis of the relative magnitude of each individual event relative to the combined rate of all events. Table 1 provides a description of these rates. In this paper we consider 15 years of MDA and vaccination administration.
Most of the parameter values used in this paper are taken from within the ranges found in the literature (Table 2). However, the data for the age-specific contact rates of hosts within the infectious reservoir (β) and age-specific contribution of hosts to the reservoir are unknown. They are estimated by using MCMC method in parameter estimation from age intensity and prevalence curves as described in references detailed in the text and Table 2. Precise details of the model fitting procedure are described in previous publications [4,14,15,17].
In the numerical evaluations of the model’s behavior (stochastic simulations), we follow the WHO guidelines for the implementation of MDA. Starting with an untreated population, we administrate MDA over a 15-year period with coverage levels and treatment intervals based on the baseline prevalence. For low baseline prevalence in SAC, we treat once every 3 years; for moderate baseline prevalence in SAC, we treat once every 2 years and for high baseline prevalence in SAC, we treat once a year. The intensity of transmission is determined by R0 (the basic reproductive number) which varies for different baseline settings. When MDA alone is used as the treatment strategy, we simulate the following treatment strategies: (i) the WHO recommended treatment coverage of 75% SAC only; (ii) 60% of SAC only; (iii) 40% of SAC only and (iv) 85% of SAC and 40% of adults.
In this paper we consider an ideal case-perfect vaccine, meaning that the rate of infection and the rate of egg production are essentially reduced by 100%, which is comparable to the efficacy of the Sm-p80 vaccine in the baboon model. This efficacy considers the prevention of worm establishment, the fecundity falling dramatically in those few worms that establish, and the inability of eggs from these worms to hatch and release viable miracidia. Vaccination is given annually to the children with the pre-specified age of administration, and the coverage levels depend on the age group that is treated and the duration of vaccine protection. In various experimental settings Sm-p80 has demonstrated robust antibody titres in baboons for up to 5–8 years [10] suggesting a reasonably long duration of protection. In this paper we simulate scenarios where (i) the vaccine gives a 5 year duration of protection (from [10]) and (ii) an ideal scenario where the vaccine gives a 20 years of protection which is longer than the duration of treatment (15 years). It should be noted here that the same results will be obtained for vaccines with a duration of protection longer than 20 years as we are only calculating the probability of achieving the WHO goals within 15 years of initiating vaccination. Also, it should be noted that the vaccine decay rate is given by 1/ (duration of protection). Duration of vaccine protection has a direct impact on the vaccine administration schedule and the coverage levels required to have a significant impact. Here we consider the epidemiology of schistosome infections and the human host age-groups contributing most to parasite transmission. The aim is to cover children from ages 5–15 by vaccinating children in cohorts. We also analyze control strategies where the vaccine is given to younger children in their first year of life. The schistosomiasis vaccine will very likely be administered in conjunction with other vaccines already present in traditional immunization programmes (HPV, DTP). Therefore, the achievable coverage will typically match that achieved for one of the other co-administered vaccines. Vaccination coverage in the first year of life ranges between 85% and 91% at global level and reduces significantly in the following years (Table 3). The coverage levels for school age children vary between 60% and 70% and for out of school individuals this range is 40%-50% [24–27].
Based on these coverage levels, for a vaccine that provides a 20-year protection against schistosomiasis, we vaccinate at age 1 (early start) or age 5 (school start), with coverage levels of 85% and 60% respectively. For a vaccine that provides a 5-year duration of protection against infection, to ensure continuous protection, we vaccinate either at ages 1, 6 and 11 with coverage levels 85%, 60% and 70% respectively, or at ages 5,10 and 15 with coverage levels 60%, 70% and 45% respectively. In this case (5-year duration of protection) we have a 3-dose schedule of vaccination, similar to the HPV administration schedule.
We consider MDA and vaccination, alone or in combination, as control strategies, where treatment is delivered at random at each round within the population with a given coverage. In other words, we do not consider individual compliance to treatment [19] in these analyses and just assume the individuals treated or vaccinated are chosen at random at each round.
At the end of the treatment period, we calculate the probability of reaching WHO morbidity and elimination as a public health problem goal, by evaluating the fraction of SAC heavy-intensity infection prevalence (≤5% heavy-intensity infection in SAC for the morbidity goal and ≤1% heavy-intensity infection in SAC for the elimination as a public health problem goal). In our results we include the prevalence of infection (population having egg count threshold > 0) and prevalence of heavy-intensity infections (population having egg count threshold > 16). The probability of reaching the 5% and 1% WHO goals are calculated as the fraction of repetitions that reach the target, by averaging across 300 simulations (to ascertain the mean expectation of the stochastic model). A summary of the treatment strategies is presented in Fig 1.
In presenting the results of the stochastic model simulations for the various scenarios described above, the impact of the candidate vaccine and/or MDA is depicted by reference to the prevalence and mean intensity of S. mansoni infection in low, moderate and high transmission settings. For each treatment strategy, the prevalence of infection and prevalence of heavy-intensity infections in SAC and adults (the morbidity goal set by WHO), as well as the probability of achieving the WHO goals at all times t, until t = 15 (the end of control interventions), are assessed.
First MDA alone is examined as the treatment strategy, using the WHO targets for treatment of 75% coverage for SAC. The results are presented in Fig 2 and Table 4. Model simulations (based on the parameter values listed in Table 2) suggest that for low prevalence regions, the 5% morbidity goal in SAC can be achieved within 5 years of treatment, while the elimination as a public health problem goal in the total population can be achieved within 10 years of treatment.
Similarly, for moderate-prevalence regions, the 5% morbidity goal in SAC can be achieved within 5 years of treatment, whereas the 1% elimination as a public health problem goal can be achieved within 15 year of MDA treatment. Again, both goals will be achieved within 15 years with a probability of unity.
In high transmission regions, we can achieve the SAC 5% morbidity goal in 85% of the simulations. However, the 1% elimination as a public health problem goal in such high transmission (large R0 values) settings can be achieved in 35% of our simulations. In these settings, increasing the SAC coverage to > 75% and/or include other age bands in the treatment is highly desirable.
In low to moderate transmission settings, using the recommended target coverage of 75% for SAC, the SAC 5% morbidity goal can be achieved within 5 years of MDA. Given the difficulties countries with endemic infection are experiencing in achieving this level of coverage, SAC coverages between 40% and 60% were also examined to explore if it is still possible to achieve the WHO goals with 15 years of MDA treatment.
The impact of MDA decreases as SAC coverage declines as indicated in Table 4. The SAC 5% morbidity goal can be achieved within 5 years at 60% SAC coverage (in low to moderate settings). However, for the <1% heavy infection in the total population goal (= elimination as a public health problem) to be achieved within 15 years the probabilities of achieving this are 90% and 70%, respectively, in low and moderate transmission regions.
Lowering the SAC coverage to 40% is predicted to achieve the WHO goals in low transmission settings. However, in moderate transmission settings, the SAC 5% morbidity goal can be achieved within 15 years of treatment with probability of 0.9, but the 1% elimination as a public health problem goal is only achieved with probability 0.4 in that time. These results highlight the importance of using different MDA coverage levels in different transmission settings, as opposed to following the recommended 75% SAC coverage for all transmission levels.
In stochastic (and deterministic) models (and in the real world) there is always a chance that the prevalence of infection will bounce back after control measures cease since in some simulation runs the breakpoint in transmission is not crossed. It is therefore important to analyze the probability of true elimination (also known as ‘transmission interruption’) which results in the prevalence within the whole community in which control measures are introduced going to zero. As in previous studies [28] it is assumed that if the overall prevalence is less than 1% it is almost certain that transmission interruption has been achieved. We find that treating only 75% of SAC cannot interrupt transmission (see Fig 2A, 2C and 2E), since the reservoir of untreated people in the adult age classes is able to seed the whole population once control ceases at year 15.
As discussed earlier, in high transmission settings it is necessary to treat both SAC and adults. Here we present the simulation results for the scenario 85% of SAC and 40% of adults are annually treated with MDA. These results are summarized in Fig 3 which shows that with this approach the WHO goals can be achieved, although the probability of complete elimination by year 15 is still low (<0.3). Longer durations of treatment and/or more frequent treatment are required to increase this probability.
In this section, the effects of both vaccination coverage, and the average duration of protection provided by the vaccine, are examined. It should be noted that, based on the animal model results, we assume the vaccine is 100% efficacious.
In the previous two sections it is shown that the WHO 5% morbidity control goals can be achieved in low to moderate transmission settings if either MDA alone or vaccination alone are administrated in endemic regions. However, these goals, particularly the 1% elimination as a public health problem goal, are unlikely to be achieved in high transmission settings. Whether it is beneficial to combine both treatments together is examined in this section. In practice, this is a likely scenario since MDA will remain the main control options for many years to come (possibly 10 to 15 years) even if Phase I, II and III trials in humans of the new vaccine go smoothly.
The simulation results suggest that giving MDA to 75% of SAC and administrating vaccination with a wide range of coverage levels (see Figs 6 and 7, Tables 7 and 8), can reach the 1% elimination as a public health problem goal in high settings with a probability of nearly 0.55 and 0.82 for vaccines with durations of protection of 20 and 5 years, respectively. The 5% SAC morbidity goal is achieved in all transmission settings. Therefore, a vaccine that provides 5 years of protection and covers three age groups, can achieve the WHO 5% morbidity control and 1% elimination as a public health problem goals. However, for a vaccine that provides 20 years protection we need to increase MDA and vaccination coverage levels, or include other age categories in the vaccination programme, to increase the probability of achieving elimination as a public health problem (<1%) in high transmission settings. However, do note that the short duration vaccine must be delivered to multiple age groups. Over 15 years an individual may need three vaccinations (or 3 short courses of vaccination) to maintain protection. As such costs and delivery may be important issues with a short duration of protection vaccine.
The results presented in this paper are very sensitive to the values of certain parameters. The two most important are the negative binomial aggregation parameter k and the magnitude of transmission before control measures are initiated (the magnitude of R0). Using k = 0.24, λ = 0.24 in low transmission settings, the model cannot support endemic parasite populations when R0 is low. As a result, the model typically cannot reproduce endemic prevalences less than about 49%. The two possible causes are: (i) Diagnostic; due to poor sensitivity in the standard diagnostic test, measured prevalences may be much lower than the real values and (ii) model transmission structure; transmission may be confined to specific age groups as elimination is approached, giving a low community-level prevalence.
To manage this limitation, we use k = 0.04 value for low transmission setting and k = 0.24 for moderate to high transmission settings.
We have chosen the extreme baseline prevalences (just below 10% for low transmission settings and just below 50% for moderate transmission settings). For these values there is a high probability to achieve the WHO goals and hence lowering the baseline prevalence does not alter the outcome.
For a baseline prevalence between 50% and 58% (high transmission settings) we obtain qualitatively similar results with the ones produced in moderate settings. Therefore, for high transmission settings, we consider endemic regions with a baseline prevalence of around 62% (R0 = 3.5) which is a realistic upper bound of prevalence for S. mansoni in most endemic regions [29], [30]. In this study, we have used parameter values fitted to data collected in Iietune village in Kenya (refer to Table 2), but the same model and analysis can be used for other endemic regions. We should note here, that if the age-related contact rates and death rates are similar to the ones we have used, the results will be similar. If the prevalence of intensity is higher (lower) in SAC, the probability of achieving the WHO goals will be lower (higher) in these regions. These results are based on data for S. mansoni, but the analysis can be easily extended to S. haematobium.
A possible key parameter in the analysis and not included in our study is the buildup of acquired immunity. To date, there aren’t enough evidences to show the presence of immunity in S. mansoni and we have assumed that the shape of age-intensity of infection is influenced only by rate of exposure to infection. It will be of great importance, in the future, to extend our model so that we can explore the effect of acquired immunity on morbidity.
Currently schistosome control strategies suggested by WHO and widely implemented in endemic regions include mass drug administration of school aged children and adults in high transmission settings. The primary goal is morbidity prevention in SAC or morbidity elimination in populations in areas of endemic infection. Snail control, snail habitat alterations and improving water, sanitation and hygiene (WASH) are also recommended (there is little information on their efficacy), but MDA is the major route for morbidity control at present.
In this paper, we have extended the individual based stochastic age structured model developed by Anderson and colleagues, which is constructed on the template of an age structured deterministic model [13–15] where its predictions have been validated using observed infection trends under defined levels of MDA in a number of field settings [31]. We specifically extend past work to include the effect of a vaccine on parasite establishment. The aim has been to explore the impact a vaccine with an efficacy of 100% might have on control efforts to attain the WHO goals for morbidity control in SAC and morbidity elimination in the total population (but not infection). Different treatment and vaccination strategies have been considered in numerical analyses; namely: MDA alone, vaccination alone, or MDA plus vaccination combined. Analyses are conducted for three different transmission settings as defined by WHO on the basis of prevalence; low (<10% baseline prevalence among SAC), moderate (10–50% baseline prevalence among SAC) and high (≥50% baseline prevalence among SAC) settings. These transmission conditions at baseline are determined by the magnitude of R0, and, concomitantly, by the overall prevalence of infection and the average intensity of infection in defined community.
We find that the optimal strategy to control or eliminate morbidity depends on the transmission setting, vaccine coverage level achieved, the duration of vaccine protection and the timeline of vaccination in different age groupings of the human host.
In low prevalence settings, MDA alone or vaccination alone, with different levels of protection, can achieve the WHO goals with a probability of close to unity. Furthermore, our results show that treating just 40% of SAC with MDA alone can achieve the morbidity control goal and potentially elimination as a public health problem goal. This is an encouraging prediction considering the difficulties endemic regions are having in achieving the WHO recommended treatment coverage for SAC at 75%.
In moderate prevalence settings, treating 60% of MDA can achieve the morbidity goal with probability of unity and possibly the elimination as a public health problem goal with probability of 0.7. Increasing the SAC coverage to 75% increases the probability of elimination to 0.96. Vaccination with a duration of protection of 5 years can achieve the morbidity control goal within 5 years of treatment and elimination as a public health problem goal within 15 years. However, a vaccine with a longer duration of protection (20 years) achieves the morbidity goal with a probability of near unity, but the probability of elimination as a public health problem goal decreases to nearly 0.55.
In high transmission settings, we obtain the following outcomes: (i) the WHO recommended MDA treatment coverage for SAC at 75% can achieve the morbidity control goal with a probability of 0.85, but there is only a 0.35 chance that we can achieve the elimination as a public health problem goal. (ii) Vaccinating 85% of 1-year olds with a vaccine that provides 20 years of protection, can achieve the morbidity control goal with probability of 0.61, but it is very unlikely that the elimination as a public health problem goal will be achieved. (iii) increasing the vaccination coverage levels (vaccinating 85% of age 1, 60% of age 6, 70% of age 11 or vaccinating 60% of age 5, 70% of age 10 and 45% of age 15) and decreasing the duration of protection to 5 years, increases the probability of achieving the WHO goals. For the morbidity control this probability increases from 0.61 to 0.89, while the probability of the elimination goal increases from 0,06 to 0.22.
Thus, in high transmission settings, vaccination alone or MDA alone cannot achieve the elimination as a public health problem target. We can modify this outcome by vaccinating across bands of age classes (i.e. including adults). However, this may risk a high frequency of adverse effects due to past or present infection in vaccinated individuals. The best strategy in these circumstances is intensive MDA plus vaccination. Treating 75% of SAC with MDA and vaccinating 60% of age 5, 70% of age 10 and 45% of age 15 (duration of vaccine protection is 5 years) can achieve the morbidity control goal with probability of unity and the elimination as a public health problem with a probability of nearly 0.84. Alternatively, increasing the SAC coverage to 85% and including 40% of adults in the treatment plan, could achieve the WHO goals with a high probability. This outcome is in line with previous results found in [3], which has reported that including adults in the treatment strategy and increasing SAC coverage levels can lower the prevalence of heavy-intensity to below 1% in SAC.
Analysing the vaccine’s administration schedule (early start versus starting vaccination on entry to school), vaccinating 5-year olds may arguably be an easier strategy to implement than vaccinating 1-year olds. Relatively few individuals will have become infected by 4 years of age, but some have. The main argument in favour of the latter age is that it is easier to reach children for vaccination via school infrastructure/attendance. Alternatively, if the vaccine is safe for very young children (< 1 year of age) then the vaccine could just be part of the national immunization schedule for infants and young children. The other benefit of vaccinating at age 1 is to avoid morbidity induced by early infection in infancy. Given the long duration of vaccine protection, model simulations suggest little difference between the two strategies. This suggests that programmatic and cost issues will be most important in public health policy formulation for the use of the vaccine.
Comparing vaccination with a long duration of protection and MDA alone, we find that good coverage of MDA across bands of age classes (i.e. SAC) is predicted to have a greater and quicker impact than cohort immunization in all settings. However, we have used different coverage levels between these two treatment strategies with less people being vaccinated than are treated with MDA. On the other hand, a vaccine with a shorter duration of protection performs better (in terms of achieving the WHO goals) because we are treating more age groups.
Unless true elimination of transmission is achieved, treatment should not cease as there is a chance that the prevalence of infection will bounce back after cessation. True elimination is not achieved in any of the scenarios considered. Unless the treatment frequencies and coverage levels are increased considerably from the scenarios examined it is very unlikely that this goal will be achieved.
Factors such as individual adherence to treatment is not taken into consideration and we have assumed a random treatment adherence at each round for a given coverage level. The simulations may therefore be on the optimistic side since a proportion of the chosen individuals for a given coverage are likely to be nonadherent over many rounds of MDA [13], [21], [23], [32], [33]. It will be of great importance to have the relevant adherence data to make more accurate predictions.
The predictions presented in this paper depend on the assumptions made concerning the precise nature of the manner in which the intensity of infection varies by age in a given endemic region, the magnitude of R0 (= transmission intensity) reflected by the baseline prevalence prior to the introduction of control measures. It will be harder to achieve the WHO targets if infection in the very young (pre-SAC) and adults is high. We have used data for S. mansoni but the same methods of analysis can be applied for S. haematobium infection.
In summary, vaccination alone or in combination with MDA, proves to be an effective method to control or eliminate schistosomiasis as a public health problem. Achievement of the WHO goals for morbidity control and elimination depends on vaccine efficacy, on the duration of vaccine protection and on the coverage levels achieved in different age classes.
|
10.1371/journal.pgen.1002366 | Arabidopsis Homologs of Retinoblastoma-Associated Protein 46/48 Associate with a Histone Deacetylase to Act Redundantly in Chromatin Silencing | RNA molecules such as small-interfering RNAs (siRNAs) and antisense RNAs (asRNAs) trigger chromatin silencing of target loci. In the model plant Arabidopsis, RNA–triggered chromatin silencing involves repressive histone modifications such as histone deacetylation, histone H3 lysine-9 methylation, and H3 lysine-27 monomethylation. Here, we report that two Arabidopsis homologs of the human histone-binding proteins Retinoblastoma-Associated Protein 46/48 (RbAp46/48), known as MSI4 (or FVE) and MSI5, function in partial redundancy in chromatin silencing of various loci targeted by siRNAs or asRNAs. We show that MSI5 acts in partial redundancy with FVE to silence FLOWERING LOCUS C (FLC), which is a crucial floral repressor subject to asRNA–mediated silencing, FLC homologs, and other loci including transposable and repetitive elements which are targets of siRNA–directed DNA Methylation (RdDM). Both FVE and MSI5 associate with HISTONE DEACETYLASE 6 (HDA6) to form complexes and directly interact with the target loci, leading to histone deacetylation and transcriptional silencing. In addition, these two genes function in de novo CHH (H = A, T, or C) methylation and maintenance of symmetric cytosine methylation (mainly CHG methylation) at endogenous RdDM target loci, and they are also required for establishment of cytosine methylation in the previously unmethylated sequences directed by the RdDM pathway. This reveals an important functional divergence of the plant RbAp46/48 relatives from animal counterparts.
| Chromatin, made of histones and DNA, is often covalently modified in the nucleus, and modifications can regulate gene transcription. RNA molecules such as small-interfering or silencing RNAs (siRNAs) and antisense RNAs (asRNAs) can trigger silencing of gene expression in eukaryotes. We have found that in the flowering plant Arabidopsis, two homologous putative histone-binding proteins associate with a histone deacetylase and function in partial redundancy in chromatin-based silencing of various loci targeted by siRNAs or asRNAs. They act in partial redundancy to silence a development-regulatory gene that controls the transition to flowering and whose silencing is triggered by asRNAs, and genomic loci containing transposable and repetitive elements whose silencing is triggered by siRNAs via the siRNA–directed DNA Methylation (RdDM) pathway. In addition, these two genes function in maintenance of DNA methylation at RdDM loci and are also required for establishment of DNA methylation in the previously unmethylated sequences, revealing that histone modifications are partly required for DNA methylation. Our findings implicate that RNA–triggered transcriptional silencing involves repressive histone modifications such as deacetylation at a target locus.
| Cytosine DNA methylation is critical for stable silencing of transposable elements (TE) and repetitive sequences and for epigenetic regulation of endogenous gene expression in eukaryotes [1]–[3]. DNA methylation is thought to play an ancestral role in the defense against invasive DNA elements to maintain genome stability and integrity [1]–[3]. In the model plant Arabidopsis, cytosine methylation occurs in three different sequence contexts: CG, CHG and CHH. CG and CHG methylation are heritably maintained respectively by DNA METHYLTRANSFERASE 1 (MET1) and the plant-specific CHROMOMETHYLASE 3 (CMT3). CHH methylation is dynamically maintained through de novo methylation by the DOMAINS-REARRANGED METHYLTRANSFERASE 2 (DRM2) and the RdDM pathway [1].
RdDM is a mechanism by which siRNAs direct de novo cytosine methylation in all sequence contexts of target DNA sequences (complementary to the siRNAs). In Arabidopsis, the plant-specific RNA polymerase Pol IV is thought to initiate silencing by generating single-stranded RNA transcripts that are subsequently converted to double-stranded RNAs (dsRNAs) by RNA-DEPENDENT RNA POLYMERASE 2 (RDR2). dsRNAs are processed by DICER 3 (DCL3) to produce 24-nt siRNAs, which are subsequently loaded to an ARGONAUTE 4 (AGO4)-containing effector complex known as RISC (for RNA-Induced Silencing Complex). Through their interaction with long non-coding RNA transcripts from target loci, generated by the RNA polymerase Pol V, the loaded RISC complexes in association with DRM2 are targeted to RdDM target loci to establish cytosine methylation in CG, CHG and CHH contexts, leading to heterochromatin formation and transcriptional silencing [for reviews, see [2], [4].
siRNAs not only direct DNA methylation, but also trigger repressive histone modifications at RdDM target loci, including histone deacetylation, H3K9 dimethylation (H3K9me2) and H3K27 monomethylation (H3K27me1). Functional loss of the RISC component AGO4 causes a strong reduction in H3K9me2 at the endogenous RdDM target loci including transposable and repetitive elements [5], [6]. Furthermore, it has been shown that at RdDM target loci H3K27 monomethylation, a hallmark for silenced heterochromatin [7], is reduced upon loss of Pol V or AGO4 activity [5]. Together with DNA methylation, these repressive histone modifications establish a silenced heterochromatin state at RdDM target loci.
Histone modifications are involved in the control of DNA methylation. For instance, the H3K9 methyltransferase KRYPTONITE (KYP)/SUVH4 and SUVH4 homologs including SUVH2, SUVH5, SUVH6 and SUVH9, catalyze dimethylation of H3K9, which is recognized and bound by CMT3, leading to the maintenance of CHG methylation [1], [8]. Histone H3 lysine-4 (H3K4) demethylation is also involved in DNA methylation. Recent studies reveal that cytosine methylation is depleted in genomic regions with di- or tri-methylated H3K4 at a genome-wide level [9]; the H3K4 demethylase known as JMJ14/PKDM7B is required for H3K4 demethylation and CHG and CHH methylation at various RdDM target loci [10]. The histone deacetylase HDA6 deacetylates lysines of core histones including H3 and H4, and is required for cytosine methylation in transgenes and silenced rRNA genes [11]–[13]. Multiple genetic screens have revealed that HDA6 is critical for transgene silencing [13], [14]. Loss of HDA6 activity causes a substantial decrease of symmetric cytosine methylation and a moderate reduction in asymmetric CHH methylation in an RdDM-silenced transgene promoter, leading to the transgene reactivation [13]. In addition, disruption of HDA6 function gives rise to histone hyperacetylation and decreased CG and CHG methylation at silenced rRNA gene promoters [11]. HDA6 plays a dual role in silencing of these loci: deacetylating core histones and mediating cytosine methylation [15]. In this way, HDA6 and DNA methylation machinery are thought to work collaboratively to silence target loci.
The histone-binding proteins RbAp46 and RbAp48 are highly homologous WD40-repeat proteins and were first identified in mammalian cells as the tumor-suppressor Rb-binding proteins [16]. Subsequent studies revealed that RbAp46/48 is an integral subunit of multiple chromatin-modifying or -assembly complexes [for a review, see [16]]. RbAp46 forms a complex with the histone acetyltransferase called HAT1 that acetylates H4, whereas RbAp48 is a subunit of the Chromatin Assembly Factor-1 (CAF-1) complex that deposits nucleosomes. Both RbAp46 and RbAp48 are components of several histone deacetylase (HDAC) co-repressor complexes such as the Sin3 complex, which deacetylate core histones to repress target gene expression. In addition, RbAp46/48 is an integral subunit of the evolutionarily conserved Polycomb Repressive Complex 2 (PRC2)-like complexes that catalyze H3K27 trimethylation (H3K27me3), resulting in transcriptional repression. Recent studies have shown that RbAp46/48 functions as a histone (H3–H4 dimer)-binding protein [17]. It is believed that the RbAp46/48-containing complexes interact with histone substrates via either RbAp46 or RbAp48 [17].
RbAp46 and RbAp48 are evolutionarily conserved in animals and plants. There are five homologs in Arabidopsis known as MSI1–MSI5 (MSI for MULTICOPY SUPPRESSOR OF IRA1) [18]. Biological functions of MSI1 and MSI4/FVE have been identified, whereas the functions of MSI2, MSI3 and MSI5 are not known. MSI1 is required for proper vegetative development and plays an essential role in gametophyte and seed development [19], [20]. The MSI1 protein is an integral subunit of the conserved Arabidopsis CAF-1 complex, and has also been found in several PRC2-like complexes [18], [20], [21]. In addition, MSI1 directly interacts with the Arabidopsis Rb homolog (RBR), a key cell-cycle regulator [22], to repress MET1 expression in female gametogenesis, presumably resulting in a reduction in CG methylation [23]. In addition to MSI1, MSI4/FVE also interacts with a plant Rb homolog [24], but the biological implication of this interaction is unclear. FVE has been shown to repress expression of the central floral repressor FLC and several cold-responsive genes in Arabidopsis [24], [25]. FLC inhibits the transition from a vegetative to a reproductive phase (i.e. flowering), and loss of FVE function causes FLC de-repression, resulting in late-flowering [24], [25]. Previous studies reveal that fve mutations cause increased levels of histone acetylation at FLC chromatin [24], [26], indicating that FVE may be involved in deacetylation of FLC chromatin to repress FLC expression. However, recent studies show that loss of FVE function also gives rise to a strong reduction in PRC2-catalyzed H3K27me3, a repressive chromatin mark, in FLC chromatin [27]. Given that the human FVE homologs, RbAp46/48, are subunits of multiple histone-modifying complexes, the mechanisms underlying FVE-mediated transcriptional repression/silencing remain elusive.
FLC plays a crucial role in flowering-time regulation in Arabidopsis and FLC expression is affected by a range of chromatin modifiers (reviewed in refs 28,29). In most rapid-cycling (i.e. early flowering) Arabidopsis ecotypes, FLC expression is repressed or silenced by a group of proteins that mediate or trigger repressive histone modifications at the FLC locus, among which, in addition to FVE, are two conserved RNA 3′end-processing factors called CstF64 and CstF77, RNA-binding proteins known as FCA and FPA, a putative H3K4 demethylase FLOWERING LOCUS D (FLD), and a putative CLF (for CURLY LEAF)-containing PRC2-like complex [for reviews, see [28], [29]]. Furthermore, recent studies show that FLC antisense transcripts trigger FLC silencing [30], [31]. There are two groups of antisense transcripts resulting from alternative polyadenylation. CstF64 and CstF77 function together with FCA and FPA to promote polyadenylation of FLC antisense transcripts at a proximal site, triggering FLC silencing [30], [31]. FLD activity is required for, and acts downstream FCA and FPA in this silencing mechanism [30], [32]. It is believed that the 3′ processing at the proximal polyadenylation site on FLC antisense transcripts leads to co-transcriptional decay of the antisense RNA downstream the proximal site, which may generate aberrant RNAs and trigger repressive histone modifications such as FLD-mediated H3K4 demethylation, and consequent silencing [30].
FLC antisense transcript-triggered silencing is mechanistically different from the siRNA-triggered silencing of RdDM target loci, although both involve RNA molecules. So far, no siRNAs targeting the FLC genomic coding region or 5′ promoter have been detected in Arabidopsis. Consistent with this, knockout of siRNA-silencing pathway components such as Pol IV, Pol V, RDR2 or AGO4 has little effect on FLC silencing [33]. In addition, cytosines in genomic FLC in most Arabidopsis ecotypes are not methylated [34]. Thus, unlike RdDM-mediated silencing, cytosine methylation is not directly involved in FLC silencing. However, both silencing mechanisms require repressive histone modifications such as histone deacetylation, H3K4 demethylation, and/or H3K9 and H3K27 methylation, and involve chromatin silencing.
In this study, we explored the role for FVE and MSI5 in chromatin silencing of various loci targeted by siRNAs or asRNAs. We show that MSI5 acts in partial redundancy with FVE to silence FLC and endogenous RdDM target loci including FWA (containing two tandem repeats), AtMu1 (DNA transposon), AtSN1 (retrotransposon) and IG/LINE (intergenic transcripts). FVE and MSI5 associate with the histone deacetylase HDA6 to form HDAC complexes, and directly interact with the target loci, leading to histone deacetylation and transcriptional silencing. Together, these results show that FVE and MSI5 play an important role in the chromatin silencing of various loci targeted by siRNAs or asRNAs in plants.
FVE and MSI5 are Arabidopsis homologs of the human histone-binding RbAp46/48 [18], [24]. The amino acid sequence similarity between FVE and RbAp48 over the entire RbAp48 is 45%, and the similarity between MSI5 and RbAp48 over the entire RbAp48 is also 45%, whereas the identity between FVE and MSI5 over the entire MSI5 is 77% (Figure S1). The high degree of sequence conservation between MSI5 and FVE suggests that these two proteins may have a similar biochemical function.
Previous studies have shown that FVE represses the floral transition in Arabidopsis [24], [25]. We sought to address the biological functions of MSI5. Two loss-of-function mutants of MSI5 carrying insertional T-DNAs were identified, in which the full-length transcription of MSI5 was severely disrupted (Figure 1A and 1B). Grown in long days (LD; 16-hr light/8-hr dark), msi5-1 did not exhibit any visible phenotypes, whereas msi5-2 flowered slightly later than wild-type Col (Figure 1C and 1D), as measured by the developmental criterion of the number of leaves formed prior to flowering, from the primary apical meristem. In short days (8-hr light/16-hr dark), both mutants flowered moderately later than Col (Figure 1E). In both long and short days, msi5-2 flowered later than msi5-1, indicating that msi5-2 is a strong allele. We further confirmed that the moderate late-flowering of msi5-2 was indeed caused by the mutation in a complementation test in which the wild-type copy of MSI5 complemented the msi5-2 mutation (Figure 1F). To examine whether MSI5 acts redundantly with FVE to repress flowering, we introduced both msi5 alleles into fve mutants. In LDs, both msi5-1;fve and msi5-2;fve flowered later than the late-flowering fve mutants (Figure 1D). Of note, msi5-2;fve flowered with 56 leaves on average which is much later than fve (34 leaves on average) (Figure 1D). Hence, MSI5 functions redundantly with FVE to promote Arabidopsis flowering.
Vernalization (an extended period of cold exposure) promotes Arabidopsis flowering. We examined the effect of cold treatment on the flowering times of msi5-1;fve. The late flowering phenotypes of this double mutant were partially suppressed by 7-day cold treatment, and after 35 days of cold exposure, the mutant flowered similar to Col (Figure 1G). It is well known that vernalization largely represses FLC expression to accelerate flowering in Arabidopsis [28], [29]. These data indicate that the late-flowering phenotypes of msi5;fve is largely dependent on FLC and that the activities of MSI5 and FVE are not required for FLC repression by vernalization.
FVE has been shown to repress FLC expression [24], [25]. To examine whether the late flowering of msi5;fve was caused by FLC de-repression, we created an flc;msi5-2;fve triple mutant. In long days, the triple mutant flowered much earlier than msi5-2;fve, but still moderately later than flc (Figure 2A). Hence, the late-flowering of fve;msi5-2 is partly dependent on FLC. We further examined the flowering times of flc, fve;flc and flc;msi5-2;fve in short days, and found that the triple mutant flowered later than fve;flc and flc (Figure 2B), suggesting that FVE and MSI5 may repress other floral repressors to promote flowering, in addition to FLC.
Besides FLC, Arabidopsis has five FLC homologs including FLOWERING LOCUS M (FLM) (also known as MAF1), and MAF2-MAF5 (MAF for MADS BOX AFFECTING FLOWERING); these genes moderately repress flowering [35]–[37]. We quantified transcript levels of FLC and FLC homologs in Col, msi5, fve, and msi5;fve seedlings. FLC expression was slightly increased in msi5-2 compared to Col, whereas it remained unchanged in msi5-1 (Figure 2C). However, both msi5 alleles caused strong increases in FLC transcript levels in the fve background (Figure 2C). Furthermore, we found that in fve mutants both MAF4 and MAF5 were de-repressed, and this de-repression was enhanced upon loss of MSI5 function in the fve background, whereas MAF1, MAF2 and MAF3 expression remained unchanged upon loss of FVE and MSI5 function (Figure 2C). Together, these data show that MSI5 acts redundantly with FVE to repress the expression of MAF4 and MAF5, in addition to FLC, and promote the floral transition.
Recent genetic analyses have revealed that FVE is partly required for proper silencing of the RdDM target loci AtSN1 (retrotransposon) and AtMu1 (DNA transposon), although the underlying mechanism is unknown [38], [39]. This prompted us first to explore whether FVE plays a broad role in silencing of the RdDM target loci including TEs and repetitive elements. We examined the effect of loss of FVE function on the silencing of two other representative RdDM loci, FWA and IG/LINE. FWA, encoding a homeodomain-containing transcription factor that can repress flowering, is sporophytically silenced by cytosine methylation in two sets of tandem repeats containing a sequence related to a SINE (for Short Interspersed Nuclear Element) retroelement located in the 5′ region of FWA [40], [41]. IG/LINE is a spurious intergenic transcript initiated from a flanking solo-LTR (for Long Terminal Repeat) that functions as a promoter [42]. Upon loss of FVE function, FWA and IG/LINE were re-activated in the fve or fve;flc seedlings, respectively (Figure 3A, 3B). We asked whether MSI5 was required for silencing of RdDM target loci. To this end, we first quantified the transcript levels of AtSN1, AtMu1 and IG/LINE in msi5-2;flc and msi5-2;fve;flc seedlings. Both msi5-2 and msi5-2;fve were introduced into the flc background to exclude the possibility that FLC de-repression may affect reactivation of these loci. Loss of MSI5 function alone had little effect on silencing of these three loci; however, upon the combined loss of FVE and MSI5 function, all three loci were strongly re-activated to levels much higher than fve alone (Figure 3B–3D). Next, we measured FWA transcript levels in msi5 and msi5;fve seedlings [in the Col background; note that FLC upregulation does not affect FWA silencing [43]]. FWA is fully silenced in the msi5 seedlings (Figure 3A), but the msi5 mutations greatly enhanced FWA reactivation upon loss of FVE function (Figure 3A), like the situation in the other three loci. Together, these data suggest that MSI5 and FVE may play a broad role in silencing of transposable and repetitive elements in Arabidopsis genome, and that MSI5 functions redundantly with FVE to silence these elements.
These distinct four loci have a common feature, that is, their de novo silencing is established by the siRNA-triggered DNA methylation pathway [40], [42], [44], [45]. To test whether MSI5 and FVE were involved in silencing of TEs other than RdDM target loci, we examined the transcript levels of Ta3 in msi5 and/or fve mutant seedlings (in the flc background), which is a pericentromeric TE that is silenced independently of siRNAs [46]. Loss of MSI5 and/or FVE function did not cause Ta3 reactivation (Figure 3E). These data indicate that MSI5 and FVE may only be required for the silencing of RdDM-targeted TEs and repetitive elements.
FWA, AtMu1 and IG/LINE are silenced by cytosine methylation [40]–[42], [44]. We sought to determine whether MSI5 and FVE are required for cytosine methylation in these loci. Using bisulfite sequencing, we examined cytosine methylation at the tandem repeats (TRs), terminal inverted repeats (TIRs) and solo-LTR, respectively, in FWA, AtMu1 and IG/LINE in msi5 and/or fve mutant seedlings (note that these repeats generate siRNAs) (Figure 4A). At the FWA locus, CG methylation was slightly reduced, but a strong reduction in CHG and CHH methylation was observed, in msi5-2;fve compared to wildtype (mCHG: 14% in WT, but 4% in msi5-2;fve; mCHH: 7% in WT, but 2% in msi5-2;fve); neither CHG nor CHH methylation was affected in msi5-2, whereas upon loss of FVE function CHG and CHH methylation was moderately reduced (Figure 4B). At AtMu1, CG methylation was not affected, but CHG methylation was greatly reduced in msi5-2;fve (in the flc background); in addition, CHH methylation was moderately reduced upon loss of FVE and MSI5 function (Figure 4C). At solo-LTR, cytosine methylation in all contexts was reduced upon the combined loss of FVE and MSI5 function (Figure 4D). The reduction of non-CG methylation at the FWA, AtMu1 and solo-LTR loci was further confirmed using the methylation-sensitive restrictive endonucleases Fnu4HI or AluI (Figure S2). Together, these results show that MSI5 and FVE primarily mediate CHH and CHG methylation at RdDM target loci.
Recent studies show that symmetric CHG methylation is largely maintained by CMT3 in concert with the H3K9 methyltransferase KYP, whereas CHH methylation cannot be maintained, but is de novo methylated by the RdDM pathway [1]. Hence, we conclude that MSI5 and FVE are required for the de novo CHH methylation and maintenance of CHG methylation at the RdDM target loci. Cytosine methylation at both AtMu1 TIRs and solo-LTR causes transcriptional silencing. The reduction in cytosine methylation at the non-transcribed and siRNA-targeted regions (TIRs and solo-LTR) upon loss of FVE and MSI5 function, suggests that these two genes silence RdDM target loci partly by mediating DNA methylation in these loci.
Both FVE and MSI5 are required for de novo CHH methylation at the endogenous RdDM target loci. We sought to examine whether they could be involved in de novo cytosine methylation in all sequence contexts on previously unmethylated sequences using an FWA transgene assay. When an unmethylated FWA transgene is introduced into Arabidopsis genome, siRNAs from the endogenous FWA are able to target the transgene and through the RdDM pathway direct de novo cytosine methylation in all sequence contexts, leading to its silencing [40], [47]. Otherwise, ectopic FWA expression would give rise to a late-flowering phenotype [47], [48].
We introduced an FWA transgene [47], [48] into flc and flc;msi5-2;fve mutant backgrounds. Consistent with our previous finding that FWA transgene is de novo silenced in the flc background [43], T1 transformants of the flc background flowered only slightly later than flc (Figure 5A). By contrast, T1 transformants of the flc;msi5-2;fve mutant flowered much later than the non-transformed control (Figure 5A). Hence, MSI5 and FVE are required for de-novo silencing of the incoming FWA transgene.
We further examined the methylation state of FWA transgene in the flc and flc;msi5-2;fve backgrounds by bisulphite sequencing, and observed that CG methylation (a primary contributor for FWA silencing), was significantly reduced in the transgene upon the combined loss of FVE and MSI5 function (Figure 5B), in contrast to the slight reduction in CG methylation of the endogenous FWA (Figure 4B). In addition, non-CG methylation of FWA transgene was also reduced in flc;msi5-2;fve compared to the flc background. Together, these results show that FVE and MSI5 are required for the establishment of cytosine methylation in all sequence contexts of the newly introduced FWA transgene and thus de novo FWA silencing.
The functional redundancy of MSI5 with FVE raised the possibility that both genes could be expressed in the same tissues. To test this, we examined the spatial expression patterns of MSI5 and FVE using translational fusions to the reporter gene β-GLUCURONIDASE (GUS); the constructs contained the promoter plus part of the protein-coding region of FVE or MSI5. In seedlings, both MSI5 and FVE were preferentially expressed in shoot apices, root tips and leaf vasculature (Figure 6A–6D). In the reproductive phase, both genes were mainly expressed in styles and the junctions of ovary and receptacle (Figure 6E, 6F). In general, FVE-GUS was expressed at a level higher than that of MSI5-GUS. We confirmed that indeed, FVE transcript levels were much higher than those of MSI5 in both seedlings and floral buds (Figure 6G). This may partly explain why FVE plays a more dominant role in gene silencing than MSI5 does. Given the high protein-sequence homology of MSI5 with FVE, the overlapping expression patterns of these two genes provide an explanation for the functional redundancy of MSI5 with FVE.
The mammalian homologs of MSI5 and FVE, RbAp46/48, are subunits of several chromatin-modifying complexes involved in gene silencing such as HDAC co-repressor complexes; RbAp46/48 binds H3–H4 dimers and is thought to recognize and bind histone substrates in these complexes [17]. Using a candidate-gene approach, we explored whether FVE and/or MSI5 could associate with HDA6, an HDAC that has been shown to be involved in FLC repression, DNA methylation maintenance and gene silencing in Arabidopsis [11], [13], [49]. Bimolecular fluorescence complementation (BiFC) [50] was employed to examine whether MSI5 and FVE could associate with HDA6 in plant cells. A non-fluorescent N-terminal EYFP (for Enhanced Yellow Fluorescent Protein) fragment was fused to the full-length FVE and MSI5 individually, whereas a non-fluorescent C-terminal EYFP fragment was fused to the full-length HDA6. nEYFP-FVE and HDA6-cEYFP were simultaneously expressed in onion epidermal cells, and fluorescence was observed in the nuclei, reflecting the physical association of FVE with HDA6 in the nucleus (Figure 7A). Similarly, we also found that MSI5 associated with HDA6 in the nuclei of onion cells (Figure 7B).
Next, we performed protein pull-down assays to confirm the association of HDA6 with FVE and MSI5. Transgenic lines (T3 homozygotes) expressing MSI5-YFP-HA (in msi5-2 background) or FVE-FLAG (in fve background) were created. The MSI5 transgene was fully functional (Figure S3A), whereas the FVE-FLAG was partially functional (Figure S3B. Total proteins were extracted from transgenic seedlings and mixed with the purified GST-HDA6 from E.coli. HDA6 was able to pull down the MSI5 fusion and FVE-FLAG from the protein extracts (Figure 7C–7D). Thus, HDA6 can directly associate with FVE and MSI5 from Arabidopsis seedlings.
We further performed co-immunoprecipitation (co-IP) experiments to determine whether HDA6 is part of a complex with FVE or MSI5 in vivo. First, we created transgenic lines expressing a functional HDA6-FLAG (Figure S3C), and a line expressing a functional HA-FVE (Figure S3D). HDA6-FLAG-expressing plants were crossed to the HA-FVE line or the MSI5-YFP-HA line, and from the resulting F1 seedlings total proteins were extracted for co-IP analysis. Indeed, we found that anti-FLAG (recognizing HDA6-FLAG) immunoprecitated the MSI5 fusion protein and HA-FVE from the seedlings (Figure 7E, 7F). Of note, we detected only a small portion of the HA-FVE protein in the HDA6-FLAG immunoprecipitates from the F1 seedlings (note that no HA-FVE was immunoprecipitated from the seedlings expressing only HA-FVE); this is most likely due to an unstable association of FVE with HDA6. Taken together, these results led us to infer that FVE or MSI5 forms an HDAC complex with HDA6 in Arabidopsis.
Consistent with the HDA6 association with FVE and MSI5, recent studies show that HDA6, like FVE and MSI5, represses FLC, MAF4 and MAF5 expression to promote the floral transition [49]. It was of interest therefore to determine whether HDA6 also silences endogenous RdDM target loci. We measured transcript levels of FWA, AtMu1, AtSN1 and IG/LINE in WT (Col) and hda6 seedlings. Indeed, loss of HDA6 activity, like loss of MSI5 and FVE function, caused re-activation of all four loci (Figure 8A–8C). Thus, HDA6, like MSI5 and FVE, is required for silencing of the RdDM target loci.
We further examined cytosine methylation state in the FWA, AtMu1 and solo-LTR loci in hda6 seedlings. At FWA, loss of HDA6 function, like of loss of FVE and MSI5 function, caused a reduction in CHG and CHH methylation (Figure 8D). At solo-LTR, cytosine methylation in all sequence contexts was reduced in hda6 compared to wildtype (Figure 8D), similar to the situation in the msi5-2;fve mutant (Figure 4D). In addition, at AtMu1, upon loss of HDA6 function cytosine methylation in all sequence contexts was reduced (Figure 8D). Together, these results show that HDA6, like MSI5 and FVE, is required for cytosine methylation at these RdDM target loci.
To investigate whether FVE could bind to the chromatin of genes that exhibit altered expression in fve mutants, we performed chromatin immunoprecipitation (ChIP) experiments using the HA-FVE line. Using anti-HA antibodies, we immunoprecipitated DNA fragments from HA-FVE-expressing seedlings (wild-type Col was used as a negative control), and quantified DNA fragments from FLC [the 5′ Intron I region that is essential for FLC silencing; see [26]], AtMu1 (the promoter region immediately downstream the 5′ TIR), solo-LTR and FWA (a region in the silencing tandem repeats) (Figure 9A). Compared with the control, the abundances of HA-FVE protein associated with FLC, solo-LTR and AtMu1 chromatin increased in the HA-FVE line (Figure 9B). In addition, FVE was strongly enriched in FWA (Figure 9B). Of note, the moderate enrichment of HA-FVE at AtMu1 and solo-LTR is likely due to that in the ChIP experiments, anti-HA may not bind effectively to the single HA epitope tag fused to FVE at these TE-containing loci. Taken together, these data suggest that FVE directly interacts with FLC and the three RdDM loci.
FVE or MSI5 forms an HDAC complex with HDA6 and may mediate histone deacetylation for transcriptional silencing. Hence, we examined H3K9 and K14 acetylation state in FVE and MSI5 targets in WT and msi5-2;fve seedlings. H3K9K14 acetylation levels were moderately increased in the AtMu1 promoter region and solo-LTR, and were strongly elevated in the 5′ Intron I of FLC and FWA tandem repeats upon combined loss of FVE and MSI5 function (Figure 9C), consistent with FVE enrichments at these regions. Together, these data suggest that FVE and MSI5 are required for histone deacetylation at FLC and the three RdDM target loci. Given the association of HDA6 with FVE and MSI5, these two proteins may act as part of the HDA6-containing HDAC complexes to mediate deacetylation of target loci and silence their expression.
In this study, we show that MSI5 acts redundantly with FVE to silence various loci targeted by siRNAs or asRNAs, including FLC and RdDM target loci. Both FVE and MSI5 form HDAC complexes with HDA6 to mediate histone deacetylation in their target loci. In addition, these two genes function in de novo CHH methylation and maintenance of symmetric cytosine methylation at the endogenous RdDM target loci, and are also required for the establishment of cytosine methylation of the previously unmethylated sequences. Our findings suggest that FVE or MSI5 acts in the context of HDA6-containing co-repressor like complexes to mediate chromatin silencing of developmental genes and RdDM loci of transposable and repetitive elements.
FVE and MSI5 are required for cytosine methylation at three representative RdDM target loci. These two proteins may mediate cytosine methylation partly via FVE/MSI5-HDA6 complex-catalyzed histone deacetylation. HDA6-catalyzed histone deacetylation is known to be required for cytosine methylation in transgenes and endogenous rRNA genes. It has been shown that silencing of a transgene promoter targeted by RdDM requires HDA6 [13]. The absence of HDA6 activity causes a substantial decrease in symmetric CG and CHG methylation and a moderate decrease in CHH methylation in the promoter region, leading to reactivation of the silenced transgene [13]. Recent studies have shown that HDA6 exhibits a complex interrelationship with cytosine methylation in silenced rRNA genes: loss of HDA6 activity leads to a decrease in symmetric CG and CHG methylation and de-repression of intergenic transcription resulting in overproduction of siRNAs and consequent increase in CHH methylation [11]. Thus, HDA6 is not essentially required for CHH methylation at rRNA genes; however, this does not exclude that HDA6 is still partly involved in this methylation at loci other than rRNAs. A very recent study has revealed that HDA6 is required for both CHH and CHG methylation at a few loci, and partly for CG methylation in some of these loci in Arabidopsis [51].
We have found histone hyperacetylation, loss of cytosine methylation and transcriptional reactivation at the RdDM target loci FWA, AtMu1 and solo-LTR (IG/LINE) upon combined loss of MSI5 and FVE function. This raises the possibility of that the loss of cytosine methylation might result from transcriptional activities at these loci. However, the reasons below argue for that the loss of DNA methylation at least partly causes re-activation of these silent loci. First, both TIRs and solo-LTR are non-transcribed regions. Second, the examined regions with a loss of cytosine methylation including TRs in FWA, TIRs in AtMu1 and solo-LTR in IG/LINE have been shown to generate siRNAs that trigger DNA methylation at these regions leading to transcriptional silencing [40], [42], [44], [45].
In this study, we have revealed that MSI5 and FVE act to silence various loci targeted by siRNAs or asRNAs. This raises the possibility of that these genes could be involved in the production of siRNAs and asRNAs for transcriptional silencing. Recent studies show that loss of FVE function does not affect the production of neither asRNAs from the FLC locus nor the siRNAs from AtMu1, solo-LTR and AtSN1 [30], [38]. We have measured the levels of Pol V-dependent silencing scaffold RNAs from solo-LTR and AtSN1 [5], [45], in mutant seedlings carrying a knockout allele of FVE and/or MSI5, and observed that the loss of FVE and/or MSI5 function does not affect the production of these RNAs (Figure S4). It has also been shown that HDA6 is not involved in siRNA production from the promoter region of a silencing-reporter transgene [13]. Based on these findings, we infer that FVE/MSI5-HDA6 complexes act downstream of or in parallel to siRNA/asRNA production for transcriptional silencing of FLC and RdDM loci. De novo CHH methylation and maintenance of symmetric CHG methylation at the endogenous RdDM loci are catalyzed by the DNA methyltransferases DRM2 and CMT3, respectively, and MSI5/FVE-HDA6 complex-mediated histone deacetylation is expected to facilitate these cytosine methylations. In animals only the CG dinucleotide is methylated [1], hence, the mammalian homologs of FVE and MSI5, RbAp46/48, certainly do not play any roles in CHG and CHH methylation. Our findings on the roles for FVE and MSI5 in cytosine methylation reveal a functional divergence of the plant RbAp46/48 relatives from animal counterparts.
In Arabidopsis, the RdDM pathway controls the establishment of cytosine methylation in the previously unmethylated sequences including CG, CHG and CHH contexts. DRM2 functions as part of the AGO4-containing RdDM effector complex to methylate cytosines in all sequence contexts [52]. Previously, it has been shown that de novo cytosine methylation of the FWA transgene newly introduced into Arabidopsis genome requires the entire RdDM-pathway components including RDR2, DCL3, AGO4, Pol IV and DRM2 [47]. In this study, we have found that FVE and MSI5 are also required for the establishment of cytosine methylation at the incoming FWA. It is likely that the MSI5/FVE-HDA6 complex-mediated histone deacetylation contributes to the establishment of a repressive chromatin environment that promotes DRM2 to catalyze cytosine methylation at the previously unmethylated sequences.
The mammalian FVE and MSI5 homologs, RbAp46/48, are subunits of multiple chromatin-modifying complexes such as hHAT1 (involved in gene activation), PRC2 and HDAC co-repressor complexes [16]. So far, FVE and MSI5 have been found to be involved only in transcriptional silencing. Previous studies have shown that loss of FVE function causes reduced H3K27me3 and histone hyperacetylation at the FLC locus, indicating that FVE might act in the context of a PRC2-like complex and/or an HDAC co-repressor-like complex for FLC silencing under normal growth conditions [24], [26], [27]. A very recent study suggests that FVE may be part of a CLF-containing PRC2-like complex to deposit H3K27me3 in FLC and silence its expression [53]; however, the findings described below argue against this notion. Firstly, a gain-of-function clf allele clf-59 [27], suppresses FLC expression in the msi5-2;fve double mutant, resulting in early flowering (Figure S5A); hence, CLF functions independently of FVE and MSI5 to regulate FLC expression. Secondly, to genetically test whether FVE could be part of a CLF-PRC2 complex, we created a clf;fve double mutant in which both clf (clf-29) and fve (fve-4) are null loss-of-function alleles [19], [24], examined FLC de-repression in clf, fve and clf;fve seedlings, and found that CLF and FVE act synergistically to silence FLC expression (Figure S5B), suggesting that FVE may not be part of the CLF complex. Thirdly, in this study we found that FVE and MSI5 silence RdDM target loci, which typically lack of H3K27me3 deposited by PRC2 complexes [54]. Lastly, we carried out co-IP experiments to determine whether FVE and CLF could be in a complex using F1 seedlings expressing a fully-functional GFP-CLF [55] and the HA-FVE fusion, but did not detected an association of CLF with FVE in seedlings (Figure S6). Together, these findings suggest that FVE and MSI5, unlike RbAp46/48, may not act as part of PRC2-like complexes to silence target-locus expression. At the FLC locus, both FVE/MSI5-HDA6 and CLF-PRC2 complexes directly repress its expression [56], and may act in concert to establish a repressive chromatin environment at FLC for its transcriptional silencing (see Figure 10 as described next).
In mammals, RbAp46/48 is an integral subunit of Class I HDAC co-repressor complexes [57], [58], and several of these complexes such as the BRAF-HDAC complex contain the H3K4 demethylase Lysine-Specific Demethylase 1 (LSD1) [59], a mammalian homolog of the Arabidopsis FLD [60]. HDA6, like the Class I HDACs, is an RPD3 (for Reduced Potassium Deficiency 3)-type histone deacetylase [13]. A recent study has revealed that HDA6 forms a complex with FLD to repress FLC expression and promote flowering [61]. In this study, we have found that HDA6 also forms a complex with FVE or MSI5. Together, these findings led us to infer that HDA6 and FLD form an HDAC co-repressor like complex with FVE or MSI5. Consistent with this, in a co-IP experiment using a line expressing a fully functional FLD-myc [62] and the MSI5-YFP-HA fusion, we have confirmed that indeed FLD is in a complex with MSI5 in Arabidopsis seedlings (Figure S7). Furthermore, like MSI5, FVE and HDA6, FLD is also required for the silencing of FLC, FLC homologs and the RdDM target loci including AtMu1, AtSN1 and IG/LINE [38], [61]. Taken together, these Arabidopsis homologs of the mammalian Class I HDAC co-repressor complex components may form HDAC co-repressor like complexes to silence developmental genes and TEs.
The loss of HDA6 function appears to cause a greater reactivation of AtMu1 than that upon the combined loss of FVE and MSI5 function, indicating that HDA6 silences this locus partly independent of MSI5 and FVE. One explanation is that MSI1, MSI2, and/or MSI3 may also participate in HDA6-mediated silencing. HDA6 plays multiple roles in Arabidopsis development. In addition to the acceleration of floral transition, HDA6 is also involved in plant senescence and acts redundantly with its homolog HDA19 to repress embryonic traits in vegetative growth [49], [63], [64], in which FVE/MSI5 appears not to be involved (data not shown). These observations indicate that HDA6 may act partially independent of FVE and MSI5 to silence developmental genes in Arabidopsis.
MSI5 acts redundantly with FVE to silence the developmental gene FLC and RdDM loci targeted by the silencing triggers asRNAs or siRNAs, respectively. Transcriptional silencing of these loci requires repressive chromatin modifications including histone deacetylation, H3K4 demethylation, H3K9 methylation, and/or H3K27 methylation.
At the FLC locus, the transcriptional silencing is triggered by FLC antisense transcripts and/or aberrant RNA molecules derived from the 3′ processing of asRNAs [30]. In the chromatin silencing at FLC, the targeted 3′ processing of FLC antisense transcripts produces RNA triggers that lead to the recruitment of repressive histone-modification activities on FLC chromatin. As noted above, FLD and HDA6 form a co-repressor like complex with FVE or MSI5. Consistent with this, FLD is required for both H3K4 demethylation and histone deacetylation on FLC chromatin [26], [32], [43]. Moreover, the loss-of-function fld and fve mutations act largely non-additively to delay the floral transition (caused by FLC de-repression) [38]. As illustrated in Figure 10A, it is very likely that the RNA molecules may trigger the recruitment of FLD-FVE/MSI5-HDA6 complexes to FLC chromatin, resulting in repressive histone deacetylation and H3K4 demethylation. In addition, the HDA6 complexes are expected to act collaboratively with the CLF-PRC2 complex that deposits repressive H3K27me3 at FLC. Together, these histone modifiers establish a repressive chromatin environment at the FLC locus leading to heterochromatin-like formation and consequent FLC silencing (Figure 10A).
At the RdDM target loci of transposable and repetitive elements, both cytosine methylation and repressive chromatin modifications such as histone deacetylation, contribute to transcriptional silencing. For instance, FWA silencing is typically caused by symmetric CG methylation [40], [41]. We have found that loss of FVE and MSI5 function leads to histone H3 hyperacetylation at the endogenous FWA and ectopic FWA activation in sporocytes, but only a slight reduction in CG methylation. This suggests that MSI5/FVE-HDA6-mediated histone deacetylation plays a direct role in FWA silencing, in addition to promoting CHG and CHH methylation at this locus. In the chromatin silencing at the RdDM target loci (Figure 10B), MSI5/FVE-containing complexes mediate histone deacetylation and possibly, H3K4 demethylation, on one hand, directly represses target locus expression, and on the other hand, together with H3K9 dimethylation and/or H3K27 monomethylation, establish a repressive chromatin environment that promotes cytosine methylation (mainly CHG and CHH methylation), which may reinforce the repressive histone modifications. Together, these modifications lead to silent heterochromatin formation and consequent transcriptional silencing. FLD, the putative H3K4 demethylase, may act as part of the FVE/MSI5-HDA6 complexes to silence some of the RdDM loci such as AtMu1, IG/LINE and AtSN1 because FLD has been shown to be required for silencing of these loci. Previously we have observed that FLD appears not to be required for FWA cytosine methylation and silencing, but two FLD homologs known as LDL1 and LDL2 mediate FWA silencing [43]. It is likely that HDA6 and FLD may form a co-repressor like complex with FVE or MSI5 to silence certain RdDM target loci, whereas at some other loci, HDA6 and FVE/MSI5 may form a complex with other components for transcriptional silencing.
Arabidopsis thaliana fve-4 [24], flc-3 [65], hda6/axe1-5 [49], clf-29 [19] and clf-59 [27] were described previously. The msi5-1 (Salk_004926) and msi5-2 (Salk_116714) alleles were isolated from the SALK collection [66]. Plants were grown under cool white fluorescent lights in long days (16-hr light/8-hr dark) or short days (8-hr light/16-hr dark).
The full-length coding sequences for HDA6, FVE and MSI5 were translationally fused with either an N-terminal EYFP fragment in the pSAT1A-nEYFP-N1/pSAT1-nEYFP-C1 vectors and/or a C-terminal EYFP fragment in the pSAT1A-cEYFP-N1/pSAT1-cEYFP-C1-B vectors (www.bio.purdue.edu/people/faculty/gelvin/nsf/index.htm). Using the Helium biolistic gene transformation system (Bio-Rad), onion epidermal cells were transiently co-transformed by appropriate plasmid pairs as indicated in Figure 7. EYFP fluorescence in the onion cells was observed and imaged using a Zeiss LSM 5 EXCITER upright laser scanning confocal microscopy (Zeiss) within 24–48 hrs after bombardment.
To create HA-FVE fusion, the full-length FVE coding sequence (1.5 kb) was first cloned into the entry vector pENTR4 (Invitrogen), and subsequently, the FVE fragment was inserted downstream of the 35S promoter and the single HA epitope in the pEarlyGate 201 vector [67] via gateway technology (Invitrogen), resulting in the p35S-HA-FVE plasmid. For MSI5-YFP-HA construction, the full-length MSI5 coding sequence (1.5 kb) was inserted downstream of the 35S promoter, but upstream of YFP followed by the single HA epitope in the pEarlyGate 101 vector [67], resulting in the p35S-MSI5-YFP-HA plasmid. To construct GST-HDA6, the full-length FVE coding sequence was cloned into downstream of GST in the protein expression vector pGEX-4T-1. To construct FVE-GUS, a 4,073-bp FVE genomic fragment (from −1,886 to +2,187; A of the start codon as +1) including a 1.9-kb native promoter plus a 2.2-kb genomic coding region was inserted upstream of the GUS reporter gene in the pMDC162 vector [68]; the genomic coding sequence was in frame with GUS. For MSI5-GUS construction, we inserted a 2,145-bp MSI5 genomic fragment (from −438 to +1,707) into upstream of the GUS reporter gene in pMDC162; the genomic coding sequence of MSI5 was in frame with GUS. For the construction of a binary plasmid harboring a wild-type copy of MSI5, a 5.1-kb genomic fragment including the 5′ promoter (1.6 kb), genomic coding sequence (3.2 kb) and 3′ end (0.3 kb), was cloned into pBGW [69]. To clone the gain-of-function clf-59 allele with a single point mutation [27], a 6.5-kb genomic fragment of clf-59 consisting of a 1.3-kb 5′ promoter, 4.5-kb genomic coding sequence and 0.7-kb 3′ end, was amplified from a Ws background and cloned into the binary vector pHGW [69].
Total RNAs were extracted from aerial parts of 10-d-old seedlings grown in long days as described previously [43]. The total RNAs were subsequently treated with ‘TURBO DNA-Free’ (Ambion) to remove residual genomic DNA. After reverse transcription, the real-time quantitative PCR was carried out on an ABI Prism 7900HT sequence detection system as previously described [43]. Primers used to amplify the cDNAs of FLC, FLM, MAF2-5, IG/LINE, and TUB2 (At_5g62690) have been previously described [38], [70]. The primer pairs used for MSI5, FVE, FWA, AtMu1, AtSN1 and Ta3 amplification are specified in Table S1. Each sample was quantified in triplicate and normalized to the endogenous control TUB2. Bars indicate standard deviations of triplicate measurements.
DNA was extracted from 10-d-old seedlings grown in long days, and subsequently, approximately 0.2-µg genomic DNA from each genotype was treated with bisulfite using the EpiTect Bisulfite kit (Qiagen) according to the manufacturer's instruction. The bottom strands of the endogenous FWA (tandem-repeat region), solo-LTR and AtMu1 (the 3′ terminal-inverted-repeat region) were amplified by PCR, and cloned into the T-Easy vector (Promega). The primers used for the endogenous FWA amplification has been described previously [43], and the primer pairs for solo-LTR and AtMu1 amplification are specified in Table S1. Analysis of cytosine methylaion of the FWA transgene in T1 transformants of flc and flc;msi5;fve was performed as described previously [40], [43].
Total proteins were extracted from 10-d old seedlings. Briefly, 0.5-g seedlings were ground in liquid nitrogen and homogenized in 1.0-ml extraction buffer (50 mM Tris-HCl pH 7.4, 100 mM NaCl, 10% glycerol, 0.1% NP-40, 1.0 mM PMSF) supplemented with 1× Roche protease inhibitor (without EDTA). Subsequently, the GST-HDA6 or GST proteins affinity-purified from the E.coli strain BL21 (DE3) together with the glutathione-linked resins (Sigma) were added into 1.0-ml protein extracts and incubated for 4 hrs at 4°C. The protein pull-downs were analyzed by immunoblotting using anti-HA (Roche, Cat#: 12-013-819-001) or anti-FLAG (Sigma, Cat#: A8592).
Immunoprecipitation experiments were performed as described previously [71]. Briefly, 0.4-g seedlings were harvested and ground in liquid nitrogen, and subsequently, total proteins were extracted and immunoprecipitated with anti-FLAG M2 affinity gel (Sigma, Cat#: A2220). Proteins in the immunoprecipitates were detected by western blotting with anti-FLAG (Sigma, Cat#: A8592) or anti-HA (Roche, Cat#: 12-013-819-001).
ChIP experiments were performed with 10-d-old seedlings largely as previously described [72], [73]. Briefly, nucleus fraction was isolated, and subsequently, immunoprecipitations were carried out using the polyclonal anti-acetylated histone H3 (Lys 9 and Lys 14) (Millipore, Cat#: 06-599B) or anti-HA (Sigma, Cat#: H6908). Quantitative measurements of genomic fragments of FLC, AtMu1, solo-LTR, MAF3 and TUB2 (as the internal normalization control) were performed using SYBR Green PCR master mix (Applied Biosystems). Quantitative measurements of FWA genomic regions and ACTIN2 (At_3g18780; served as the internal normalization control for FWA enrichment) were performed on an ABI Prism 7900HT sequence detection system using TaqMan MGB probes (FAM dye–labeled) as described previously [43]. Each ChIP sample was quantified in triplicate. The primers used to amplify FLC and TUB2 were described previously [43], and the primer pairs for AtMu1, solo-LTR and MAF3 amplification are specified in Table S1. Rationale for calculation of the fold enrichment of HA-FVE in the HA-FVE line over the control line (Col) is as follows: in the HA-FVE line a gene of interest (eg. FLC) was first normalized to TUB2 or ACTIN2, and in the control line the gene of interest was similarly normalized; subsequently, the normalized value from the HA-FVE line was further normalized by the value from the control line to obtain a value of fold enrichment for the gene of interest. A similar rationale was adopted for the calculation of fold enrichment of acetylated H3 in msi5-2;fve over Col.
|
10.1371/journal.pgen.1007588 | Numerous recursive sites contribute to accuracy of splicing in long introns in flies | Recursive splicing, a process by which a single intron is removed from pre-mRNA transcripts in multiple distinct segments, has been observed in a small subset of Drosophila melanogaster introns. However, detection of recursive splicing requires observation of splicing intermediates that are inherently unstable, making it difficult to study. Here we developed new computational approaches to identify recursively spliced introns and applied them, in combination with existing methods, to nascent RNA sequencing data from Drosophila S2 cells. These approaches identified hundreds of novel sites of recursive splicing, expanding the catalog of recursively spliced fly introns by 4-fold. A subset of recursive sites were validated by RT-PCR and sequencing. Recursive sites occur in most very long (> 40 kb) fly introns, including many genes involved in morphogenesis and development, and tend to occur near the midpoints of introns. Suggesting a possible function for recursive splicing, we observe that fly introns with recursive sites are spliced more accurately than comparably sized non-recursive introns.
| The splicing of RNA transcripts is an essential step in the production of mature mRNA molecules, involving removal of intron sequences and joining of flanking exon sequences. Introns are usually removed as a single unit in a two-step catalytic reaction. However, a small subset of introns in flies are removed via splicing of multiple distinct consecutive segments in a process known as recursive splicing. This pathway was thought to be quite rare since intermediates of recursive splicing are seldom detected. In this study, we developed three new computational approaches to identify sequence reads, read pairs and patterns of read accumulation indicative of recursive splicing in Drosophila melanogaster cells using data from sequencing of nascent RNA captured within minutes after transcription. We used these methods to identify hundreds of previously unknown sites of recursive splicing, occurring commonly in fly introns longer than 40kb and often in genes involved in morphogenesis and development. We observed that recursive splicing is associated with increased splicing accuracy of long introns, which are otherwise often spliced inaccurately, potentially explaining its widespread occurrence in long fly introns.
| RNA splicing is a crucial step in the mRNA lifecycle, during which pre-mRNA transcripts are processed into mature transcripts by the excision of intronic sequences. Introns are normally excised as a single lariat unit. However, some introns in the Drosophila melanogaster genome are known to undergo recursive splicing, in which two or more adjacent sections of an intron are excised in separate splicing reactions, each producing a distinct lariat [1,2]. Recursively spliced segments are bounded at one or both ends by recursive sites, which consist of juxtaposed 3' and 5' splice site motifs around a central AG/GT motif (with “/” indicating the splice junction) [1,3]. This mechanism appears to be restricted to very long Drosophila introns [3,4]. However, because recursive splicing yields an exon ligation product identical to that which would have been produced from excision of the intron in one step, the genome-wide prevalence and function of recursive splicing have been difficult to ascertain [3,4].
Recursive splicing was initially observed in the splicing of a 73 kb intron in the Drosophila Ultrabithorax (Ubx) gene, where the intron is removed in four steps through intermediate splicing of the 5' splice site to two microexons and one recursive site before pairing with the proper 3' splice site [1]. Bioinformatic searches for recursive sites predicted a couple hundred possible recursive sites in Drosophila, predominantly in introns larger than 10 kb [3], but sites in only four introns, all from developmentally important genes (Ubx, kuzbanian (kuz), outspread (osp), and frizzled (fz)), could be experimentally validated [1–3]. Biochemical characterization showed that recursive splicing is the predominant processing pathway for splicing of these introns, which are generally constitutively spliced [1–4]. More recently, an analysis by Duff and coworkers of all ~10 billion RNA-seq reads generated by the Drosophila ModENCODE project identified 130 recursively spliced introns in flies [4]. Using this larger catalog of recursive sites, they confirmed that recursive splicing is a conserved mechanism to excise constitutive introns, requires canonical splicing machinery, and only occurs in the longest 3% of Drosophila introns [4]. Similar analyses of mammalian RNA-seq datasets have resulted in the identification of just a handful of recursively spliced introns, mostly in genes involved in brain development, despite the greater abundance of long introns in vertebrate genomes [5].
The scarcity of validated examples suggests that recursive splicing is quite rare, even in Drosophila. However, the transient nature of recursive splicing intermediates makes it difficult to detect evidence for recursive splicing using standard RNA-seq data. Support for recursive splicing has come from RNA-seq reads that span a junction between a known splice site and a putative recursive splice site internal to an intron, or from observation of a sawtooth pattern of reads resulting from the splicing out of recursive segments [4,5]. Previous studies using polyA-selected RNA-seq data– which derive predominantly from mature transcripts– had limited ability to detect such evidence. However, nascent RNA sequencing, which profiles pre-mRNA transcripts shortly after they are transcribed, should enable much more efficient capture of reads from intermediates of splicing, including recursive splicing. Using such data should allow for more unbiased and systematic discovery of recursive splicing.
To globally detect transient splicing intermediates indicative of recursive splicing, we applied novel computational approaches to high-throughput sequencing data from short time period metabolic labeling of RNA. This approach detected about four times as much recursive splicing as had been previously observed. This expanded catalog of sites and associated analyses suggests a function for recursive splicing in improving splicing accuracy.
Pre-mRNA splicing can initiate immediately after transcription of an intron is completed, and can occur in as short a time as one or a few seconds [6–9]. Since recursive splicing involves the splicing of intermediate intronic segments, it may begin soon after the transcription of the first intronic recursive site. Thus, to have the greatest chance of capturing recursive splicing intermediates, it is essential to capture nascent transcripts as soon as possible after transcription, before introns have been fully spliced. Here, we used nascent RNA sequencing data from our recent study, which used incorporation of a metabolic label to isolate RNA at short time points after transcription [9]. The experimental approach to collect these data involved 5, 10, or 20 min labeling with 4-thiouridine (4sU) in Drosophila S2 cells and 4sU biotinylation to selectively isolate nascent RNA, followed by RNA sequencing with paired-end 51 nt reads [9]. These data were complemented by steady state RNA-seq data representing predominantly mature mRNA (Methods). The progressive labeling strategy used for these data results in isolation of transcripts that initiated during the labeling period, in addition to transcripts that were elongated during this period but initiated prior to the addition of the label [9]. While this likely does not significantly bias the distribution of fragment lengths sequenced, there is an overall 5' to 3' bias of reads across the entire transcript.
We hypothesized that this high-coverage nascent RNA data would more readily identify recursive sites and better characterize the prevalence of recursive splicing. For this purpose, we used a computational pipeline to detect three key signatures of recursive splice sites (Fig 1). First, we used a custom python script to search for splice junction reads derived from putative recursive sites (RatchetJunctions), as previously described (Methods; Fig 1A) [4,5]. Ratchet junction reads contain a segment adjacent to an annotated 5' or 3' splice site juxtaposed to a segment adjacent to an unannotated intronic recursive site, providing direct evidence for the presence of a recursive splicing event.
Second, we developed a new computational tool, RatchetPair, to identify read pairs that map to distant genomic sites in a manner such that presence of intervening recursive splicing can be inferred from the size distribution of inserts in the sequenced library (Methods; Fig 1A). Unlike ratchet junction reads, recursive junction spanning read pairs do not pinpoint a specific recursive site. Instead, a recursive site is inferred based on the empirical distribution of fragment lengths and genomic sequence information. To do so, we adapted the GEM algorithm [10], originally designed to infer protein binding sites from ChIP-seq data, to assign a probability that each read-pair was indicative of a recursive site in a given region (Methods). This modified GEM algorithm was run with all read pairs and splice junction reads pooled together to derive the empirical distribution of fragment lengths.
Third, we developed the first automated software, RatchetScan, for inference of recursive sites from sawtooth patterns in read density (Fig 1B). This type of pattern is an expected product of co-transcriptional recursive splicing and has been associated with many recursive introns [4,5,11]. Briefly, assuming that RNA is spliced shortly after transcription elongation past the recursive splice site or 3' splice site, the splicing of recursive segments during transcription of subsequent sequences will result in a sawtooth distribution of reads across the intron with recursive sites commonly located near the right-hand base of each “tooth”. It is important to note that these approaches do not differentiate between unproductive splicing (followed by degradation of the intron-containing transcript) and productive splicing of the full intron.
RachetScan predicts the locations of recursive sites in three distinct steps. First, RNA-seq data was processed to summarize read density in each sub-intronic region (S1A Fig). We then developed a Markov Chain Monte Carlo- (MCMC-) based inference algorithm to detect presence of sawtooth patterns in introns. This algorithm is suitable for efficient exploration of complex intronic read patterns encountered when considering a variable number of possible recursive splice sites in each intron. We considered all nucleotides as potential recursive splice sites, rather than only focus on sites at the center of strong juxtaposed recursive motifs, allowing us to independently use sequence information to assess the false-positive rate of our method. Our RachetScan algorithm is initiated with a randomly chosen state, consisting of a set of proposed recursive sites in the intron (S1B Fig). In each round, a new state is proposed by perturbing the current state, with three classes of perturbations: (1) a new recursive site is added; (2) a recursive site is removed; or (3) a recursive site location is locally shifted, each with defined probabilities. Using a scoring function and transition rules (detailed in Methods), the algorithm decides to either accept the new proposed state or maintain the current state. This procedure was iterated over 107 rounds and the current state was sampled every 50 rounds, where the number of samples recorded in each state is proportional to the probability that the intron is best fit by the model corresponding to that state. Finally, recursive sites are predicted based on the output of the inference algorithm and sequence information (S1C Fig; S2 Fig). This approach does not infer the order of splicing of recursive segments (but see below).
Combining these three approaches and using reads pooled across all replicates and labeling periods, our analysis detected 539 candidate recursive sites in 379 fly introns (S1 Table). From this set, we curated a set of 243 “high confidence” recursive sites in 157 introns (identified by at least 2 methods with greater than 5 supporting junctions or read pairs or a sawtooth FDR of 5% and visual inspection of read densities), and a “medium confidence” set of 296 sites (identified by at least 1 method with greater than 5 supporting junctions or read pairs or a sawtooth FDR of 20%; Fig 2A; Methods). Approximately 60% of our high-confidence sites (144 sites) were identified using all three approaches. Overall, 98 introns contained multiple recursive sites, with up to seven high-confidence sites observed in a single intron. For instance, intron 1 of the tenascin major (Ten-m) gene contains five recursive sites, two of which were previously unknown (Fig 1C). Of the recursive sites previously reported by Duff and colleagues, 124 occurred in genes expressed in S2 cells. Our approach detected 119 (96%) of these known sites, as well as 126 novel high confidence sites and 294 novel medium confidence sites (Fig 2A), thus increasing the number of recursive sites defined in this cell type by ~4-fold (S2C Fig). For three recursive segments, we were also able to detect reads that spanned the intronic lariat resulting from the second step of splicing (S2 Table; Methods). For 13 sites in 3 recursively spliced introns, we performed RT-PCR validation experiments using primers flanking recursive segments, followed by sequencing (Methods). These experiments validated 8 previously identified sites and 5 novel recursive sites in nascent RNA from Drosophila S2 cells, including 3 sites in an ~55 kb intron of Tet that was not previously known to be recursively spliced (S3A Fig, S3D Fig; S3 Table). Both the high confidence and the medium confidence candidate recursive sites exhibited a strong juxtaposed 3'/5' splice site motif (S4 Fig). The greater numbers of sites detected by our approach (2–4 times more sites in this cell type), using less than 1/20th as many reads as used by Duff and colleagues, affirms the potential of nascent RNA analysis for identification of recursive splice sites.
Using this updated catalog of recursive sites, we observed that many very long introns (> 40 kb in length) have recursive sites, with 63% of such introns containing at least one high-confidence recursive site, and an additional 7% containing medium-confidence site(s) (Fig 2B). This observation suggests that recursive splicing is the prevalent mechanism by which very large fly introns are excised. We assessed the sensitivity of our detection pipeline by running it on subsamples of reads ranging from 0.1% to 100% of the total reads (Fig 2C). The shape of the resulting curve tapered off at higher coverage levels but never plateaued: new recursive sites were still being detected as read depth increased from 50% to 100% of sequenced reads and therefore would likely increase further at higher read depths. A somewhat higher proportion of recursive sites were detected in high-expressed genes (TPM > 20) than low-expressed genes (TPM ≤ 20). However, subsampling of the reads mapping to high-expressed genes to levels comparable to those observed for low-expressed genes resulted in a substantially lower fraction of recursive sites at each depth, suggesting that recursive splicing is more prevalent in low-expressed than high-expressed genes (Fig 2C). Together, these data suggest that the true fraction of very long introns that contain recursive sites may be substantially higher than our observed fraction of 63–70%, i.e. that recursive splicing is likely present in almost all very long fly introns.
Recursive splice sites can be required for the processing of long introns [3]. However, it is possible that most recursive sites are functionally neutral, and that mRNA production is not impacted by their presence. The size of our dataset enabled us to examine four properties of recursive sites that could help to distinguish between these possibilities: sequence conservation; distribution in the fly genome; distribution within introns; and efficiency of splicing. In each case, the patterns observed suggest that recursive sites often have functional impact.
Both high and medium confidence recursive sites exhibited twice the level of evolutionary conservation observed in and around control AGGT motifs in long introns (Fig 3A), implying strong selection to maintain most or all of these sites. Recursively spliced introns were enriched in genes involved in functions related to development, morphogenesis, organismal, and cellular processes, with stronger enrichments observed for genes containing high-confidence recursive sites (Fig 3B; S4 Table). Both of these observations are consistent with results from a previous study based on a smaller sample of recursive introns [4].
Longer introns might contain more recursive sites purely by chance. Indeed, while the majority of recursively spliced introns had just one recursive site, the number of sites increased roughly linearly with intron length (Fig 3C). However, the positioning of recursive sites within introns was significantly biased away from a random (uniform) distribution. Instead, recursive sites in introns with only one such site tended to be located closer to the midpoint of the intron than expected by chance (Kolmogorov-Smirnov P = 0.003; Fig 3D). Furthermore, the first recursive site in introns with two or three such sites tended to be located approximately 33% and 25% of the way from the 5' end of the intron, respectively (Fig 3E). The distribution of recursive sites within introns suggests that they are positioned so as to break larger introns into “bite-sized” chunks of intermediate size (typically ~9–15 kb in length; S5A and S5B Fig) rather than at random locations that would more often produce much longer and much shorter segments. Recursively spliced introns were also enriched in first introns, which are longer than non-first introns, relative to subsequent introns in fly genes (hypergeometric P < 0.05).
To ask whether recursive splicing contributes to the efficiency of processing of very long introns, we evaluated the order and timing of recursive splicing events (Methods). We observed a steady increase in the proportion of exon-exon junction reads relative to recursive junctions across the time course, reflecting the progress of splicing (Fig 4A). Among recursive junction reads, we observed far higher counts of reads spanning the 5' splice site and the recursive site (RS), relative to RS-RS or RS-3' splice site junctions, consistent with recursive segments being predominantly excised in 5' to 3' order (Fig 4A; S5C Fig). This order of splicing is consistent with recursive splicing occurring co-transcriptionally. Using targeted RT-PCR amplification of segment combinations in nascent RNA from 3 recursively spliced introns, we were only able to detect products spanning the 5' splice site and recursive site (S3B Fig). Surprisingly, we did detect a product spanning the recursive site and 3' splice site for the third recursive site of the Ten-m intron in steady-state cDNA (S3C Fig; validated by sequencing, S3D Fig), indicating that splicing of downstream recursive segments can sometimes occur before splicing of initial segments. Finally, we also detected one read that spans a lariat resulting from a RS-RS junction (S2 Table), as well as four reads (for three junctions) that span lariats resulting from the excision of recursive introns in one segment (5'-3' junction). The observation of these reads indicates that these introns are not always recursively spliced, though we note that these lariats are from introns that are much shorter than typical recursive introns (1.7–2.5 kb).
Previously we developed a framework for estimating rates of splicing from nascent RNA sequencing data across different labeling periods [9]. Here, we adapted this approach to estimate the splicing half-lives of individual recursive segments (Methods; S5D Fig; S5 Table), which have a mean length of 9.1 kb (S5A Fig). Recursive segment half-lives were the slowest for the first segment in the intron, with faster half-lives for successive segments (S5E Fig). Overall, recursive segments had 1.5-fold longer half-lives than non-recursive introns of the same lengths (Fig 4B; Mann-Whitney P = 1.5×10−9). Estimating the mean splicing half-life of a recursive intron as the maximum of a set of exponentials (to approximate the waiting time to splice all recursive segments), we found that recursive introns are spliced more slowly than non-recursive introns of similar size (S5F Fig; Mann-Whitney P < 2.2 × 10−16), consistent with the larger number of biochemical steps involved in recursive versus non-recursive splicing of an intron.
To ask whether recursive splicing occurs while the intron is continuing to be transcribed, we calculated the ratio of the half-life of the first segment to the estimated time needed to transcribe the remainder of the intron (Methods). For 49% of recursively spliced introns, the first segment half-life is shorter than the time to transcribe the full recursive intron (Fig 4C), implying common co-transcriptional splicing in about half of cases. We observed that longer recursive introns were more likely to be spliced co-transcriptionally.
The accuracy of splicing is likely to be at least as important as its speed, since splicing to an arbitrary (incorrect) splice site will most often produce an mRNA that is unstable or encodes a protein that is aberrant or nonfunctional. As a simple measure of potential splicing errors, we tallied the fraction of nascent RNA reads (from the 5 minute labeling period) that spanned “non-canonical” splice junctions, involving pairs of intron terminal dinucleotides other than the three canonical pairs “GT-AG”, “GC-AG” and “AT-AC” that account for ~99.9% of all known fly introns. For the bulk of non-recursive introns (most of which are < 100 nt in length), the frequency of such non-canonical splicing was negligible (Fig 4D, black curve). However, for non-recursive introns with lengths matching the much more extended lengths of recursively spliced introns, potential splicing errors were much more frequent (Fig 4D, gray curve), suggesting that the fly spliceosome loses accuracy as intron length (and the number of possible decoy splice sites) increases. Notably, recursive introns had ~37% fewer non-canonical junctions compared to similarly sized non-recursive introns (Fig 4D, gold curve, Kolmogorov-Smirnov P = 0.015). Therefore, presence of recursive splice sites may increase the accuracy of splicing, perhaps at the expense of splicing speed.
Analysis of intermediates can provide insight into otherwise hidden biochemical pathways. Here, application of new computational approaches to nascent RNA sequencing data, which is highly enriched for splicing intermediates, enabled us to identify about four times more recursive sites in the Drosophila genome than were known previously. The surprisingly widespread occurrence of recursive splicing raises questions about what functions it may serve.
A priori, this pathway might improve the speed or accuracy of splicing, or might impact regulation. Our analyses suggest that recursive splicing does not in fact increase splicing rates, and may actually slow splicing somewhat, likely because of the additional steps involved. However, we observe that the Drosophila splicing machinery appears to make a relatively high rate of errors in the splicing of longer introns, and that presence of recursive sites may substantially improve splicing accuracy. In splicing of a non-recursive 30 kbp intron, the 5' splice site is synthesized about 20 minutes before its correct partner 3' splice site, creating a long window during which splicing can only occur to incorrect 3' splice sites, likely contributing to the higher error rate seen for long fly introns. Presence of a recursive site may help to organize the processing of the intron, keeping the splicing machinery associated with the 5' splice site engaged in a productive direction and avoiding engagement with decoy 3' splice sites. It was previously observed that masking a recursive splice site in a zebrafish cadm2 intron does not change the overall splicing of the intron but reduces cadm2 mRNA levels [5]. This observation could be explained if the recursive site promotes accurate splicing and prevents unproductive splicing pathways that result in unstable products targeted by RNA decay pathways such as nonsense-mediated mRNA decay.
Recursive sites may also participate in splicing regulation. A previous study of a handful of recursively spliced introns in humans identified RS-exons that are initially recognized during recursive splicing via an “exon definition” model of splice site recognition [5], while an alternative “intron definition” pathway has been proposed for recursive splice site recognition in flies [4]. An exon definition model would require presence of a 5' splice site downstream of each recursive site. Consistent with this model, we observed that recursive segments following recursive sites are enriched for strong 5' splice site motifs relative to first recursive segments and relative to non-recursive introns matched for length (Fig 4E). Use of an exon definition pathway in the initial steps of spliceosome assembly might also contribute to splicing accuracy, with the downstream 5' splice site helping to specify the recursive site [9]. It could also produce alternative mRNA isoforms containing an additional exon [5].
Exon definition of recursive segments through transient RS-exons requires that the recursive site first be recognized as a 3' splice site and subsequently as a 5' splice site for splicing of the subsequent segment (assuming that simultaneous recognition of an RS in both modes is sterically prohibited). For this ordered recognition to occur (and for sequential splicing of recursive sites generally), binding of dU2AF/U2 snRNP must outcompete binding of U1 to the RS prior to its splicing to the upstream exon. Consistent with this expectation, the 3' splice site motifs of RS are very strong, stronger than non-recursive 3' splice sites, and they have higher information content than RS 5' splice sites (S6 Fig).
Developmental genes are enriched for long introns, which are more likely to be recursively spliced, but explanations for this pattern remain murky. It is possible that intron length is used to tune the timing of expression of these genes relative to the rapid embryonic cell cycle [12,13]. Alternatively, long introns may be needed to accommodate large transcriptional enhancers or complex three-dimensional organization of these gene loci related to their dynamic transcriptional regulation, or to facilitate alternative splicing. Thus, it is unclear whether recursive splicing is a feature of developmental genes or exists to facilitate the splicing of long introns that independently persist in developmental genes. In addition to producing unstable mRNAs, splicing errors may also produce stable mRNAs that encode aberrant protein forms, including dominant negative forms. Perhaps recursive splicing has been selected for in these genes to improve splicing accuracy and avoid production of aberrant developmental regulatory proteins at critical stages to improve the robustness of development.
We used RNA-seq data from our recent study of splicing kinetics in Drosophila S2 cells (GEO GSE93763; [9]). These data included 3 independent replicates of S2 cells labeled for 5, 10 and 20 minutes with 500 μM 4-thiouridine, isolation of labeled RNA, and library preparation using random hexamer priming following ribosomal RNA subtraction. Separation of total RNA into newly transcribed and untagged pre-existing RNA was performed as previously described [14,15], where 4sU-labeled RNA was selectively biotinylated and captured using streptavidin coated magnetic beads. cDNA for two independent biological replicates of “total” RNA were prepared using an equal mix of random hexamers and oligo-dT primers from unlabeled S2 cells [9]. Libraries were sequenced with paired-end 51 nt reads (100 nt reads for the “total” RNA samples), generating an average of 126M read pairs per library. Reads were filtered and mapped to the Drosophila melanogaster dm3 reference assembly as described in [9].
Gene expression values (TPMs) in each replicate library were calculated using Kallisto [16] and the transcriptome annotations from FlyBase Drosophila melanogaster Release 5.57 [17].
We used three features of recursive sites found in our nascent sequencing data to identify recursive sites: (1) splice junction reads derived from putative recursive sites (“RachetJunctions”), (2) recursive-site spanning pairs, specifically read pairs that map to sites flanking putative recursive segments such that the fragment length can only be accounted for by the presence recursive intermediate (“Rachet Pair”), and (3) a sawtooth pattern in intronic read density (“RachetScan”). Details of the computational and statistical methods for each of these approaches and our pipeline for recursive site detection are described below.
Out of the full set of recursive sites that were identified across all three methods, we filtered down to a final set of sites with the following criteria: (1) in genes with TPM ≥ 1 in the total RNA libraries, (2) in introns with at least 3 reads spanning the 5' to 3' splice sites (using the largest annotated intron), and (3) not overlapping with an annotated 5' splice site in the that intron. We ran our final pipeline on reads pooled across replicates and labeling periods to increase detection power. This resulted in a total of 539 recursive sites identified by any method. High-confidence sites were identified by the criteria used by Duff et al. [4]. We wrote a script to plot the read density around putative recursive sites and manually filtered each site based on the presence of a recognizable sawtooth pattern. This resulted in the identification of 243 high-confidence sites.
Conservation of recursive sites was estimated using per nucleotide phastCons scores [18] from a 15-way Drosophila alignment downloaded from UCSC Genome Browser.
We calculated position weight matrices (PWM) for the intronic portions of Drosophila 5' and 3' splice sites using all annotated splice sites. These weight matrices were then juxtaposed with the 3' splice site PWM followed by the 5' splice site PWM to create a recursive splice site motif PWM. Individual motif occurrences were scored using a normalized bit score [23]. The bit score for each motif occurrence is defined as the sum across the log probabilities for each nt being drawn from the motif. We calculated normalized scores by subtracting the minimum possible score and dividing by the range of possible bit scores.
We searched for reads crossing the 5'SS-branch point junction using code previously developed in our lab (https://github.com/jpaggi/findbps). In short, our approach works by: (1) identifying reads that do not have a valid alignment; (2) splitting the unalignable reads just before the 7-mer best matching the consensus 5'SS motif; and (3) mapping this split read as a pair, requiring that the second segment align upstream of the first segment. We require the following features to be present: (1) each segment must be at least 15 nt long; (2) only 1 mismatch is allowed per segment; (3) segments must be separated by less than 1 Mbp; and (4) the pair has a unique alignment.
We then filtered the resulting alignments for cases where the second segment aligned immediately downstream of a 5'SS or recursive site and the first segment aligned within 100 nt upstream of a 3'SS or recursive site. Overall, we detected 323 5'SS-branch point junction reads across 319 introns. The putative branch points show a motif favoring an A at the branchpoint and a U at the -2 position, consistent with the human branchpoint consensus motif. We observed 7 5'SS-branchpoint junction reads from introns that we report to be recursively spliced. These counts are consistent with analysis by Duff et al., which identified 46 recursive lariat junction reads amongst 10.2 billion reads. If such reads occurred at the same frequency in our data, we would expect to observe 1.8 recursive lariat junction reads. All implicated branch points are adenosines. These 7 reads implicate a lariat associated with the following categories of splicing events (S2 Table): (1) 5'SS-RS: two reads associated with two unique junctions; (2) RS-RS: one read associated with one junction; and (3) 5'SS-3'SS: four reads associated with three unique junctions. The 5'SS-3'SS lariat junction reads suggest that recursive splicing is not always used for these introns. All three such junctions derived from introns of lengths far shorter than typical recursive introns (1762 nt, 1929 nt, and 2548 nt), suggesting that non-recursive splicing may compete with recursive splicing of introns in this size range.
Nascent RNA was isolated after 5 minutes of labeling with 4sU (as described above) and reverse transcribed to first-strand cDNA using ProtoScript II Reverse Transcriptase (M0368S, NEB) primed with random hexamers according to manufacturer’s protocol. cDNA was diluted 1:5 and 1uL was used as template for PCR reactions using primers designed to amplify recursive segments anchored by either the intronic 5' splice site or intronic 3' splice site (S3 Table). PCR amplification was performed using Taq DNA Polymerase (10342020, Invitrogen) for 40 cycles. PCR products were visualized on a 1.5% agarose gel relative to Azura PureView 50bp DNA ladder (AZ1155, Azura). PCR products were purified with a DNA Clean & Concentrator kit (D4033, Zymo) and Sanger sequenced to confirm the junction boundaries.
In order to assess the sensitivity of our recursive site detection pipeline, we subsampled our reads to various proportions of the total read coverage and re-assessed the number of recursive sites detected. To do so, we used the samtools view – s command [24] to subsample each fastq file from all samples to the following fractions: 0.1%, 0.5%, 1%, 2.5%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90%. For each of these subsampled read sets, we re-ran the entire recursive site detection pipeline as described above to assess the number of recursive sites detected.
To assess the impact of gene expression levels on our power to detect recursive sites, we separated long introns into those from lowly expressed genes (TPM ≤ 20) and highly expressed genes (TPM > 20). Using the subset of reads mapping to these genes, we repeated the subsampling procedure and entire recursive site detection pipeline described above to characterize the percentage of lowly or highly expressed long introns that have recursive sites.
Finally, to understand whether the lower proportion of lowly expressed introns that have recursive sites is due to technical or biological reasons, we subsampled reads from long introns within highly expressed genes to match the read distribution of a comparable number of long introns from lowly expressed genes. Specifically, we isolated all reads from long introns in highly expressed genes and used pysam [20] to randomly subsample these reads to match the distribution of reads from lowly expressed introns. Using only this subset of reads from highly expressed genes, we again repeated the subsampling procedure and the entire recursive site detection pipeline described above to characterize the percentage of highly expressed introns that have recursive when reads from these introns are subsampled to a lower read coverage.
Previous studies have searched exclusively for recursive junction reads consistent with the 5' to 3' removal of recursive segments [4,5]. In order to determine if recursive splicing does indeed follow a 5' to 3' order, we quantified junction reads consistent with alternative orders of recursive splicing. These reads fall into two categories: junction reads between two intronic AGGTs and junction reads from an intronic AGGT to an annotated 3' splice site.
We constrained our search to combinations or recursive sites producing recursive segments of at least 1 kb. Nearly all recursive segments detected in our study were greater than 1 kb, thus adding this constraint mainly served to filter out spurious hits likely caused by alignment errors and unannotated splicing events. We considered all events with support from at least 3 uniquely aligning reads with recursive splice sites scoring above 0.85 in the scoring metric described above. Requiring at least three uniquely aligning reads matches the cutoff used for our previous analysis, where we found that recursive splice sites generally have strong motifs that score greater than 0.85.
These analyses produced thirteen candidate intronic AGGT to annotated 3' splice site recursive junction reads, and no candidate intronic AGGT to AGGT recursive sites. These candidate recursive splice sites were evaluated visually in a genome browser. Two of these sites corresponded to recursive splice sites detected by both methods in our study. One of these sites has sixty recursive junction reads supporting a 5' to 3' order, while only five junction reads support a 3' to 5' order. The second site has 829 and 13 junction reads for the 5' to 3' and 3' to 5' orders, respectively. All other candidate alternative ordering sites did not appear to be represent viable recursive site candidates, due to either a lack of sawtooth pattern, low intron expression, or extensive repeats complicating the alignment. These data suggest that recursive splicing overwhelmingly, but perhaps not always, proceeds in a 5' to 3' order.
We quantified splicing rates for each recursive segment independently by applying an approach for 4sU RNA-seq data that we previously described [9]. Specifically, we used reads that overlapped recursive sites and junction reads (split between either the recursive site and an annotated splice site, between two recursive sites, or between the 5' and 3' splice sites; as detailed in S5D Fig), as measures of uncompleted and completed segment splicing, respectively. The junction dynamics approach from Pai et al. 2017 [9] was applied to each set of reads to obtained a splicing half-life for each recursive segment. For full introns matched for length, we used splicing half-lives calculated in Pai et al. 2017 [9]. We estimated co- vs. post-transcriptional splicing of the first recursive segment by comparing the segment splicing half-life to the time to transcribe the remainder of the intron. Specifically, the time to complete intron transcription was estimated as:
3′transcription=lengthofintron(nt)−lengthoffirstsegment(nt)1,500nt/min
and the splicing delay was calculated as the ratio of the first segment’s splicing half-life to the 3' transcription time.
We estimated the accuracy of splicing in Drosophila introns by identifying non-annotated junction reads with non-canonical splice site sequences within annotated introns within the nascent RNA reads from the 5 minute labeling period. To do so, we first re-mapped the raw 4sU-seq reads with the STAR v2.5 software [25], with the mapping parameter—outSAMattribute NH HI AS nM jM to mark the intron motif category for each junction read in the final mapped file.
The jM attribute adds a jM:B:c SAM attribute to split reads arising from exon-exon junctions. All junction reads were first isolated and separated based on the value assigned to the jM:B:c tag. Junction reads spanning splice sites in the following categories were considered to be annotated or canonical: (1) any annotated splice site based on FlyBase D. melanogaster Release 5.57 gene structures [jM:B:c,[20–26]], (2) intron motifs containing “GT-AG” (or the reverse complement) [jM:B:c,1 or jM:B:c,2], (3) intron motifs containing “GC-AG” (or the reverse complement) [jM:B:c,3 or jM:B:c,4], and (4) intron motifs containing “AT-AC” (or the reverse complement) [jM:B:c,5 or jM:B:c,6]. Junction reads with jM:B:c,0 were considered to arise from non-canonical non-annotated splice sites. We calculated the frequency of inaccurate splice junctions for each intron as a ratio of the density of reads arising from non-canonical non-annotated splice sites to the density of all junction reads from the intron.
We calculated the strength of splice sites using a maximum entropy model as implemented in maxEntScan [26] using 9 nucleotides around the 5' splice site (-3:+6) and 23 nucleotides around the 3' splice site (-20:+3). These models were optimized on mammalian splice site preferences, but seem to be reasonable for Drosophila as well and have been used in gene prediction in fly genomes.
Gene Ontology enrichment analyses were performed using a custom script to avoid significant gene ontology terms with overlapping gene sets. Specifically, the script used the Flybase gene ontology annotation downloaded from the Gene Ontology Consortium website [27] and searches for the gene ontology term with the most significant enrichment of genes with recursively spliced introns (relative to a background of all genes with introns greater than 10,000 kb). Genes that belong to the most significant gene ontology term are then removed from the foreground and background sets of genes and the process is repeated iteratively until no genes are left in the foreground set. P-values are computed using a Fisher-exact test and then corrected using a Benjamini-Hochberg multiple test correction.
Source code for our pipeline to identify recursive splicing sites is available at https://github.com/jpaggi/recursive.
|
10.1371/journal.pcbi.0040036 | Computational and Experimental Analysis of Redundancy in the Central Metabolism of Geobacter sulfurreducens | Previous model-based analysis of the metabolic network of Geobacter sulfurreducens suggested the existence of several redundant pathways. Here, we identified eight sets of redundant pathways that included redundancy for the assimilation of acetate, and for the conversion of pyruvate into acetyl-CoA. These equivalent pathways and two other sub-optimal pathways were studied using 5 single-gene deletion mutants in those pathways for the evaluation of the predictive capacity of the model. The growth phenotypes of these mutants were studied under 12 different conditions of electron donor and acceptor availability. The comparison of the model predictions with the resulting experimental phenotypes indicated that pyruvate ferredoxin oxidoreductase is the only activity able to convert pyruvate into acetyl-CoA. However, the results and the modeling showed that the two acetate activation pathways present are not only active, but needed due to the additional role of the acetyl-CoA transferase in the TCA cycle, probably reflecting the adaptation of these bacteria to acetate utilization. In other cases, the data reconciliation suggested additional capacity constraints that were confirmed with biochemical assays. The results demonstrate the need to experimentally verify the activity of key enzymes when developing in silico models of microbial physiology based on sequence-based reconstruction of metabolic networks.
| Geobacter sulfurreducens is a member of the Geobacteraceae family of micro-organisms that breathe metals and have a unique mode of metabolism. Stimulation of the activity of this species in the environment has been shown to result in the removal of radioactive contaminants in groundwater. Similarly, the respiration of these micro-organisms also has been linked to electricity generation in a microbial fuel cell. Both the rate of electricity generation and the efficiency of ground water clean-up can be enhanced through the improved understanding of the growth and metabolism of Geobacteraceae. In order to better understand the growth and metabolism of this organism, we had constructed a large-scale mathematical model of the metabolic network of this organism. Using this model, we identified reaction alternates that sustain metabolism in the event of gene deletions. We then experimentally confirmed the role of these metabolic reactions through gene deletion mutants and biochemical assays and improved the predictive ability of the mathematical model. Such an integrated computational and experimental approach can be used to study the activity and function of metabolic network in a rapid manner for other poorly characterized organisms of environmental relevance.
| Geobacter species are of interest because of their natural role in carbon and mineral cycling, their ability to remediate organic and metal contaminants in the subsurface, and their capacity to harvest electricity from waste organic matter [1–3]. Geobacter sulfurreducens [4] is the most commonly investigated species of this genus because a genetic system [5], the complete genome sequence [6], whole genome microarrays [7] and genome-scale proteomics [8] are available. Furthermore, functional genomics studies have provided insight into the mechanisms of extracellular electron transport onto important electron acceptors such as Fe(III) oxides and electrodes [9–14].
G. sulfurreducens can use either acetate or hydrogen as the sole electron donors for Fe(III) reduction, and fumarate or malate can also be used as terminal electron acceptors [4]. An understanding of acetate metabolism in Geobacter species is required because acetate, secreted by fermenting organisms, is the dominant electron donor for Geobacteraceae in soils and sediments [15], and because recent studies have shown that the addition of acetate to uranium-contaminated aquifers can stimulate in situ bioremediation of uranium contamination by Geobacter species [16,17]. Previous studies have demonstrated that Geobacter species, and the closely related Desulfuromonas acetoxidans, oxidize acetate via the TCA cycle [18–20]. However, many other aspects of acetate metabolism, and central metabolism in general, are still poorly understood. To better understand the physiology of G. sulfurreducens, a constraint-based genome-scale metabolic model was constructed and used to investigate the unique physiology associated with the reduction of extracellular electron acceptors, such as Fe(III) [21]. The genome-scale model enabled the assessment of the impact of global proton balance during Fe(III) reduction on biomass and energy yields, and successfully predicted the lower biomass yields observed during the growth of a mutant in which the fumarate reductase had been deleted [22].
Furthermore, the network reconstruction revealed the existence of a number of redundant or alternate pathways in the central metabolism of G. sulfurreducens [21]. Recent genetic and in silico studies have shown that the presence of such redundant metabolic pathways, as well as isozymes, can enable metabolic networks to withstand genetic perturbations [23–26]. Experimental evidence for alternate optimal pathways have been observed in E. coli, where four metabolic gene deletion mutants had significantly different metabolic flux distributions, but similar overall growth rates [25]. Hence, the systematic investigation of the role of redundant pathways using in silico models can provide key insights into the properties of the metabolic networks.
Here we report on a coupled computational and experimental evaluation of potential redundant pathways during acetate metabolism in G. sulfurreducens. We demonstrate the need for redundancy in the acetate assimilation pathways, due to a coupling between the TCA cycle and acetate activation to acetyl-CoA, and also the inactivity of some of the predicted alternatives for pyruvate oxidation to acetyl-CoA. We also show that by using this information to constrain the model, its predictive capacity can be improved.
A combined computational and experimental approach was used for characterizing key redundant metabolic pathways in G. sulfurreducens (see Figure 1). In order to identify all of the active redundant pathways in a specific environment, flux variability analysis (FVA) was used first to enumerate the reactions that participate in these pathways followed by Extreme Pathway Analysis (ExPA) to identify the alternate pathways. The FVA step is required as the direct application of the ExPA algorithm to a genome-scale network is computationally intractable [27]. This analysis was initially performed considering the use of acetate as electron donor and carbon source, with either fumarate or Fe(III) citrate as the electron acceptor (Figure 2).
FVA identified 32 reactions whose flux can vary with no effect on the growth rate and thus are predicted to participate in redundant pathways in the central metabolism of G. sulfurreducens (Table S2). From these 32 reactions, eight sets of redundant pathways that are comprised of reactions that are predicted to function as equivalent alternatives for optimal growth in G. sulfurreducens metabolism were identified with the Extreme Pathway Analysis algorithm [27]. These included reaction sets for the conversion of: pyruvate to acetyl-CoA, succinyl-CoA and acetate to succinate and acetyl-CoA glutamate to alphaketoglutarate, glutamate to alphaketoglutarate and glutamine, alphaketoglutarate to succinyl-CoA, AMP to ADP, and folate to tetrahydrofolate (see Figure 3 and Figure S1).
Some other reactions that could constitute alternatives providing redundancy, were not identified by this method because such reactions were considered inactive by the model when growing on acetate. One example is found in the malic enzyme that could participate in the conversion of malate to pyruvate (Figure S1), providing redundancy in the malate to pyruvate conversion, but it was not identified by the computational approach because this reaction is predicted to be inactive during optimal growth on acetate with either fumarate or Fe(III) citrate serving as the electron acceptor. However, this enzyme is predicted to be active under other conditions, such as during the absence of malate dehydrogenase activity (discussed in detail below).
Genetic and physiological analysis can help resolve the activities of the redundant pathways. To determine if the metabolic flexibility predicted by pathway analysis and modeling correctly described the physiology of G. sulfurreducens, and to further understand the role of some of the identified redundant pathways, genes encoding pyruvate ferredoxin oxidoreductase (Por, Figure 3A), phosphotransacetylase (Pta, Figure 3B) and acetyl-CoA transferase (Ato, Figure 3B), were inactivated (Table 1 and Figure S2). In addition to these reactions that are all stoichiometrically equivalent and are optimal alternatives, we also considered two cases of sub-optimal alternatives in the central metabolism for the evaluation of the predictive capacity of the model. These two pathways were identified by considering reactions which, when deleted in silico [21], were predicted to result in sub-optimal growth relative to the wild type growth rate. One of these was the oxidative decarboxylation of malate by the malic enzyme, and the subsequent carboxylation of pyruvate, which could potentially substitute for the activity of the malate dehydrogenase (Figure 3C), but at the cost of an extra ATP. The other sub-optimal alternatives considered were the synthesis of PEP from pyruvate (Figure 3D), via the PEP dikinase (Ppdk, EC 2.7.9.1), the PEP synthase (Ppsa, EC 2.7.9.2), or the PEP carboxykinase (Ppck, EC 4.1.1.32)/pyruvate carboxylase (Pc, EC 6.4.1.1) pathway. Of the three pathways, the Ppdk pathway is energetically more favorable than either the Ppck or the Ppsa pathway, both of which are stoichiometrically equivalent. This is because the Ppdk pathway can lead to proton translocation via the diphosphatase reaction (Ppa) and thereby contributes to maintaining the proton gradient and ATP synthesis. For the analysis, we inactivated the genes encoding malate dehydrogenase (Mdh, Figure 3C) and PEP carboxykinase (Ppck, Figure 3D; Table 1).
The enzymes whose genes were inactivated were selected as they would provide additional information on the role of key central metabolic pathways (acetate activation, gluconeogenesis and anapleurotic pathways), which are conserved across different Geobacteraceae, and would enhance our understanding of the physiology of these acetate-utilizing bacteria.
Growth of the mutants with acetate as the donor, the condition used for redundant pathway identification, was evaluated. In addition, to further evaluate the predictive capacity of the model, the mutants were grown on combinations of acetate, pyruvate, and hydrogen as the carbon source/electron donor with either fumarate or Fe(III) citrate as the electron acceptor (Figure 2). This resulted in a total of 12 different growth conditions.
The wild type strain could grow on all combinations of donors/acceptors except when pyruvate was the sole donor/carbon source with either fumarate or Fe(III) citrate as the acceptor (Figures 4A and 5A). However, G. sulfurreducens could grow in the presence of pyruvate and hydrogen with either acceptor. This indicated that pyruvate can be transported and used as a carbon source but it cannot serve as carbon and electron donor. The reason for this phenotype is not known and contrasted with predictions of the previously published model based on the presence of the transporter and enzymes needed for pyruvate oxidation [21]. Thus, it was necessary to incorporate an additional constraint on the pyruvate transport flux in order to ensure that pyruvate could contribute to growth as a carbon source but could not serve as the sole electron donor in silico. The pyruvate uptake constraint was chosen, as this constraint is active only during growth with pyruvate, and hence does not impact growth predictions in any other environment. The rate of pyruvate uptake was constrained to 0.15 mmol pyruvate/grams of dry weight per hour (gdwh), which is the rate required to meet the non-growth associated ATP maintenance demand (0.45 mmol ATP/gdwh) during growth with fumarate as the electron acceptor [21].
The pyruvate ferredoxin oxidoreductase (Por) reaction was evaluated because examination of the mutant phenotypes when growing on pyruvate as carbon source, would provide information about the functionality of the alternative reactions, pyruvate dehydrogenase (Pdh) and pyruvate formate lyase (Pfl), which could potentially substitute for pyruvate ferredoxin oxidoreductase activity, but with the production of different reduced electron carriers (Figure 3A). It is interesting to note that Por can be used not only for the anaerobic oxidation of pyruvate, but also for pyruvate synthesis [28], a role that likely occurs in this organism for the synthesis of three carbon compounds from acetate. In fact, previous in silico analysis has suggested that Por plays an important role in carbon fixation, converting acetate to pyruvate in G. sulfurreducens [21].
There were three putative Por encoded in the genome, two of them of the heterodimeric type, (gene clusters GSU1859–62 and GSU2052–54), and one of the homodimeric type (GSU0097 gene). The first two are more similar to indolpyruvate ferredoxin oxidoreductases, enzymes involved in the metabolism of aromatic amino acids (42% identity for AAM31789.1 from Methanosarcina mazei Go1, 42% identity for CAE09839 from Wolinella succinogenes DSM 1740). Furthermore, in both cases a gene coding for a phenylacetate CoA ligase, an enzyme involved in the metabolism of phenylalanine or in the aromatic catabolism of phenylacetic acid, is present immediately downstream, suggesting a putative role for the product of these genes in the reduction of a 2-oxoacid of the aromatic type. Therefore, the best Por candidate was GSU0097, encoding a putative homodimeric type enzyme [28]. It is similar to NifJ, a well characterized Por present in nitrogen fixing photosynthetic bacteria (59% identity to Q06879 from Nostoc sp. PCC 7120), where this enzyme has a role in providing electrons to ferredoxin or flavodoxin, the electron donors for nitrogenase [29].
The mutant deficient in GSU0097, designated POR1, lacked pyruvate-ferredoxin oxidoreductase activity (Table 2), indicating that the enzyme encoded by GSU0097 is the only functional pyruvate-ferredoxin oxidoreductase in G. sulfurreducens, under the conditions tested. Surprisingly, the POR1 strain was unable to grow with acetate as the carbon source and electron donor or with pyruvate as the carbon source and hydrogen as the electron donor (Figures 4B and 5B). The mutant did grow when both acetate and pyruvate were included in the medium. These results demonstrated that Por activity is the only way to produce sufficient pyruvate when growing on acetate. They also show that the predicted redundant pathways, pyruvate formate lyase, pyruvate dehydrogenase and another suboptimal alternative pathway through aldehyde dehydrogenase for the conversion of pyruvate to acetyl-CoA (Figure 3A and Table S3), are not functional, at least under the growth conditions evaluated. When cell extracts of the wild type strain were assayed (Table 2), we could not detect Pfl or Pdh activity, in accordance with the growth results. Therefore, the pyruvate formate lyase and pyruvate dehydrogenase reactions were inactivated in the model. In addition, the flux through the aldehyde dehydrogenase (Table S3) was constrained to be no greater than the corresponding value for the wild type case in order to limit the flux through this alternative pathway. New simulations with these additional constraints correctly predicted all the POR1 phenotypes (Table 3).
In the G. sulfurreducens metabolic network, gluconeogenic synthesis of phosphoenolpyruvate can occur through three possible ways with different energetic demands to provide the required PEP (Figure 3D). Hence, we evaluated the role of these pathways in gluconeogenesis through the analysis of a deletion mutant in phosphoenolpyruvate (PEP) carboxykinase (Ppck). The one gene in the G. sulfurreducens genome with clear homology to the Ppck enzymes from other organisms is GSU3385 (53% identical to BAD30010 from Corynebacterium glutamicum; [30]. The mutant in which GSU3385 was deleted, designated PPCK1, lacked the Ppck enzyme activity (Table 2), demonstrating that the GSU3385 gene codes for this enzyme. When the wild-type was assayed, activity was 10-fold higher with GDP than with ADP, suggesting that the enzyme belongs to the class of monomeric GTP dependent Ppck enzymes [31].
Strain PPCK1 grew at least as well as wild type with fumarate (Figure 4C) as the electron acceptor, but did not grow on Fe(III) citrate (Figure 5C). Growth on fumarate is consistent with model simulations which suggest that pyruvate phosphate dikinase (Ppdk) is the primary source for phosphoenolpyruvate (PEP) generation. The Ppdk reaction also produces diphosphate, which is hydrolyzed (Figure 3D) to translocate a proton across the cell membrane, resulting in an energetic advantage. The lack of growth on Fe(III) citrate was not predicted by the model. A possible reason for this is that Ppck activity could contribute to lower oxaloacetate levels, which is important because the conversion of malate to oxaloacetate by malate dehydrogenase is thermodynamically unfavorable (standard free energy change is + 29.7 kJ/mol [32]). A high oxaloacetate to malate ratio is not likely to occur when fumarate is the electron acceptor because malate pools are probably maintained at higher concentrations due to excess fumarate in the cell. Indeed growth on fumarate often results in secretion of malate [33]. Therefore, PPCK1 lethality when Fe(III) citrate is the electron acceptor is not because the phosphoenolpyruvate (PEP) requirements of the cells are not met by redundant pathways, but probably because of a disruption in oxaloacetate homeostasis. However, since metabolite concentrations are not represented in the Flux Balance Analysis metabolic model, oxaloacetate homeostasis constraint cannot be incorporated in the model.
The inactivation of the Mdh reaction (EC 1.1.1.37, Figure 3C) was included in this analysis to further investigate the role of redundant pathways in the anaplerotic reactions connecting the TCA cycle to the glycolytic-gluconeogenic pathway. Deleting the one gene in the G. sulfurreducens genome with homology to malate dehydrogenase (GSU1466) eliminated malate dehydrogenase activity (Table 2). The malate dehydrogenase-deficient strain, designated MDH1, could grow with hydrogen as the electron donor, but not acetate (Figures 4D and 5D). Growth on hydrogen in the absence of malate dehydrogenase is expected, as it is consistent with other mutants with defects in TCA-cycle enzymes [22,34] because the electrons obtained from hydrogen are likely to flow directly to the menaquinone pool, avoiding the need for reducing equivalents derived from the TCA cycle to generate energy.
The lack of growth with acetate as the electron donor when Fe(III) citrate is the electron acceptor was predicted by the model. Although there is a predicted alternative pathway for conversion of malate to oxaloacetate involving the malic enzyme and pyruvate carboxylase (Figure 3C), this alternative pathway is not optimal because it consumes ATP in the pyruvate carboxylase step. Simulations predicted that this extra ATP cost would prevent growth with Fe(III) citrate as the electron acceptor because of the already low energy yields with Fe(III).
For growth on fumarate, the model predicted that, in order to compensate for the lack of Mdh, the flux through the pyruvate carboxylase would have to increase over 50 fold relative to the wild type flux distribution. The measured activity of the pyruvate carboxylase was low in the wild type, ca. 5 μmol/mg of protein/min. When pyruvate carboxylase flux is constrained at levels for the wild type cells, the model correctly predicts that MDH1 should not be able to grow with acetate as the carbon source, even with fumarate as the electron donor (Table 3).
In order to evaluate the potential redundancy in pathways for converting acetate to acetyl-CoA (Figure 3B), the gene coding for the phosphotransacetylase (GSU2706) was deleted. There was only one gene putatively coding for Pta in the G. sulfurreducens genome (72% identical to AP00550643 from Desulfuromonas acetoxidans DSM 684) whereas there were two genes whose products putatively code for Ack, GSU2707 and GSU3348 [6], so we decided to inactivate GSU2706. The GSU2706 mutant, designated, PTA1, lacked phosphotransacetylase activity whereas high activity was detected in the wild type (Table 2).
PTA1 was unable to grow with acetate as the electron donor (Figures 4E and 5E). This contrasted with the predictions of the model which indicated that acetyl-CoA transferase should be the primary provider of acetyl-CoA and that growth was possible on acetate and fumarate. Even the addition of pyruvate as a carbon source to meet the gluconeogenic carbon requirements did not rescue the mutant growth on acetate. PTA1 did grow when pyruvate and hydrogen were the carbon source and electron donor, respectively (Figures 4E and 5E). However, when acetate was supplied as the carbon source to cultures growing with pyruvate and hydrogen, growth on Fe(III) citrate was completely inhibited and growth on fumarate was partially inhibited (Figures 4E and 5E).
In order to further evaluate this phenotype of PTA1, the predicted alternative pathway for succinyl-CoA synthesis catalyzed by succinyl-CoA synthetase was investigated. No succinyl-CoA synthetase activity was detected in cell extracts (Table 2), consistent with a previous report [33], even though genes coding for two putative subunits of this enzyme are present in the genome [6]. When the model was adjusted to remove the succinyl-CoA synthetase reaction it was found that for every acetyl-CoA molecule produced by acetyl-CoA transferase activity, one must be utilized in the citrate synthase reaction of the TCA cycle (Figure 6), effectively coupling the acetyl-CoA transferase flux with that of the TCA cycle. Thus, in order for acetate to be utilized for biomass production an alternative pathway for acetyl-CoA production is required. The acetate kinase/phosphotransacetylase pathway is apparently the only route for producing this acetyl-CoA. Growth is possible with hydrogen as the electron donor and pyruvate as the carbon source because pyruvate can be used for gluconeogenesis and hydrogen provides reducing equivalents.
The model simulations with an inactivated succinyl-CoA synthetase match the experimental observations in all the cases, except for growth in the presence of both acetate and pyruvate, or acetate, pyruvate and hydrogen. It may be that PTA1 accumulates acetyl phosphate, in a manner similar to mutants in other organisms that lack phosphotransacetylase [35]. Acetyl phosphate has been proposed to be a metabolic signal participating in the regulation of gene expression in other bacteria [36]. Therefore, it may be that acetate is affecting growth through a regulatory effect triggered by acetyl phosphate accumulation. This possibility is currently under investigation.
In order to further evaluate the mechanisms for formation of acetyl-CoA, the acetyl-CoA transferase activity (Figure 3B) was investigated. The model predicted that deleting the acetyl-CoA transferase activity would not permit growth on acetate, but that an acetyl-CoA transferase activity-deficient mutant should be able to grow with hydrogen as the electron donor. In the presence of hydrogen, the TCA cycle activity is not required for the generation of redox equivalents and consequently the acetyl-CoA transferase is predicted to be nonessential.
Two genes potentially encoding acetyl-CoA transferase enzymes were found in the G. sulfurreducens genome. These genes are also similar to well characterized acetyl-CoA hydrolases from S. cerevisiae (GSU0490 is 57% identical; [37,38]), and Neurospora crassa (GSU0174 is 58% identical [39]). The mutants lacking GSU0490 and GSU0174 were designated ATO1 and ATO2 respectively, and the double mutant in which both genes were inactivated was designated ATO3. Cell extracts of ATO1 and ATO2 had diminished acetyl-CoA transferase activity whereas the double mutant, ATO3, had no activity. The hydrolase activity was not affected in the mutants, demonstrating that GSU0490 and GSU0174 genes codes for acetyl-CoA transferases.
The single mutants ATO1 and ATO2 were able to grow under all the conditions tested (Figures 4F and 4G and 5F and 5G), as would be expected, because these mutants still have acetyl-CoA transferase activity. They did have a lower growth rate in the pyruvate-H2 medium than the wild type, an observation which could not be explained based on the Flux Balance Analysis model. ATO3, which completely lacked acetyl-CoA transferase activity, could only grow with hydrogen as the electron donor (Figures 4H and 5H). This follows the prediction of the model that growth on acetate is not possible because of the loss of the key step in succinate conversion in the TCA cycle. The ability of the mutant to grow with hydrogen as the electron donor and acetate as the carbon source indicates that the acetate kinase pathway for acetate activation is sufficient for this purpose.
The results demonstrate the continued need to experimentally verify the activity of key enzymes when developing in silico models of microbial physiology based on sequence-based reconstruction of metabolic networks. The initial version of the in silico model (Mahadevan et al., 2006) predicted the growth phenotype in 47 of the 72 growth conditions evaluated (Table 3). One of the most significant errors in the modeling was the prediction that pyruvate could serve as the sole electron donor to support growth. This prediction assumed that G. sulfurreducens had the necessary pyruvate transporters to support the required flux of pyruvate into the cell and the associated metabolic pathways to catabolize the assimilated pyruvate. The finding that G. sulfurreducens did not grow solely on pyruvate required that the pyruvate uptake be constrained to allow the use of pyruvate only as a carbon source (Materials and Methods). With this constraint the model accurately predicted the growth of the wild type for all conditions evaluated. However, there were 16 instances in which the model predicted the growth of mutants, because of predicted redundant pathways, when the mutants did not grow. These discrepancies between predictions and growth studies were helpful in identifying predicted pathways that are not actually functional or can not be considered truly redundant in G. sulfurreducens, at least under the growth conditions evaluated. These results have led to the following additional constraints in the model: the elimination of pyruvate formate lyase, pyruvate dehydrogenase, and succinyl-CoA synthetase reactions. Further constraints that limit the flux through pyruvate carboxylase, pyruvate transport and aldehyde dehydrogenase were also incorporated (see Figure 6). The model with the updated constraints was able to correctly predict 64 out of the 72 cases (89%) (Table 3). While these results are significant, it is important to emphasize that the model is essentially a reflection of our knowledge of metabolism and the improvement in the predictive capabilities of the model is the direct outcome of the characterization of the role of redundant pathways. Even with these additional constraints, the model could not predict the phenotypes in 5 conditions involving the PPCK (phosphoenolpyruvate carboxykinase) mutant (all during Fe(III) citrate reduction) and in 3 conditions involving the PTA (phosphotransacetylase) mutant, possibly due to accumulation of acetyl-phosphate, a regulatory feature, which can not yet be incorporated into the model due to lack of sufficient information.
In addition, an alternative approach for simulating the phenotype of knock-out mutations [40], in which minimization of the difference in flux distribution (MOMA) rather than growth maximization was the objective, was evaluated. However, applying MOMA with the additional constraints resulted in 14 incorrect predictions, of which 6 were false negatives. The remaining 8 were the same false positives predicted with the Flux Balance Analysis approach with the objective of growth maximization. The false negatives resulted from the fact that MOMA does not always identify non-zero growth solutions even if they are feasible, whereas such false negative predictions are less common in the Flux Balance Analysis approach.
In summary, the results demonstrate that the iterative comparison of the in silico and the in vivo phenotypes has led to additional information on the role and activity of the central metabolic pathways in G. sulfurreducens. Such integrated analysis of computational and experimental data can provide valuable insights on the activity and function of metabolic pathways in a rapid manner for poorly characterized organisms of environmental significance.
The previously described constraint-based in silico model of G. sulfurreducens [21] served as the basis for this analysis. Growth under different environments was simulated by modifying the constraints on the exchange fluxes of the corresponding growth medium constituents: electron donors such as hydrogen, pyruvate, and acetate, and electron acceptors such as Fe(III) citrate and fumarate. The presence of hydrogen in the environment was modeled by allowing a hydrogen influx of 10 mmol/gdwh. In the case of pyruvate, the in silico prediction that pyruvate could be utilized as the sole carbon and electron source if the pyruvate transporters are present, was inconsistent with physiological data, which indicated that G. sulfurreducens can grow on pyruvate only if an electron donor, such as hydrogen is also present (see Results). In order to ensure that pyruvate could contribute to growth as a carbon source but could not serve as the sole electron donor in silico, the rate of pyruvate uptake was constrained to 0.15 mmol pyruvate/gdwh, the rate required to meet the non growth associated ATP maintenance demand (0.45 mmol ATP/gdwh) during growth with fumarate as the electron acceptor [21]. The growth with acetate and fumarate was modeled by allowing the corresponding acetate and fumarate uptake rates to have flux values up to 5 mmol/gdwh and 25 mmol/gdwh respectively. Fe(III) citrate reduction in the presence of acetate was simulated by allowing acetate and Fe(III) uptake rates to have values up to 10 mmol/gdwh and 150 mmol/gdwh. These rates were chosen so that they were similar to experimentally observed values reported earlier [21,41]. Optimal growth rates with these constraints were calculated to be 0.055 hr−1 and 0.043 hr−1 for fumarate and Fe(III) citrate reduction, respectively, and were found to be consistent with experimental observations. If a particular substrate was not present in an environment, the uptake rate corresponding to the substrate was constrained to be zero.
Two different approaches to the prediction of a deletion mutant growth rate have been proposed in the past based on the assumptions of the cellular objective after a gene deletion [40,42,43]. Growth rates of in silico deletion mutants were calculated by these two approaches for all the mutants considered: (a) linear optimization which calculates the maximum possible growth rate for the mutant in the presence of a specific environmental condition [43] and (b) Minimization of Metabolic adjustment using the algorithm of Segre et al., 2002, which calculates the growth rate by using the optimal wild type flux distribution as a reference and minimizing the distance to the wild type solution (calculated by the Euclidean distance between the mutant and wild type flux distribution) in the flux coordinates. In silico deletion mutants with predicted growth rates lower than 0.001 hr−1 were considered to be lethal.
The range of variation in fluxes for all various reactions in the model at the predicted optimal (maximum possible) growth rate was calculated with the previously described (Mahadevan et al., 2003) Flux Variability Analysis (FVA) algorithm. Briefly, this algorithm involves the maximization and minimization of every flux in the network subject to the stoichiometric constraints and an additional constraint that forces an optimal growth rate. The solution of this series of optimization problems results in the maximum and the minimum value of flux allowed for every reaction in the network, given the constraint that the growth rate is optimal [27]. Simulations were performed for growth with acetate as the electron donor and limiting substrate, and either fumarate or Fe(III) citrate serving as the electron acceptor. The acetate flux was assumed to be 5 mmol/gdwh for the case of fumarate reduction and 10 mmol/gdwh for the simulation of growth during Fe(III) citrate reduction so that the rates were representative of uptake rates observed experimentally [41]. In order to eliminate scaling issues in the FVA formulation, the upper bound and lower bound for fluxes that did not have any constraint was fixed to 1,000 and −1,000 mmol/gdwh respectively instead of –Infinity and Infinity. The fluxes that varied (range greater than 0.01 mmol/gdwh) were used to identify redundant pathways using a modified Extreme Pathway calculating algorithm [44] as previously described [27].
G. sulfurreducens strains used in this study (Table 1) were grown as previously described [5] in anaerobic pressure tubes with acetate (15 mM) as the electron donor and either Fe(III) citrate (56 mM) or fumarate (40 mM) as the electron acceptor. When indicated, 15 mM sodium pyruvate, hydrogen gas (0.59 atmospheres) or mixtures of these were used to replace or supplement the acetate. When hydrogen was used as electron donor, 10 ml hydrogen gas was injected into the headspace.
Mutations were introduced into the chromosome of G. sulfurreducens strain DL1 (ATCC 51573) [4] by homologous recombination. Construction of linear DNA fragments for gene disruption by recombinant PCR, electroporation and mutant isolation were performed as previously described [5].
Primers used for the construction of the various linear fragments utilized for gene disruption are listed and described in Table S1. The resulting genotypes of the various mutants constructed in this study are depicted in Figure S2. The plasmids used for generating the mutants are summarized in Table 1 [45–47]. All mutants were isolated in NBAFYE plates (NBAF medium amended with yeast extract) with the appropriate anoxic sterile antibiotic (200 μg/ml kanamycin, 20 μg/ml gentamycin, or 10 μg/ml chloramphenicol) as needed, and supplemented with 0.1% peptone, and 15 mM pyruvate to alleviate possible metabolic limitations generated by the gene inactivations. The plates were incubated in an anaerobic chamber under a 7% H2, 10% CO2, and 83% N2 atmosphere at 30 °C. A single isolate of each mutant was selected for detailed analysis and maintained with the adequate antibiotic. All the insertion-deletions were confirmed by PCR analysis.
In order to generate a mutant strain completely lacking the acetate:succinate Coenzyme-A transferase activity, a double mutant (ATO3) was constructed by electroporation of ATO2 mutant with the DNA fragment used to construct ATO1 (Supplementary Information: Figure S2).
Protein content was determined by a modification of the method of [48] using bovine serum albumin as protein standard. Growth of the cultures containing fumarate as electron acceptor was estimated by measuring turbidity at 600 nm. Fe(II) concentrations were determined with the ferrozine assay as previously described [49].
Genomic DNA was purified using the MasterPure Complete DNA & RNA purification kit (Epicentre Technologies) PCR product purification and gel extraction were carried out using the PCR purification kit and the Qiaquick gel extraction kit (Qiagen). Primers were purchased from Sigma-Genosys. All PCR amplifications were done using Taq DNA polymerase (Qiagen).
Cell-free extracts were prepared from 100 ml mid-log cultures. Cells were harvested by centrifugation (12 min, 2450 × g rpm, 5 °C), washed with 50 mM potassium phosphate buffer, pH 7.3, containing 2.5 mM dithiothreitol, and centrifuged again. The cells were resuspended in 3 ml of the same buffer and were disrupted by sonication (20 times, 100 W, 10 s, in an ice water bath); the cell debris was removed by centrifugation (14,000 × g, 5 min, 5 °C), and the supernatant was further clarified by ultracentrifugation (125,000 × g, 1.5 h, 5 °C). The cell-free extracts used to determine pyruvate-ferredoxin oxidoreductase and pyruvate formate lyase activities were prepared under strict anoxic conditions using the same protocol except that the cells were disrupted in a French press at 40,000 kPa (two passages). All enzymatic assays were carried out at 30 °C. All specific activities are expressed in units per milligram of protein (1U = 1 μmol min−1).
Malate dehydrogenase activity (Mdh) was measured in the direction of oxaloacetate reduction [33] by monitoring the decrease of NADH absorption at 340 nm (E340 = 6.22 mM−1cm−1) in 1 ml assay mixtures containing 50 mM Tris-HCl, pH 8, 0.2 mM NADH, 2.5 mM oxaloacetate and cell extract.
Pyruvate carboxylase (Pc) and phosphoenolpyruvate carboxykinase (Ppck) activities were monitored at 340 nm in assays coupled to the NADH dependent reduction of oxaloacetate by malate dehydrogenase. The Pyc assay mixture (1 ml) contained 100 mM Tris-HCl pH 7.8, 5 mM MgCl2, 50 mM NaHCO3, 5 mM sodium pyruvate, 2 mM ATP, 0.1 mM NADH, 2 U porcine malate dehydrogenase (Sigma-Aldrich), and cell extract [50]. The Ppck activity assay mixture (1 ml) contained 100 mM Hepes buffer pH 7.8, 10 mM MgCl2, 0.5 mM MnCl2, 1 mM DTT, 50 mM NaHCO3, 0.25 mM NADH, 2.5 mM phosphoenolpyruvate, 2.5 mM GDP (or ADP), 10 U porcine malate dehydrogenase (Sigma-Aldrich), and cell extract [31].
Pyruvate-ferredoxin oxidoreductase (Por) activity was measured at 600 nm as the pyruvate-dependent reduction of methyl viologen (E600 = 12 mM−1cm−1) using a modified version of the method reported by [51]. The reaction mixture (1 ml) contained 50 mM Tris-HCl pH 7.5, 2.5 mM MgCl2, 1 mM thiamine pyrophosphate, 1 mM Coenzyme A (CoASH), 10 mM methyl viologen, and 10 mM sodium pyruvate. This assay was carried out in under anoxic conditions in sealed 1 ml cuvettes and all solutions were sparged with oxygen-free nitrogen.
Pyruvate formate lyase activity was determined at 340 nm using a coupled assay [52]. The reaction mixture contained 0.1 M Tris-HCl (pH 6.5), 0.2 mM CoA, 10 mM DTT, 1 mM NAD, 5 mM L-malate, 4 U of porcine citrate synthase (Sigma-Aldrich), 20 U of porcine malate dehydrogenase (Sigma-Aldrich), 50 mM pyruvate. The assay was carried out under anoxic conditions.
Pyruvate dehydrogenase was assayed by monitoring reduction of NAD at 340 nm [53]. The assay was carried out in 0.1M Tris-HCl (pH 6.5) containing 0.2mM of magnesium chloride, 0.01 mM calcium chloride, 0.3 mM thiamine pyrophosphate, 0.12 mM coenzyme A, 2.0 mM NAD, and 5 mM pyruvate.
Phosphotransacetylase (Pta) activity was measured in a coupled assay by monitoring reduction of NAD at 340 nm as previously described [54]. The reaction mixture contained 250 mM Tris-HCl (pH 7.8), 15 mM malic acid, 4.5 mM MgCl2, 2 mM CoASH, 22.5 mM NAD, 10 mM acetyl phosphate, 12 U of porcine malate dehydrogenase (Sigma-Aldrich), and 1.4 U of porcine citrate synthase (Sigma-Aldrich).
Acetate-succinate CoA transferase (Ato) activity was determined as described by Sohling and Gottschalk [55], via a coupled assay in which the product of the Ato reaction, acetyl-CoA, is condensed with oxaloacetate by the enzyme citrate synthase and the liberation of CoASH is monitored by measuring the reduction of 5,5′-dithio-bis(2-nitrobenzoic acid) (DTNB) at 412 nm (E412 = 13.6 mM−1cm−1). The reaction mixture contained 100 mM potassium phosphate buffer, pH 7.0, 200 mM, 1 mM oxaloacetate, 1 mM DTNB, 0.1 mM succinyl-CoA, and 3 U of porcine citrate synthase (Sigma-Aldrich).
Succinyl-CoA synthetase (Sucoas) was assayed in 50 mM Tris-HCl, pH 7.2, 100 mM KCl, 10 mM MgCl2, 0.4 mM ATP, 0.1 mM CoASH, and 20 mM sodium succinate [56]. Succinate-dependent succinyl-CoA formation was monitored at 235 nm (E235 = 4 mM−1cm−1).
Succinyl-CoA hydrolase (Sucoh) activity was measured by monitoring the liberation of free CoASH with DTNB at 412 nm [57]. The assay mixture (1 ml) contained 100 mM potassium phosphate buffer, pH 7.4, 0.125 mM DTNB, and 1 mM of acetyl-CoA.
One approach to maintaining robustness of metabolism is through alternate pathways that can readily substitute in the event of loss of function. Metabolic modeling based on linear programming known as Flux Balance Analysis is effective in predicting large-scale growth phenotypes in different organisms. Flux Variability Analysis is an algorithm to identify those reactions that participate in alternate metabolic pathways. In this algorithm, in addition to the stoichiometric and capacity constraints, a growth rate constraint that forces the growth rate to be optimal is incorporated. Then, the flux through every reaction in the model is maximized and minimized in the formulation. If a reaction has an alternate pathway, then the flux through this reaction can be zero and yet optimal growth can be maintained as the flux gets rerouted through the alternate pathway. The range of flux is calculated as the difference between the maximum and the minimum possible value. Therefore, any reaction that has a non-zero range in the FVA problem has an alternate pathway. These reactions are compiled and augmented with the reactions in the reverse direction and used as the input to the Extreme Pathway Analysis algorithm. The redundant pathways are then identified from the list of predicted pathways. Further details are available in a previously published manuscript (Mahadevan and Schilling, 2003). However, if the number of reactions identified is small such as in the case of G. sulfurreducens, these pathways can be determined by manual inspection as shown in Table S2.
The GenBank Database (http://www.ncbi.nlm.nih.gov/) accession numbers for the G. sulfurreducens proteins described in this report are as follows: Por (GSU0097), AAR33432.1; Ppck (GSU3385), AAR36775.1; Pta (GSU2706), AAR36078.1; Ato1 (GSU0490), AAR33822.1; Ato2 (GSU0174), AAR33509.1; and Mdh (GSU1466), AAR34840.1. |
10.1371/journal.pcbi.1000380 | Pairwise Maximum Entropy Models for Studying Large Biological
Systems: When They Can Work and When They Can't | One of the most critical problems we face in the study of biological systems is
building accurate statistical descriptions of them. This problem has been
particularly challenging because biological systems typically contain large
numbers of interacting elements, which precludes the use of standard brute force
approaches. Recently, though, several groups have reported that there may be an
alternate strategy. The reports show that reliable statistical models can be
built without knowledge of all the interactions in a system; instead, pairwise
interactions can suffice. These findings, however, are based on the analysis of
small subsystems. Here, we ask whether the observations will generalize to
systems of realistic size, that is, whether pairwise models will provide
reliable descriptions of true biological systems. Our results show that, in most
cases, they will not. The reason is that there is a crossover in the predictive
power of pairwise models: If the size of the subsystem is below the crossover
point, then the results have no predictive power for large systems. If the size
is above the crossover point, then the results may have predictive power. This
work thus provides a general framework for determining the extent to which
pairwise models can be used to predict the behavior of large biological systems.
Applied to neural data, the size of most systems studied so far is below the
crossover point.
| Biological systems are exceedingly complicated: They consist of a large number of
elements, those elements interact in nonlinear and highly unpredictable ways,
and collective interactions typically play a critical role. It would seem
surprising, then, that one could build a quantitative description of biological
systems based only on knowledge of how pairs of elements interact. Yet, that is
what a number of studies have found. Those studies, however, focused on
relatively small systems. Here, we ask the question: Do their conclusions extend
to large systems? We show that the answer depends on the size of the system
relative to a crossover point: Below the crossover point the results on the
small system have no predictive power for large systems; above the crossover
point the results on the small system may have predictive power. Moreover, the
crossover point can be computed analytically. This work thus provides a general
framework for determining the extent to which pairwise models can be used to
predict the behavior of large biological systems. It also provides a useful
heuristic for designing experiments: If one is interested in understanding truly
large systems via pairwise interactions, then make sure that the system one
studies is above the crossover point.
| Many fundamental questions in biology are naturally treated in a probabilistic
setting. For instance, deciphering the neural code requires knowledge of the
probability of observing patterns of activity in response to stimuli [1];
determining which features of a protein are important for correct folding requires
knowledge of the probability that a particular sequence of amino acids folds
naturally [2],[3]; and determining the patterns of foraging of
animals and their social and individual behavior requires knowledge of the
distribution of food and species over both space and time [4]–[6].
Constructing these probability distributions is, however, hard. There are several
reasons for this: i) biological systems are composed of large numbers of elements,
and so can exhibit a huge number of configurations—in fact, an
exponentially large number, ii) the elements typically interact with each other,
making it impossible to view the system as a collection of independent entities, and
iii) because of technological considerations, the descriptions of biological systems
have to be built from very little data. For example, with current technology in
neuroscience, we can record simultaneously from only about 100 neurons out of
approximately 100 billion in the human brain. So, not only are we faced with the
problem of estimating probability distributions in high dimensional spaces, we must
do this based on a small fraction of the neurons in the network.
Despite these apparent difficulties, recent work has suggested that the situation may
be less bleak than it seems, and that an accurate statistical description of systems
can be achieved without having to examine all possible configurations [2], [3], [7]–[11]. One merely has to measure
the probability distribution over pairs of elements and use those to build the full
distribution. These “pairwise models” potentially offer a
fundamental simplification, as the number of pairs is quadratic in the number of
elements, not exponential. However, support for the efficacy of pairwise models has,
necessarily, come from relatively small subsystems—small enough that the
true probability distribution could be measured experimentally [7]–[9],[11]. While
these studies have provided a key first step, a critical question remains: will the
results from the analysis of these small subsystems extrapolate to large ones? That
is, if a pairwise model predicts the probability distribution for a subset of the
elements in a system, will it also predict the probability distribution for the
whole system? Here we find that, for a biologically relevant class of systems, this
question can be answered quantitatively and, importantly,
generically—independent of many of the details of the biological system
under consideration. And the answer is, generally, “no.” In this
paper, we explain, both analytically and with simulations, why this is the case.
To gain intuition into the extrapolation problem, let us consider a specific
example: neuronal spike trains. Fig. 1A shows a typical spike train for a small population of
neurons. Although the raw spike times provide a complete description, they are
not a useful representation, as they are too high-dimensional. Therefore, we
divide time into bins and re-represent the spike train as 0 s and 1 s: 0 if
there is no spike in a bin; 1 otherwise (Fig. 1B) [7]–[9],[11]. For
now we assume that the bins are independent (an assumption whose validity we
discuss below, and in more detail in the section “Is there anything
wrong with using small time bins?”). The problem, then, is to find where is a binary variable indicating no spike () or one or more spikes () on neuron . Since this, too, is a high dimensional problem (though less
so than the original spike time representation), suppose that we instead
construct a pairwise approximation to , which we denote , for a population of size . (The pairwise model derives its name from the fact that it
has the same mean and pairwise correlations as the true model; see Eq. (15).)
Our question, then, is: if is close to for small , what can we say about how close the two distributions are for
large ?
To answer this question quantitatively, we need a measure of distance. The
measure we use, denoted , is defined in Eq. (3) below, but all we need to know about it
for now is that if then , and if is near one then is far from . In terms of , our main results are as follows. First, for small , in what we call the perturbative regime, is proportional to . In other words, as the population size increases, the
pairwise model becomes a worse and worse approximation to the true distribution.
Second, this behavior is entirely generic: for small , increases linearly, no matter what the true distribution is.
This is illustrated schematically in Fig. 2, which shows the generic behavior of . The solid red part of the curve is the perturbative regime,
where is a linearly increasing function of ; the dashed curves show possible behavior beyond the
perturbative regime.
These results have an important corollary: if one does an experiment and finds
that is increasing linearly with , then one has no information at all about the true
distribution. The flip side of this is more encouraging: if one can measure the
true distribution for sufficiently large that saturates, as for the dashed blue line in Fig. 2, then there is a chance that
extrapolation to large is meaningful. The implications for the interpretation of
experiments is, therefore, that one can gain information about large behavior only if one can analyze data past the perturbative
regime.
Under what conditions is a subsystem in the perturbative regime? The answer turns
out to be simple: the size of the system, , times the average probability of observing a spike in a bin,
must be small compared to 1. For example, if the average probability is 1/100,
then a system will be in the perturbative regime if the number of neurons is
small compared to 100. This last observation would seem to be good news: if we
divide the spike trains into sufficiently small time bins and ignore temporal
correlations, then we can model the data very well with a pairwise distribution.
The problem with this, though, is that temporal correlations can be ignored only
when time bins are large compared to the autocorrelation time. This leads to a
kind of catch-22: pairwise models are guaranteed to work well (in the sense that
they describe spike trains in which temporal correlations are ignored) if one
uses small time bins, but small time bins is the one regime where ignoring
temporal correlations is not a valid approximation.
In the next several sections we quantify the qualitative picture presented above:
we write down an explicit expression for , explain why it increases linearly with when is small, and provide additional tests, besides assessing the
linearity of , to determine whether or not one is in the perturbative
regime.
A natural measure of the distance between and is the Kullback-Leibler (KL) divergence [12], denoted and defined as(1)
The KL divergence is zero if the two distributions are equal; otherwise it is
nonzero.
Although the KL divergence is a very natural measure, it is not easy to interpret
(except, of course, when it is exactly zero). That is because a nonzero KL
divergence tells us that , but it does not give us any real handle on how good, or bad,
the pairwise model really is. To make sense of the KL divergence, we need
something to compare it to. A reasonable reference quantity, used by a number of
authors [7]–[9], is the KL divergence
between the true distribution and the independent one, the latter denoted . The independent distribution, as its name suggests, is a
distribution in which the variables are taken to be independent,(2)where is the distribution of the response of the neuron, . With this choice for a comparison, we define a normalized
distance measure—a measure of how well the pairwise model explains the data—as(3)
Note that the denominator in this expression, , is usually referred to as the multi-information [7],[13],[14].
The quantity lies between 0 and 1, and measures how well a pairwise model
does relative to an independent model. If it is 0, the pairwise model is equal
to the true model (); if it is near 1, the pairwise model offers little
improvement over the independent model; and if it is exactly 1, the pairwise
model is equal to the independent model (), and so offers no improvement.
How do we attach intuitive meaning to the two divergences and ? For the latter, we use the fact that, as is easy to show,(4)where and are the entropies [15],[16] of and , respectively, defined, as usual, to be . For the former, we use the definition of the KL divergence to write(5)
The quantity has the flavor of an entropy, although it is a true entropy
only when is maximum entropy as well as pairwise (see Eq. (6) below).
For other pairwise distributions, all we need to know is that lies between and . A plot illustrating the relationship between , the two entropies and , and the entropy-like quantity , is shown in Fig.
3.
Note that for pairwise maximum entropy models (or maximum entropy models for
short), has a particularly simple interpretation, since in this case really is an entropy. Using to denote the pairwise entropy of a maximum entropy model, for
this case we have(6)as is easy to see by inserting Eqs. (4) and (5) into (3). This
expression has been used previously by a number of authors [7],[9].
The extrapolation problem discussed above is the problem of determining in the large limit. This is hard to do in general, but there is a
perturbative regime in which it is possible. The small parameter that defines
this regime is the average number of spikes produced by the whole population of
neurons in each time bin. It is given quantitatively by where is the bin size and the average firing rate,(7)with the firing rate of neuron .
The first step in the perturbation expansion is to compute the two quantities
that make up : and . As we show in the section “Perturbative
Expansion” (Methods), these are
given by(8a)(8b)where(9a)(9b)
Here and in what follows we use to denote terms that are proportional to in the limit . The in Eq. (9a) has been noted previously [7], although the
authors did not compute the prefactor, .
The prefactors and , which are given explicitly in Eqs. (42) and (44), depend on
the low order statistics of the spike trains: depends on the second order normalized correlation
coefficients, depends on the second and third order normalized correlation
coefficients (the normalized correlation coefficients are defined in Eq. (16)
below), and both depend on the firing rates of the individual cells. The details
of that dependence, however, are not important for now; what is important is
that and are independent of and (at least on average; see next section).
Inserting Eq. (8) into Eq. (3) (into the definition of ) and using Eq. (9), we arrive at our main result,(10a)(10b)
Note that in the regime , is necessarily small. This explains why, in an analytic study
of non-pairwise model in which was small, Shlens et al. found that was rarely greater than 0.1 [8].
We refer to quantities with a superscript zero as “zeroth
order.” Note that, via Eqs. (4) and (5), we can also define zeroth
order entropies,(11a)(11b)
These quantities are important primarily because differences between them and the
actual entropies indicate a breakdown of the perturbation expansion (see in
particular Fig. 4 below).
Assuming, as discussed in the next section, that and are approximately independent of , , and , Eq. (10) tells us that scales linearly with in the perturbative regime—the regime in which . The key observation about this scaling is that it is
independent of the details of the true distribution, . This has a very important consequence, one that has major
implications for experimental data: if one does an experiment with small and finds that is proportional to , then the system is, with very high probability, in the
perturbative regime, and one does not know whether will remain close to as increases. What this means in practical terms is that if one
wants to know whether a particular pairwise model is a good one for large
systems, it is necessary to consider values of that are significantly greater than , where(12)
We interpret as the value at which there is a crossover in the behavior of
the pairwise model. Specifically, if , the system is in the perturbative regime and the pairwise
model is not informative about the large behavior, whereas if , the system is in a regime in which it may be possible to make
inferences about the behavior of the full system.
As we show in Methods (see in particular
Eqs. (42) and (44)), the prefactors and depend on which neurons out of the full population are used.
Consequently, these quantities fluctuate around their true values (in the sense
that different subpopulations produce different values of and ), where “true” refers to an average over
all possible sub-populations. Here we assume that the neurons are chosen randomly from the full population, so any
set of neurons provides unbiased estimates of and . In our simulations, the fluctuations were small, as indicated
by the small error bars on the blue points in Fig. 5. However, in general the size of the
fluctuations is determined by the range of firing rates and correlation
coefficients, with larger ranges producing larger fluctuations.
Because does not affect the mean values of and , it is reasonable to think of these quantities—or at
least their true values—as being independent of . They are also independent of , again modulo fluctuations. Finally, as we show in the section
“Bin size and the correlation coefficients” (Methods), and are independent of in the limit that is small compared to the width of the temporal correlations
among neurons. We will assume this limit applies here. In sum, then, to first
approximation, and are independent of our three important quantities: , , and . Thus, we treat them as effectively constant throughout our
analysis.
Although the behavior of in the perturbative regime does not tell us much about its
behavior at large , it is possible that other quantities that can be calculated
in the perturbative regime, , , and (the last one exactly), are informative, as others have
suggested [7]. Here we show that this is not the
case—they also are uninformative.
The easiest way to relate the perturbative regime to the large regime is to ignore the corrections in Eqs. (8a) and (8b),
extrapolate the expressions for the zeroth order terms, and ask what their large behavior tells us. Generic versions of these extrapolations,
plotted on a log-log scale, are shown in Fig. 4A, along with a plot of the independent
entropy, (which is necessarily linear in ). The first thing we notice about the extrapolations is that
they do not, technically, have a large behavior: one terminates at the point labeled , which is where (and thus, via Eq. (0a), ; continuing the extrapolation implies negative true zeroth
order entropy); the other at the point labeled , which is where (and thus, via Eq. (5) and the fact that , ).
Despite the fact that the extrapolations end abruptly, they still might provide
information about the large regime. For example, and/or might be values of at which something interesting happens. To see if this is the
case, in Fig. 4B we plot the
naive extrapolations of and (that is, the zeroth order quantities given in Eq. (11), and ), on a linear-linear plot, along with . This plot contains no new information compared to Fig. 4A, but it does elucidate
the meaning of the extrapolations. Perhaps its most striking feature is that the
naive extrapolation of has a decreasing portion. As is easy to show mathematically,
entropy cannot decrease with (intuitively, that is because observing one additional neuron
cannot decrease the entropy of previously observed neurons). Thus, , which occurs well beyond the point where the naive
extrapolation of is decreasing, has essentially no meaning, something that has
been pointed out previously by Bethge and Berens [10]. The other
potentially important value of is . This, though, suffers from a similar problem: when , is negative.
How do the naively extrapolated entropies—the solid lines in Fig. 4B—compare to
the actual entropies? To answer this, in Fig. 4B we show the true behavior of and versus (dashed lines). Note that is asymptotically linear in , even though the neurons are correlated, a fact that forces to be linear in , as it is sandwiched between and . (The asymptotically linear behavior of is typical, even in highly correlated systems. Although this
is not always appreciated, it is easy to show; see the section “The
behavior of the true entropy in the large N limit,”
Methods.) Comparing the dashed and solid lines, we see that the naively
extrapolated and true entropies, and thus the naively extrapolated and true
values of , have extremely different behavior. This further suggests that
there is very little connection between the perturbative and large regimes.
In fact, these observations follow directly from the fact that and depend only on correlation coefficients up to third order (see
Eqs. (42) and (44)) whereas the large behavior depends on correlations at all orders. Thus, since and tell us very little, if anything, about higher order
correlations, it is not surprising that they tell us very little about the
behavior of in the large limit.
To check that our perturbation expansions, Eqs. (8–10), are correct,
and to investigate the regime in which they are valid, we performed numerical
simulations. We generated, from synthetic data, a set of true distributions,
computed the true distance measures, , , and numerically, and compared them to the zeroth order ones, , , and . If the perturbation expansion is valid, then the true values
should be close to the zeroth order values whenever is small. The results are shown in Fig. 5, and that is, indeed, what we
observed. Before discussing that figure, though, we explain our procedure for
constructing true distributions.
The set of true distributions we used were generated from a third order model (so
named because it includes up to third order interactions). This model has the form(13)where is a normalization constant, chosen to ensure that the
probability distribution sums to 1, and the sums over , and run from 1 to . The parameters and were chosen by sampling from distributions (see the section
“Generating synthetic data,” Methods), which allowed us to
generate many different true distributions. In all of our simulations we
calculate the relevant quantities directly from Eq. (13) . Consequently, we do
not have to worry about issues of finite data, as would be the case in realistic
experiments.
For a particular simulation (corresponding to a column in Fig. 5), we generated a true distribution
with , randomly chose 5 neurons, and marginalized over them. This
gave us a 10-neuron true distribution. True distributions with were constructed by marginalizing over additional neurons
within our 10-neuron population. To achieve a representative sample, we
considered all possible marginalizations (of which there are 10 choose , or ). The results in Fig. 5 are averages over these marginalizations.
For neural data, the most commonly used pairwise model is the maximum entropy
model. Therefore, we use that one here. To emphasize the maximum entropy nature
of this model, we replace the label “pair”
that we have been using so far with
“maxent.” The maximum entropy distribution
has the form(14)
Fitting this distribution requires that we choose the and so that the first and second moments match those of the true
distribution. Quantitatively, these conditions are(15a)(15b)where the angle brackets, and , represent averages with respect to and , respectively. Once we have and that satisfy Eq. (15), we calculate the KL divergences, Eqs.
(1) and (4), and use those to compute .
The results are shown in Fig.
5. The rows correspond to our three quantities of interest: , , and (top to bottom). The columns correspond to different values of , with the smallest on the left and the largest on the right. Red circles are the
true values of these quantities; blue ones are the zeroth order predictions from
Eqs. (9) and (10b).
As suggested by our perturbation analysis, the smaller the value of , the larger the value of for which agreement between the true and zeroth order values
is good. Our simulations corroborate this: the left column of Fig. 5 has , and agreement is almost perfect out to ; the middle column has , and agreement is almost perfect out to ; and the right column has , and agreement is not good for any value of . Note that the perturbation expansion breaks down for values
of well below (defined in Eq.(12)): in the middle column of Fig. 5 it breaks down when , and in the right column it breaks down when . This is not, however, especially surprising, as the
perturbation expansion is guaranteed to be valid only if .
These results validate the perturbation expansions in Eqs. (8) and (10), and show
that those expansions provide sensible predictions—at least for some
parameters. They also suggest a natural way to assess the significance of
one's data: plot , , and versus , and look for agreement with the predictions of the
perturbation expansion. If agreement is good, as in the left column of Fig. 5, then one is in the
perturbative regime, and one knows very little about the true distribution. If,
on the other hand, agreement is bad, as in the right column, then one is out of
the perturbative regime, and it may be possible to extract meaningful
information about the relationship between the true and pairwise models.
That said, the qualifier “at least for some parameters” is an
important one. This is because the perturbation expansion is essentially an
expansion that depends on the normalized correlation coefficients, and there is
an underlying assumption that they don't exhibit pathological behavior.
The order normalized correlation coefficient for the distribution , denoted , is written(16)A potentially problematic feature of the correlation coefficients
is that the denominator is a product over mean activities. If the mean
activities are small, the denominator can become very small, leading to very
large correlation coefficients. Although our perturbation expansion is always
valid for sufficiently small time bins (because the correlation coefficients
eventually becomes independent of bin size; see the section “Bin size
and the correlation coeffcients,” Methods), “sufficiently
small” can depend in detail on the parameters. For instance, at the
maximum population size tested () and for the true distributions that had , the absolute error of the prediction had a median of
approximately 16%. However, about 11% of the runs had
errors larger than 60%. Thus, the exact size of the small parameter
at which the perturbative expansion breaks down can depend on the details of the
true distribution.
Estimation of the KL divergences and from real data can be hard, in the sense that it takes a large
amount of data for them to converge to their true values. In addition, as
discussed above, in the section “The prefactors
gind and
gpair”, there are fluctuations in associated with finite subsampling of the full population of
neurons. Those fluctuations tend to keep from being purely linear, as can seen, for example, in the
blue points in Fig. 5F and
5I. We therefore provide a second set of relationships that can be used
to determine whether or not a particular data set is in the perturbative regime.
These relationships are between the parameters of the maximum entropy model, the and , and the mean activity and normalized second order correlation
coefficient (the latter defined in Eq. (19) below).
Since the quantity plays a central role in our analysis, we replace it with a
single parameter, which we denote ,(17)In terms of this parameter, we find (using the same perturbative
approach that led us to Eqs. (8–10); see the section
“External fields, pairwise couplings and moments,” Methods), that(18a)(18b)where , the normalized second order correlation coefficient, is
defined in Eq. (16) with ; it is given explicitly by(19)(We don't need a superscript on or a subscript on the angle brackets because the first and
second moments are the same under the true and pairwise distributions.) Equation
(18a) can be reconstructed from the low firing rate limit of analysis carried
out by Sessak and Monasson [17], as can the first three terms in the
expansion of the log in Eq. (18b).
Equation (18) tells us that the of the and , the external fields and pairwise couplings, is very weak. In
Fig. 6 we confirm this
through numerical simulations. Equation (18b) also provides additional
information—it gives us a functional relationship between the pairwise
couplings and the normalized pairwise correlations function, . In Fig.
7A–C we plot the pairwise couplings, , versus the normalized pairwise correlation coefficient, (blue dots), along with the prediction from Eq. (18b) (black
line). Consistent with our predictions, the data in Fig. 7A–C essentially follows a
line—the line given by Eq. (18b).
A relationship between the pairwise couplings and the correlations coefficients
has been sought previously, but for the more standard Pearson correlation
coefficient [7],[9],[11]. Our analysis explains why it was not found.
The Pearson correlation coefficient, denoted , is given by(20)
In the small limit—the limit of interest—the right hand
side of Eq. (20) is approximately equal to . Because depends on the external fields, and (see Eq. (18a)) and there is a one-to-one
relationship between and (Eq. (18b)), there can't be a one-to-one relationship
between and . We verify the lack of a relationship in Fig. 7D and 7E, where we again plot , but this time versus the standard correlation coefficient, . As predicted, the data in Fig. 7D and 7E is scattered over two
dimensions. This suggests that , not , is the natural measure of the correlation between two neurons
when they have a binary representation, something that has also been suggested
by Amari based on information-geometric arguments [18].
Note that the lack of a simple relationship between the pairwise couplings and
the standard correlation coefficient has been a major motivation in building
maximum entropy models [7],[11]. This is for good reason: if there is a
simple relationship, knowing the adds essentially nothing. Thus, plotting versus (but not ) is an important test of one's data, and if the two
quantities fall on the curve predicted by Eq. (18b), the maximum entropy model
is adding very little information, if any.
As an aside, we should point out that the is a function of the variables used to represent the firing
patterns. Here we use 0 for no spike and 1 for one or more spikes, but another,
possibly more common, representation, derived from the Ising model and used in a
number of studies [7],[9],[11], is to use −1 and +1
rather than 0 and 1. This amounts to making the change of variables . In terms of , the maximum entropy model has the form where and are given by(21a)(21b)
The second term on the right side of Eq. (21a) is proportional to , which means the external fields in the Ising representation
acquire a linear that was not present in our 0/1 representation. The two
studies that reported the of the external fields [7],[9] used
this representation, and, as predicted by our analysis, the external fields in
those studies had a component that was linear in .
An outcome of our perturbative approach is that our normalized distance measure, , is linear in bin size (see Eq. (10b)). This suggests that one
could make the pairwise model look better and better simply by making the bin
size smaller and smaller. Is there anything wrong with this? The answer is yes,
for reasons discussed above (see the the section “The extrapolation
problem”); here we emphasize and expand on this issue, as it is an
important one for making sense of experimental results.
The problem arises because what we have been calling the
“true” distribution is not really the true distribution of
spike trains. It is the distribution assuming independent time bins, an
assumption that becomes worse and worse as we make the bins smaller and smaller.
(We use this potentially confusing nomenclature primarily because all studies of
neuronal data carried out so far have assumed temporal independence, and
compared the pairwise distribution to the temporally independent—but
still correlated across neurons—distribution [7]–[9],[11]. In
addition, the correct name “true under the assumption of temporal
independence,” is unwieldy.) Here we quantify how much worse. In
particular, we show that if one uses time bins that are small compared to the
characteristic correlation time in the spike trains, the pairwise model will not
provide a good description of the data. Essentially, we show that, when the time
bins are too small, the error one makes in ignoring temporal correlations is
larger than the error one makes in ignoring correlations across neurons.
As usual, we divide time into bins of size . However, because we are dropping the independence assumption,
we use , rather than , to denote the response in bin . The full probability distribution over all time bins is
denoted . Here is the number of bins; it is equal to for spike trains of length . If time bins are approximately independent and the
distribution of is the same for all (an assumption we make for convenience only, but do not need;
see the section “Extending the normalized distance measure to the time
domain,” Methods), we can write(22)
Furthermore, if the pairwise model is a good one, we have(23)
Combining Eqs. (22) and Eq. (23) then gives us an especially simple expression
for the full probability distribution: .
The problem with small time bins lies in Eq. (22): the right hand side is a good
approximation to the true distribution when the time bins are large compared to
the spike train correlation time, but it is a bad approximation when the time
bins are small (because adjacent time bins become highly correlated). To
quantify how bad, we compare the error one makes assuming independence across
time to the error one makes assuming independence across neurons. The ratio of
those two errors, denoted , is given by(24)
It is relatively easy to compute in the limit of small time bins (see the section
“Extending the normalized distance measure to the time
domain,” Methods), and we find that(25)
As expected, this reduces to our old result, , when there is only one time bin (). When is larger than 1, however, the second term is always at least
one, and for small bin size, the third term is much larger than one.
Consequently, if we use bins that are small compared to the temporal correlation
time of the spike trains, the pairwise model will do a very bad job describing
the full, temporally correlated spike trains.
Probability distributions over the configurations of biological systems are extremely
important quantities. However, because of the large number of interacting elements
comprising such systems, these distributions can almost never be determined directly
from experimental data. Using parametric models to approximate the true distribution
is the only existing alternative. While such models are promising, they are
typically applied only to small subsystems, not the full system. This raises the
question: are they good models of the full system?
We answered this question for a class of parametric models known as pairwise models.
We focused on a particular application, neuronal spike trains, and our main result
is as follows: if one were to record spikes from multiple neurons, use sufficiently
small time bins and a sufficiently small number of cells, and assume temporal
independence, then a pairwise model will almost always succeed in matching the true
(but temporally independent) distribution—whether or not it would match
the true (but still temporally independent) distribution for large time bins or a
large number of cells. In other words, pairwise models in the
“sufficiently small” regime, what we refer to as the
perturbative regime, have almost no predictive value for what will happen with large
populations. This makes extrapolation from small to large systems dangerous.
This observation is important because pairwise models, and in particular pairwise
maximum entropy models, have recently attracted a great deal of attention: they have
been applied to salamander and guinea pig retinas [7], primate retina
[8],
primate cortex [9], cultured cortical networks [7], and cat visual
cortex [11].
These studies have mainly operated close to the perturbative regime. For example,
Schneidman et al. [7] had (for the data set described in their Fig. 5), Tang et al. [9] had to 0.4 (depending on the preparation), and Yu et al. [11] had . For these studies, then, it would be hard to justify
extrapolating to large populations.
The study by Shlens et al. [8], on the other hand, might be more amenable to
extrapolation. This is because spatially localized visual patterns were used to
stimulate retinal ganglion cells, for which a nearest neighbor maximum entropy
models provided a good fit to their data. (Nearest neighbor means is zero unless neuron and neuron are adjacent.) Our analysis still applies, but, since all but the
nearest neighbor correlations are zero, many of the terms that make up and vanish (see Eqs. (42) and (44)). Consequently, the small parameter
in the perturbative expansion becomes (rather than ), where is the number of nearest neighbors. Since is fixed, independent of the population size, the small parameter
will not change as the population size increases. Thus, Shlens et al.may have tapped
into the large population behavior even though they considered only a few cells at a
time in their analysis. Indeed, they have recently extended their analysis to more
than 100 neurons, and they still find that nearest neighbor maximum entropy models
provide very good fits to the data [19].
That the pairwise model is always good if is sufficiently small has strong implications: if we want to
build a good model for a particular , we can simply choose a bin size that is small compared to . However, one of the assumptions in all pairwise models used
on neural data is that bins at different times are independent. This produces a
tension between small time bins and temporal independence: small time bins
essentially ensure that a pairwise model will provide a close approximation to a
model with independent bins, but they make adjacent bins highly correlated.
Large time bins come with no such assurance, but they make adjacent bins
independent. Unfortunately, this tension is often unresolvable in large
populations, in the sense that pairwise models are assured to work only up to
populations of size where τcorr is the typical
correlation time. Given that is at least several Hz, for experimental paradigms in which
the correlation time is more than a few hundred ms, is about one, and this assurance does not apply to even
moderately sized populations of neurons.
These observations are especially relevant for studies that use small time bins
to model spike trains driven by natural stimuli. This is because the long
correlation times inherent in natural stimuli are passed on to the spike trains,
so the assumption of independence across time (which is required for the
independence assumption to be valid) breaks badly. Knowing that these models are
successful in describing spike trains under the independence assumption, then,
does not tell us whether they will be successful in describing full, temporally
correlated, spike trains. Note that for studies that use stimuli with short
correlation times (e.g., non-natural stimuli such as white noise), the temporal
correlations in the spike trains are likely to be short, and using small time
bins may be perfectly valid.
The only study that has investigated the issue of temporal correlations in
maximum entropy models does indeed support the above picture [9]: for
the parameters used in that study ( to 0.4), the pairwise maximum entropy model provided a good
fit to the data ( was typically smaller than 0.1), but it did not do a good job
modeling the temporal structure of the spike trains.
As mentioned in the Introduction, in addition to the studies on neuronal data,
studies on protein folding have also emphasized the role of pairwise
interactions [2],[3]. Briefly, proteins
consist of strings of amino acids, and a major question in structural biology
is: what is the probability distribution of amino acid strings in naturally
folding proteins? One way to answer this is to approximate the full probability
distribution of naturally folding proteins from knowledge of single-site and
pairwise distributions. One can show that there is a perturbative regime for
proteins as well. This can be readily seen using the celebrated HP protein model
[20], where a protein is composed of only two types of
amino acids. If, at each site, one amino acid type is preferred and occurs with
high probability, say with , then a protein of length shorter than will be in the perturbative regime, and, therefore, a good
match between the true distribution and the pairwise distribution for such a
protein is virtually guaranteed.
Fortunately, the properties of real proteins generally prevent this from
happening: at the majority of sites in a protein, the distribution of amino
acids is not sharply peaked around one amino acid. Even for
those sites that are sharply peaked (the evolutionarily-conserved sites), the
probability of the most likely amino acid, , rarely exceeds 90% [21],[22]. This puts proteins consisting of only a few
amino acids out of the perturbative regime, and puts longer
proteins—the ones usually studied using pairwise models—well
out of it.
This difference is fundamental: because many of the studies that have been
carried out on neural data were in the perturbative regime, the conclusions of
those studies—specifically, the conclusion that pairwise models
provide accurate descriptions of large populations of neurons—is not
yet supported. This is not the case for the protein studies, because they are
not in the perturbative regime. Thus, the evidence that pairwise models provide
accurate descriptions of protein folding remain strong and exceedingly
promising.
In our analysis, we sidestepped two issues of practical importance: finite
sampling and alternative measures for assessing the quality of the pairwise
model. These issues are beyond the scope of this paper, but in our view, they
are natural next steps in the analysis of pairwise models. Below we briefly
expand on them.
Finite sampling refers to the fact that in any real experiment, one has access to
only a finite amount of data, and so does not know the true probability
distribution of the spike trains. In our analysis, however, we assumed that one
did have full knowledge of the true probability distribution. Since a good
estimate of the probability distribution is crucial for assessing whether the
pairwise model can be extrapolated to large populations, it is important to
study how such estimates are affected by finite data. Future work is needed to
address this issue, and to find ways to overcome data limitation—for
example, by finding efficient methods for removing the finite data bias that
affects information theoretic quantities such as the Kullback-Leibler
divergence.
There are always many possible ways to assess the quality of a model. Our choice
of was motivated by two considerations: it is based on the
Kullback-Leibler divergence, which is a standard measure of
“distance” between probability distributions, and it is a
widely used measure in the field [7]–[10]. It
suffers, however, from a number of shortcomings. In particular, can be small even when the pairwise model assigns very
different probabilities to many of the configurations of the system. It would,
therefore, be important to study the quality of pairwise models using other
measures.
To understand how the true entropy behaves in the large limit, it is useful to express the difference of the entropies
as a mutual information. Using to denote the true entropy of neurons and to denote the mutual information between one neuron and the
other neurons in a population of size , we have(26)
If knowing the activity of neurons does not fully constrain the firing of neuron , then the single neuron entropy, , will exceed the mutual information, , and the entropy will be an increasing function of . For the entropy to be linear in , all we need is that the mutual information saturates with . Because of synaptic noise, this is a reasonable assumption
for networks of neurons: even if we observed all the spikes from all the
neurons, there would still be residual noise associated with synaptic failures,
jitter in release time, variability in the amount of neurotransmitter released,
stochastic channel dynamics, etc. Consequently, in the large limit, we may replace by its average, denoted . Also replacing by its average, denoted , we see that for large , the difference between and approaches a constant. Specifically,(27)where this expression is valid in the large limit and the corrections are sublinear in .
Our main quantitative result, given in Eqs. (8–10), is that the KL
divergence between the true distribution and both the independent and pairwise
distributions can be computed perturbatively as an expansion in powers of in the limit . Here we carry out this expansion, and derive explicit
expressions for the quantities and .
To simplify our notation, here we use for the true distribution. The critical step in computing the
KL divergences perturbatively is to use the Sarmanov-Lancaster expansion [23]–[28] for ,(28)where(29a)(29b)(29c)(29d)
This expansion has a number of important, but not immediately obvious,
properties. First, as can be shown by a direct calculation,(30)where the subscripts and indicate an average with respect to and , respectively. This has an immediate corollary,
This last relationship is important, because it tells us that if a product of contains any terms linear in one of the , the whole product averages to zero under the independent
distribution. This can be used to show that(31)from which it follows that
Thus, is properly normalized. Finally, a slightly more involved
calculations provides us with a relationship between the parameters of the model
and the moments: for ,(32a)(32b)
Virtually identical expressions hold for higher order moments. It is this last
set of relationships that make the Sarmanov-Lancaster expansion so useful.
Note that Eqs. (32a) and (32b), along with the expression for the normalized
correlation coefficients given in Eq. (16), imply that(33a)(33b)
These identities will be extremely useful for simplifying expressions later on.
Because the moments are so closely related to the parameters of the distribution,
moment matching is especially convenient: to construct a distribution whose
moments match those of up to some order, one simply needs to ensure that the
parameters of that distribution, , , , etc., are identical to those of the true distributions up to
the order of interest. In particular, let us write down a new distribution, ,(34a)(34b)
We can recover the independent distribution by letting , and we can recover the pairwise distribution by letting . This allows us to compute in the general case, and then either set to zero or set to .
Expressions analogous to those in Eqs. (31–33) exist for averages with
respect to ; the only difference is that is replaced by .
Using Eqs. (28) and (34a) and a small amount of algebra, the KL divergence
between and may be written(35)where(36)
To derive Eq. (35), we used the fact that (see Eq. (31)). The extra term was included to ensure that and its first derivatives vanish at , something that greatly simplifies our analysis later on.
Our approach is to Taylor expand the right hand side of Eq. (35) around , compute each term, and then sum the whole
series (we do not assume that either or is small). Using to denote the coefficients of the Taylor series, we have(37)
Note that because and its first derivatives vanish at , all terms in this sum have .
Because both and are themselves sums, an exact calculation of the terms in Eq.
(37) would be difficult. However, as we show below, in the section
“Averages of powers of ξp and
ξq” (see in particular
Eqs. (52) and (54)), they can be computed as perturbation expansions in powers
of , and the result is(38)where and are given by(39). The last equality in Eq. (39) follows from a small amount of
algebra and the definition of the correlation coefficients given in Eq. (16).
Equation (38) is valid only when , which is the case of interest to us (since the Taylor
expansion of has only terms with ).
The important point about Eq. (38) is that the and dependence follows that of the original Taylor expansion.
Thus, when we insert this equation back into Eq. (37), we recover our original
function—all we have to do is interchange the sums. For example,
consider inserting the first term in Eq. (38) into Eq. (37),
Performing the same set of manipulations on all of Eq. (38) leads to(40)
This expression is true in general (except for some technical considerations; see
the section “Averages of powers of
ξp and
ξq”); to restrict it to the KL
divergences of interest we set to and to either or . In the first case ( set to ), , which implies that , and thus . Because has a quadratic minimum at , when , the second two terms on the right hand side of Eq. (40) are . We thus have, to lowest nonvanishing order in ,(41)with the correction coming from the last sum in Eq. (40). Defining(42)where, recall , and inserting Eq. (42) into Eq. (41), we recover Eq. (8a).
In the second case ( set to ), the first and second moments of and are equal. This implies, using Eq. (32), that , and thus . Because (see Eq. (36)), the first three terms on the right hand side
of Eq. (40)—those involving and but not —vanish. The next order term does not vanish, and yields(43)
Defining(44)and inserting this expression into Eq. (43), we recover Eq.
(8b).
In this section we derive, to leading order in , expressions relating the external fields and pairwise
couplings of the maximum entropy model, and , to the first and second moments; these are the expressions
reported in Eq. (18). The calculation proceeds along the same lines as in the
previous section. There is, though, one extra step associated with the fact that
the quadratic term in the maximum entropy distribution given in Eq. (14) is
proportional to , not . However, to lowest order in , this doesn't matter. That's becausewhere is defined as in Eq. (29d) except with replaced by , and we used the fact that . The second term introduces a correction to the external
fields, . However, the correction is , so we drop it. We should keep in mind, though, that our final
expression for will have corrections of this order.
Using Eq. (14), but with replaced by where it appears with , we may write (after a small amount of algebra)(45)where is the same as the function defined in Eq. (29a) except that is replaced by , the subscript “ind”
indicates, as usual, an average with respect to , and the two functions and are defined by(46)and(47)
Given this setup, we can use Eqs. (55) and (56) below to compute the moments
under the maximum entropy model. That's because both and its first derivative vanish at , which are the two conditions required for these equations to
be valid. Using also the fact that , Eqs. (55) and (56) imply that(48a)(48b)(48c)where the first term in Eq. (48b) came from Eq. (29d) with replaced by , the term “” in Eq. (48c) came from , and for the second two expressions we used the fact that, to
lowest order in , the denominator in Eq. (45) is equal to 1.
Finally, using Eq. (19) for the normalized correlation coefficient, dropping the
subscript “maxent” (since the first and
second moments are the same under the maxent and true distributions), and
inverting Eqs. (48b) and (48c) to express the external fields and coupling
coefficients in terms of the first and second moments, we arrive at Eq (18).
Here we compute , which, as can be seen in Eq. (37), is the key quantity in our
perturbation expansion. Our starting point is to (formally) expand the sums that
make up and (see Eqs. (29b) and (34b)), which yields(49)
The sum over is a sum over all possible configurations of the . The coefficient are complicated functions of the , etc. Computing these functions is straightforward, although
somewhat tedious, especially when is large; below we compute them only for and 3. The reason starts at 2 is that ; see Eq. (37).
We first show that all terms with superscript are . To do this, we note that, because the right hand side of Eq.
(49) is an average with respect to the independent distribution, the average of
the product is the product of the averages,(50)Then, using the fact that with probability and with probability (see Eq. (29c)), we have(51)
The significance of this expression is that, for , , independent of . Consequently, if all the in Eq. (50) are greater than 1, then the right hand side is . This shows that, as promised above, the superscript labels the order of the terms.
As we saw in the section “The KL divergence in the Sarmanov-Lancaster
representation”, we need to go to third order in , which means we need to compute the terms on the right hand
side of Eq. (49) with and 3. Let us start with , which picks out only those terms with two unique indices.
Examining the expressions for and given in Eqs. (29b) and (34b), we see that we must keep only
terms involving , since has three indices, and higher order terms have more. Thus, the contribution to the average in Eq. (49), which we denote , is given by
Pulling and out of the averages, using Eq. (33a) to eliminate and in favor of and , and applying Eq. (51) (while throwing away some of the terms
in that equation that are higher than second order in ), the above expression may be written(52)
Note that we were not quite consistent in our ordering with respect to , in the sense that we kept some higher order terms and not
others. We did this so that we could use rather than , as the former is directly observable.
For the calculation is more involved, but not substantially so.
Including terms with exactly three unique indices in the sum on the right hand
side of Eq. (49) gives us(53)
This expression is not quite correct, since some of the terms contain only two
unique indices—these are the terms proportional to —whereas it should contain only terms with exactly
three unique indices. Fortunately, this turns out not to matter, for reasons we
discuss at the end of the section.
To perform the averages in Eq. (53), we would need to use multinomial expansions,
and then average over the resulting powers of . For the latter, we can work to lowest order in , which means we only take the first term in Eq. (51). This
amounts to replacing every with (and similarly for and ), and in addition multiplying the whole expression by an
overall factor of . For example, if and , one of the terms in the multinomial expansion is . This average would yield, using Eq. (51) and considering only
the lowest order term, .
This procedure also is not quite correct, since terms with only one factor of , which average to zero, are replaced with . This also turns out not to matter; again, we discuss why at
the end of the section.
We can, then, go ahead and use the above “replace blindly”
algorithm. Note that the factors of , and turn and into normalized correlation coefficients (see Eq. (33)), which
considerably simplifies our equations. Using also Eq. (39) for , Eq. (53) becomes(54)
We can now combine Eqs. (52) and (54), and insert them into Eq. (49). This gives
us the first two terms in the perturbative expansion of ; the result is written down in Eq. (38) above.
Why can we ignore the overcounting associated with terms in which an index
appears exactly zero or one times? We clearly can't do this in general,
because for such terms, replacing with fails—either because the terms didn't exist
in the first place (when one of the indices never appeared) or because they
averaged to zero (when an index appeared exactly once). In our case, however,
such terms do not appear in the Taylor expansion. To see why, note first of all
that, because of the form of , its Taylor expansion can be written where is finite at (see Eq. (36)). Consequently, the original Taylor expansion of , Eq. (37), should contain a factor of ; i.e.,where the are the coefficients of the Taylor expansion of . The factor , when expanded, has the form
As we saw in the previous section, we are interested in the third order term only
to compute , for which . Therefore, the above multiplicative factor reduces to . It is that last factor of that is important, since it guarantees that for every term in
the Taylor expansion, all indices appear at least twice. Therefore, although Eq.
(53) is not true in general, it is valid for our analysis.
We end this section by pointing out that there is a very simple procedure for
computing averages to second order in . Consider a function that has a minimum at and also obeys . Then, based on the above analysis, we have(55)
Two easy corollaries of this are: for and positive integers,(56a)(56b)where the sum in Eq. (56a) runs over only, and we used the fact that both and are symmetric with respect to the interchange of and .
As can be seen in Eq. (13), the synthetic data depends on three sets of
parameters: , and . Here we describe how they were generated.
To generate the , we draw a set of firing rates, , from an exponential distribution with mean 0.02 (recall that , which we set to 15, is the number of neurons in our base
distribution). From this we chose the external field according to Eq. (18a),
In the perturbative regime, a distribution generated with these values of the
external fields has firing rates approximately equal to the ; see Eq. (18a) and Fig. 6.
To generate the and , we drew them from Gaussian distributions with means equal to
0.05 and 0.02 and standard deviations of 0.8 and 0.5, respectively. Using
non-zero values for means that the true distribution is not pairwise.
One of our main claims is that is linear in bin size, . This is true, however, only if and are independent of , as can be seen from Eq. (10b). In this section we show that
independence is satisfied if is smaller than the typical correlation time of the responses.
For larger than such correlation times, and do depend on , and is no longer linear in . Note, though, that the correlation time is always finite, so
there will always be a bin size below which the linear relationship, , is guaranteed.
Examining Eqs. (42) and (44), we see that and depend on the normalized correlation coefficients, and (we drop superscripts, since our discussion will be generic).
Thus, to understand how and depend on bin size, we need to understand how the normalized
correlation coefficients depend on bin size. To do that, we express them in
terms of standard cross-correlograms, as the cross-correlograms contain, in a
very natural way, information about the temporal timescales in the spike train.
We start with the second order correlation coefficient, since it is simplest. The
corresponding cross-correlogram, which we denote , is given by(57)where is the time of the kth spike on
neuron (and similarly for ), and is the Dirac . The normalization in Eq. (57) is slightly
non-standard—more typical is to divide by something with units of
firing rate (, or ), to give units of spikes/s. The normalization we use is
convenient, however, because in the limit of large , approaches one.
It is slightly tedious, but otherwise straightforward, to show that when is sufficiently small that only one spike can occur in a time
bin, is related to via(58)
The (unimportant) factor comes from the fact that the spikes occur at random locations
within a bin.
Equation (58) has a simple interpretation: is the average height of the central peak of the
cross-correlogram relative to baseline. How strongly depends on is thus determined by the shape of the cross-correlogram. If
it is smooth, then approaches a constant as becomes small. If, on the other hand, there is a sharp peak at , then for small , so long as is larger than the width of the peak. (The factor of included in the scaling is approximate; it is a placeholder
for an effective firing rate that depends on the indices and . It is, however, sufficiently accurate for our purposes.) A
similar relationship exists between the third order correlogram and the
correlation coefficient. Thus, is also independent of in the small limit, whereas if the central peak is sharp it scales as .
The upshot of this analysis is that the shape of the cross-correlogram has a
profound effect on the correlation coefficients and, therefore, on . What is the shape in real networks? The answer typically
depends on the physical distance between cells. If two neurons are close, they
are likely to receive common input and thus exhibit a narrow central peak in
their cross-correlogram. Just how narrow depends on the area. Early in the
sensory pathways, such as retina [29]–[31] and LGN [32], peaks
can be very narrow—on the order of milliseconds. Deeper into cortex,
however, peaks tend to broaden, to at least tens of milliseconds [33],[34].
If, on the other hand, the neurons are far apart, they are less likely to
receive common input. In this case, the correlations come from external stimuli,
so the central peak tends to have a characteristic width given by the temporal
correlation time of the stimulus, typically 100 s of milliseconds.
Although clearly both kinds of cross-correlograms exist in any single population
of neurons, it is convenient to analyze them separately. We have already
considered networks in which the cross-correlograms were broad and perfectly
flat, so that the correlation coefficients were strictly independent of bin
size. We can also consider the opposite extreme: networks in which the the
cross-correlograms (both second and higher order) among nearby neurons exhibit
sharp peaks while those among distant neurons are uniformly equal to 1. In this
regime, the correlation coefficients depend on : as discussed above, the second order ones scale as and the third as . This means that the arguments of and are large (see Eqs. (42) and (44)). From the definition of in Eq. (36), in this regime both are approximately linear in
their arguments (ignoring log corrections). Consequently, and . This implies that and scale as and , respectively, and so , independent of . Thus, if the bin size is large compared to the correlation
time, will be approximately independent of bin size.
In this section we derive the expression for given in Eq. (25). Our starting point is its definition, Eq.
(24). It is convenient to define to be a concatenation of the responses in time bins,(59)where, as in the section “Is there anything wrong with
using small time bins?”, the superscript labels time, so is the full, temporally correlated, distribution.
With this definition, we may write the numerator in Eq. (24) as(60)where is the entropy of , the last sum follows from a marginalization over all but one
element of , and is the true distribution at time (unlike in the section “Is there anything wrong with
using small time bins?”, here we do not assume that the true
distribution is the same in all time bins). Note that is independent of time, since it is computed from a time
average of the true distribution. That time average, which we call , is given in terms of as
Inserting this definition into Eq. (60) eliminates the sum over , and replaces it with . For simplicity we consider the maximum entropy pairwise
model. In this case, because is in the exponential family, and the first and second moments
are the same under the true and maximum entropy distributions, we can replace with . Consequently, Eq. (60) becomes
This gives us the numerator in the expression for (Eq. (24)); using Eq. (4) to write , the full expression for becomes(61)where we added and subtracted to the numerator.
The first term on the right hand side of Eq. (61) we recognize, from Eq. (6), as . To cast the second into a reasonable form, we define to be the entropy of the distribution that retains the
temporal correlations within each neuron but is independent across neurons.
Then, adding and subtracting this quantity to the numerator in Eq. (61), and
also adding and subtracting , we have(62)
The key observation is that if , then
Comparing this with Eqs. (8a) and (9a), we see that is a factor of times larger than . We thus have(63)
Again assuming , and defining , the last term in this expression may be written(64)
Inserting this into Eq. (63) and using Eqs. (4), (8a) and (9a) yields Eq. (25).
We have assumed here that ; what happens when , or larger? To answer this, we rewrite Eq. (61) as(65)
We argue that in general, as increases, becomes increasingly different from , since the former was derived under the assumption that the
responses at different time bins were independent. Thus, Eq. (25) should be
considered a lower bound on .
|
10.1371/journal.ppat.1004726 | TRIM26 Negatively Regulates Interferon-β Production and Antiviral Response through Polyubiquitination and Degradation of Nuclear IRF3 | Virus infection leads to the activation of transcription factor IRF3 and subsequent production of type I inteferons, which induce the transcription of various antiviral genes called interferon stimulated genes (ISGs) to eliminate viral infection. IRF3 activation requires phosphorylation, dimerization and nuclear translocation. However, the mechanisms for the termination of IRF3 activation in nucleus are elusive. Here we report the identification of TRIM26 to negatively regulate IFN-β production and antiviral response by targeting nuclear IRF3. TRIM26 bound to IRF3 and promoted its K48-linked polyubiquitination and degradation in nucleus. TRIM26 degraded WT IRF3 and the constitutive active mutant IRF3 5D, but not the phosphorylation deficient mutant IRF3 5A. Furthermore, IRF3 mutant in the Nuclear Localization Signal (NLS), which could not move into nucleus, was not degraded by TRIM26. Importantly, virus infection promoted TRIM26 nuclear translocation, which was required for IRF3 degradation. As a consequence, TRIM26 attenuated IFN-β promoter activation and IFN-β production downstream of TLR3/4, RLR and DNA sensing pathways. TRIM26 transgenic mice showed much less IRF3 activation and IFN-β production, while increased virus replication. Our findings delineate a novel mechanism for the termination of IRF3 activation in nucleus through TRIM26-mediated IRF3 ubiquitination and degradation.
| Innate immunity is the first line of defense to protect host from infection of invading pathogens. Production of type I inteferons by the innate immune cells is pivotal for the cellular antiviral immune responses. After virus infection, IFN-β transcription requires IRF3, which is activated through phosphorylation, dimerization and nuclear translocation. Although IRF3 activation and IFN-β production are essential for the host to prevent viral infection, aberrant or excessive IFN-β production may lead to the pathogenesis of human autoimmune diseases. Therefore, IRF3 activation and IFN-β production must be terminated at the appropriate time points after viral infection. Degradation of IRF3 in the nucleus represents a novel mechanism to terminate IFN-β production. Here we identified TRIM26 as a novel E3 ligase to target nuclear IRF3. TRIM26 attenuated IFN-β production through polyubiquitination and degradation of nuclear IRF3. In vivo experiments with TRIM26 transgenic mice further confirmed the negative function of TRIM26 on IFN-β production and antiviral responses. Given that IRF3 is a common molecule downstream of TLR3, RLRs and intracellular DNA signaling pathways, our study identified a novel mechanism to limit RNA and DNA virus-induced signaling, inflammation and tissue injury and provided new clues for the treatment of autoimmune diseases.
| Innate immunity is essential for the host to protect from infection of invading pathogens. Activation of the innate immune response depends on the detection and recognition of pathogen-associated molecular patterns (PAMPs) by germline DNA-encoded pattern-recognition receptors (PRRs). The well studied PRRs include Toll-like receptors (TLRs), RIG-I-like receptors (RLRs), NOD-like receptors (NLRs) and intracellular DNA sensors [1], [2]. Among them, several types of PRR have been identified to recognize viral nucleic acid including RNA and DNA. For example, membrane-bound TLR3 recognizes extracellular viral double-stranded RNA in endosomes. Another type of RNA sensor is the cytosolic RLRs including RIG-I and MDA5, which detect intracellular viral dsRNA [3], [4]. Recently, several intracellular DNA sensors such as cGAS, IFI16, DDX41 and LRRFIP1 have been identified capable of sensing DNA from various microbes [5]-[8]. Upon binding with dsRNA, TLR3 triggers a signaling pathway mediated by Toll/IL-1R (TIR) domain-containing adaptor that induces IFN-β (TRIF) [9], [10]. While, RIG-I and MDA5 recruit a CARD-containing adapter protein MAVS (also known as VISA, IPS-1 and Cardif) to initiate the antiviral signaling pathway [11]-[14]. Although the nature of DNA sensors needs further investigation, the adaptor protein STING (also called MPYS, MITA, and ERIS) in DNA sensing pathway are well defined [15]-[17]. After viral infection, these key adaptors TRIF, MAVS and STING recruit the kinases TBK1 and IKKε to activate the transcription factor interferon-regulatory factor 3 (IRF3), leading to the production of type I inteferons and antiviral immune responses [18], [19]. Although IRF3 activation and IFN-β production are essential for the host to prevent viral infection, aberrant or excessive IFN-β production can lead to the pathogenesis of human autoimmune diseases such as SLE [20]. Therefore, IRF3 activation and IFN-β production must be terminated at the appropriate time points after viral infection.
IRF3 activation requires phosphorylation on multiple phosphorylation sites in the C-terminal of IRF3. After phosphorylation, IRF3 forms homo-dimmer and then moves into the nucleus, where it binds to target genes harboring interferon stimulation-response element (ISRE) [21]-[23]. Several mechanisms including dephosphorylation and polyubiquitination have been demonstrated to terminate IRF3 activation. Phosphorylated IRF3 was found to be dephosphorylated by phosphatase PP2A recruited by RACK1 adaptor protein. Therefore, RACK1 and PP2A limit virus-induced type I interferon signaling [24]. Phosphorylation in the C-terminal phosphor-acceptor has been reported to facilitate IRF3 proteasomal degradation after infection with SeV [23], [25]. But, the identity of the E3 ligase responsible for nuclear IRF3 ubiquitination and degradation is not defined.
TRIM26 is a member of the tripartite motif (TRIM) protein family composed of more than 70 members in human [26]. TRIM proteins share a similar characteristic structure, which includes a RING (R) domain, one or two B-boxes (B), and a coiled coil (CC) domain in the N-terminal and a domain in the C-terminal with variable structures. Here, we identified a novel function for TRIM26 as an E3 ubiquitin ligase for nuclear IRF3. TRIM26 bound to and induced IRF3 polyubiquitination in nucleus after virus infection. TRIM26 promoted the degradation of WT IRF3 and the phosphorylation active mutant IRF3 5D, but not the phosphorylation deficient mutant IRF3 5A. Nuclear translocation of IRF3 and TRIM26 was required for IRF3 degradation. TRIM26 transgenic mice had decreased IRF3 activation, IFN-β production and antiviral immune response. Our findings demonstrated that TRIM26 is essential for the termination of nuclear IRF3 activation.
TRIM26 is located in the MHC class I region [27], but its biological functions in the immune response remain elusive. To explore the function of TRIM26 in antiviral immune responses, the effect of TRIM26 on the activation of IFN-β expression downstream of various PRRs was investigated using IFN-β promoter reporter. LPS- and poly(I:C)-induced IFN-β promoter activation was attenuated in RAW264.7 macrophages transfected with TRIM26 expression plasmid, compared to that transfected with control vector (Fig. 1A). Similarly, transfection of TRIM26 expression plasmid also decreased LPS- and poly(I:C)-induced IFN-β activation in HEK293 cells stably expressing TLR4 and TLR3, respectively (Fig. 1A). Recognition of RNA virus through RIG-I like receptors (RLRs) may lead to the expression of IFN-β. Transfection of TRIM26 expression plasmid decreased SeV-induced IFN-β promoter activation in HEK293 cells (Fig. 1B). poly(I:C) present in the cytosol has been shown to activate IFN-β production through RIG-I and MDA-5 in HEK293 cells [4]. Consistent with the SeV infection data, IFN-β promoter activation induced by poly(I:C) transfection was also decreased upon TRIM26 overexpression (Fig. 1B). Recognition of DNA molecule through intracellular DNA sensors may also lead to the expression of IFN-β. To investigate whether DNA-induced IFN-β expression was affected by TRIM26, ISD (interferon-stimulating DNA) and poly(dA:dT) were transfeceted into Hela cells. Overexpression of TRIM26 attenuated ISD- and poly(dA:dT)-induced IFN-β promoter activation (Fig. 1C). Recent studies indicated that recognition of intracellular DNA molecule through cyclic GMP-AMP synthase (cGAS) led to the expression of IFN-β [5]. Overexpression of TRIM26 also attenuated cGAS-induced IFN-β promoter activation (Fig. 1C). These reporter assays strongly suggested that TRIM26 acts on molecules that are shared by various nucleic acid-induced signaling pathways to negatively regulate IFN-β expression.
To directly investigate the inhibitory role of TRIM26 in IFN-β production, two TRIM26 specific siRNAs were designed and transfected into peritoneal primary macrophages. Western blot analysis showed that the expression of TRIM26 protein was decreased after transfection with TRIM26 specific siRNA 1 and 2 (Fig. 1D). Transfection of TRIM26 siRNAs enhanced LPS- and poly(I:C)-induced IFN-β production (Fig. 1E). Importantly, TRIM26 siRNA 1, which has a higher efficiency to knockdown TRIM26 protein expression, has a greater potential to increase LPS- and poly(I:C)-induced IFN-β production (Fig. 1E). Therefore, TRIM26 siRNA 1 was used in the following experiments. Similarly, SeV- and ISD-induced IFN-β production was also increased in TRIM26 siRNA-transfected primary peritoneal macrophages (Fig. 1F). These results further confirmed the above reporter data and demonstrated that TRIM26 negatively regulates IFN-β production downstream of various PRRs.
IFN-β plays critical roles in the innate immune responses against viral infection. To directly investigate the effect of TRIM26 on antiviral responses, vesicular stomatitis virus (VSV) was used. Transfection of TRIM26 expression plasmid into Hela cells attenuated VSV-induced IFN-β expression (S1A Fig.), while VSV RNA replicates in the cells were increased in TRIM26-transfected cells (Fig. 1G). Accordingly, plaque assays showed that overexpression of TRIM26 substantially increased viral replication compared to control vector-transfected cells (Fig. 1G). In contrast, transfection of TRIM26 siRNA into primary peritoneal macrophages increased VSV-induced IFN-β expression (S1B Fig.), while intracellular VSV RNA replicates was decreased (Fig. 1H). Plaque assays showed that transfection of TRIM26 siRNA significantly decreased VSV viral replication (Fig. 1H). Taken together, these data demonstrated that TRIM26 negatively regulates IFN-β production and antiviral immune responses.
IRF3 is the main transcription factor involved in IFN-β production. To investigate the effect of TRIM26 on IRF3 activation, several set of experiments were performed. LPS and poly(I:C) stimulation increased the activation of ISRE reporter in HEK293 cells stably expressing TLR4 and TLR3, respectively (Fig. 2A). While, transfection with TRIM26 expression plasmid attenuated LPS- and poly(I:C)-induced activation of ISRE reporter (Fig. 2A). Similarly, SeV and ISD-induced ISRE activation was also decreased by TRIM26 overexpression (Fig. 2A). Transfection of TRIF, MAVS, STING+cGAS and TBK1 could induce IRF3 phosphorylation in HEK293 cells (Fig. 2B). Overexpression of TRIM26 substantially decreased TRIF-, MAVS-, STING+cGAS- and TBK1-induced IRF3 phosphorylation (Fig. 2B). In contrast, knockdown of endogenous TRIM26 expression by siRNA in primary peritoneal macrophages increased LPS-induced IRF3 phosphorylation (Fig. 2C). Similarly, SeV-induced IRF3 phosphorylation was also increased in TRIM26 siRNA-transfected macrophages (Fig. 2D). These data suggested that TRIM26 inhibits IRF3 activation to negatively regulate IFN-β production.
To determine the molecular targets of TRIM26, the effect of TRIM26 overexpression on the activation of IFN-β promoter mediated by various molecules was examined in reporter assays. As shown in Fig. 2E, TRIF-, RIG-I-, MAVS-, TBK1- and IRF3-induced IFN-β promoter activation was inhibited by TRIM26 overexpression in a dose dependent manner. Similarly, TRIF-, RIG-I-, MAVS-, TBK1- and IRF3-induced activation of ISRE reporter was also inhibited by TRIM26 overexpression (Fig. 2F). These data indicated that TRIM26 targets IRF3 directly to regulate IFN-β production.
The presence of RING-finger domain indicates TRIM26 may function as an E3 ligase. Thus, the ability of TRIM26 to induce IRF3 polyubiquitination and degradation was investigated. Co-Immunoprecipitation (Co-IP) showed that endogenous TRIM26 formed a complex with IRF3 upon LPS stimulation and SeV infection, while there is no interaction in untreated macrophages (Fig. 3A). Co-transfection of Flag-TRIM26 and Myc-IRF3 demonstrated that Flag-TRIM26 interacted with WT IRF3 and the constitutive active mutant IRF3 5D, but not with the non-active IRF3 mutant 5A (Fig. 3B).
To directly test TRIM26-mediated IRF3 ubiquitination, Myc-IRF3 was transfected into HEK293 cells together with Flag-TRIM26. Co-IP showed that the level of IRF3 ubiquitination was markedly increased in the presence of TRIM26 expression plasmid and MG-132 (Fig. 3C). Notably, two TRIM26 mutants in the RING-finger domain (C16A and C16/36A) lost the ability to promote IRF3 polyubiquitination (Fig. 3C), indicating TRIM26 promotes IRF3 ubiquitination though the RING-finger domain. Consistent with the inability of C16A to promote IRF3 ubiquitination, TRIF- and TBK1-induced IFN-β promoter activation was not affected by TRIM26 C16A (S2A Fig.). Furthermore, TRIM26 C16A did not facilitate VSV replication compared to WT TRIM26 (S2B Fig.). In vitro binding and ubiquitination assays demonstrated that TRIM26 could directly interact with IRF3 and promote IRF3 ubiquitination. Importantly, TRIM26-mediated IRF3 ubiquitination was dependent on the RING-finger domain because IRF3 ubiquitination induced by TRIM26 C16A was greatly attenuated compared to that induced by WT TRIM26 (Fig. 3D). To further confirm TRIM26-induced IRF3 ubiquitination, primary peritoneal macrophages were transfected with TRIM26 siRNA, and IRF3 ubiquitination was measured after SeV infection and LPS stimulation. SeV infection and LPS stimulation induced IRF3 ubiqutination (Fig. 3E). However, knockdown TRIM26 expression with siRNA substantially attenuated SeV- and LPS-induced IRF3 ubiquitination (Fig. 3E).
To investigate the form of polyubiquitin chains linked to IRF3, WT HA-Ubiquitin and its mutants K48 and K63, which has only one lysine residue in ubiquitin at position 48 and 63, respectively, were used. TRIM26-induced IRF3 ubiquitination could be easily detected in WT HA-Ub and K48 transfected cells (Fig. 3F). While, there is much less IRF3 ubiquitination in K63 transfected cells (Fig. 3F), suggesting that TRIM26 mainly conjugates K48-linked polyubiquitin chains to IRF3. K48-linked ubiquitination often leads to the degradation of target proteins by the 26S proteasome. Consistent with these observations, TRIM26-induced degradation of IRF3 was reversed by proteasome inhibitor MG-132, but not by lysosome inhibitor Chloroquine (Fig. 3G, lane 5 vs. lane 3 and 6). Importantly, TRIM26 C16A could not promote the degradation of IRF3 compared to WT TRIM26 (Fig. 3G, lane 4 vs. lane 3).
Previous studies have demonstrated that K70 and K87 are the two main ubiquitination sites in IRF3 [28]. To investigate whether TRIM26 promotes IRF3 ubiquitination through K70 and K87, IRF3 mutant K70/87A was transfected into HEK293 cells together with Flag-TRIM26. Compared to WT IRF3, TRIM26-induced IRF3 ubiquitination was decreased in IRF3 mutant K70/87A (Fig. 3H). Notably, TRIM26 did not induce the degradation of IRF3 mutant K70/87A compared to WT IRF3 (Fig. 3H, input, lane 3 vs. lane 4). All together, these data demonstrated that TRIM26 interacts with and promotes K48-linked polyubiqutination of IRF3 at K70/87, leading to IRF3 proteasomal degradation.
Western blot analysis showed that TRIM26 protein is strongly expressed in various organs including lung, thymus, liver, spleen, small intestine and brain, but not in heart and kidney (S3A Fig.). Sendai virus (SeV) infection, which activates the RLR signaling, increased TRIM26 protein expression in primary peritoneal macrophages and Hela cells (Fig. 4A). TRIM26 expression was also induced in primary peritoneal macrophages upon LPS and poly(I:C) stimulation, which activate TLR4 and TLR3, respectively (Fig. 4B). TRIM26 mRNA expression was also induced by LPS and poly(I:C) stimulation and SeV infection in macrophages (S3B Fig.). Scanning of the mice TRIM26 promoter sequence identified a putative ISRE sequence (GATTTCACTTTCC, -162 bp to-150 bp), indicating TRIM26 expression may rely on the transcription factors STAT1 and STAT2 downstream of IFN-β signaling. Indeed, IFN-β stimulation increased TRIM26 mRNA and protein expression (Fig. 4C, upper panel and S3B Fig.). Blockage of IFN-β signaling by antibody against IFN-β receptor 1 (IFNR1) also attenuated LPS-induced TRIM26 expression (Fig. 4C, lower panel, right). Transfection of STAT1 or STAT2 specific siRNA substantially attenuated LPS-induced TRIM26 expression (Fig. 4C, lower panel, left). To investigate the cellular localization of TRIM26, GFP-TRIM26 was constructed and transfected into HEK293 cells. TRIM26 showed diffused expression in the cytoplasm without stimulation. While, a considerable proportion of TRIM26 moved into nucleus after infection with SeV and VSV (Fig. 4D). Biochemical assays with cytoplasmic and nuclear proteins also confirmed the translocation of TRIM26 into nucleus after SeV infection in Hela cells (Fig. 1E). This translocation may be mediated by IFN-β because IFN-β alone could induce TRIM26 nuclear translocation in HEK293 cells (Fig. 4D). Similarly, LPS induced TRIM26 translocation into nucleus in HEK293 cells stably expressing TLR4 (S3C FigC). Endogenous TRIM26 was also translocated into nucleus upon LPS stimulation in peritoneal primary macrophages (Fig. 4F). All together, these data suggested that TRIM26 expression is induced upon viral infection, which also induces the nuclear translocation of TRIM26.
IRF3 normally shuttles between nucleus and cytoplasm, but is present dominantly in the cytoplasm prior infection. To investigate the location of TRIM26-mediated IRF3 ubiquitination and degradation, cytoplasmic and nuclear fractions were prepared from RAW264.7 macrophages after SeV infection. Co-IP showed that endogenous IRF3 interacted with TRIM26 in the nucleus, but not in the cytoplasm (Fig. 5A). SeV infection mainly promoted IRF3 ubiquitination in the nucleus in RAW264.7 cells and HEK293 cells (S4A and S4B Figs.). TRIM26 siRNA knockdown attenuated IRF3 ubiquitination in the nucleus (lane 8 vs. lane 7), while the level of IRF3 ubiquitination in the cytoplasm kept constant (Fig. 5B, lane 4 vs. lane 3), indicating that TRIM26 mainly interacts with and promotes IRF3 ubiquitination in the nucleus.
To investigate whether TRIM26 targets active IRF3 in nucleus, IRF3 WT and mutants 5D and 5A were transfected into HEK293 cells together with Flag-TRIM26. TRIM26 was found to degrade IRF3 WT and 5D in a dose dependent manner (Fig. 5C). In contrast, IRF3 5A could not be degraded by TRIM26 in the same settings (Fig. 5C). Notably, IRF3 5A was not degraded by TRIM26 even under the condition of infection with SeV and VSV (S4C Fig.). To investigate where IRF3 was degraded by TRIM26 in nucleus, cytoplasmic and nuclear fractions were prepared from IRF3 and TRIM26 or control vector transfected cells after infection with SeV or left uninfected. A considerable proportion of IRF3 was translocated into nucleus upon IRF3 overexpression without SeV infection (Fig. 5D, lane 5). SeV infection induced further nuclear translocation of IRF3 (Fig. 5D, lane 7). Overexpression of TRIM26 promoted the degradation of nuclear IRF3 (Fig. 5D, lane 6 and 8), but not the cytoplasmic IRF3 (Fig. 5D, lane 2 and 4). Similar to WT IRF3, a considerable proportion of 5D was translocated into nucleus upon overexpression (S4D Fig.). However, a very small proportion of IRF3 5A was translocated into nucleus compared to WT IRF3 and 5D (S4D Fig.). Overexpression of TRIM26 also promoted the degradation of IRF3 5D in nucleus (S4D Fig, lane 12), but not in cytoplasm (lane 10). IRF3 5A was not degraded in both cytoplasm and nucleus (S4D Fig.). These data suggested that TRIM26 mainly promotes the degradation of active form IRF3 in nucleus.
The localization of IRF3 is mediated by both the nuclear localization signal (NLS) and nuclear exporter signal (NES) [29]. IRF3 mutant KR77/78NG in the NLS lost the ability to translocate into nucleus after viral infection [29]. To investigate TRIM26-mediated IRF3 degradation is really in nucleus, IRF3 WT and IRF3 KR77/78NG were transfected into HEK293 cells together with TRIM26 expression plasmid. Compared to WT IRF3, the protein level of IRF3 KR77/78NG was greatly increased (Fig. 6A, lane 3 vs. lane 1). Importantly, KR77/78NG could not be degraded by TRIM26 (Fig. 6A, lane 4 vs. lane 3). TRIM26 also could not degrade KR77/78NG mutant even in the active form IRF3 5D (Fig. 6A, lane 8 vs. lane 7), and the protein level of IRF3 5D-KR77/78NG was greatly increased compared to that of IRF3 5D (Fig. 6A, lane 7 vs. 5). Consistent with the inability to degrade IRF3 KR77/78NG, Flag-TRIM26 was not interacted with IRF3 KR77/78NG, and 5D KR77/78NG (S5A Fig.). IRF3 mutant ΔDBD with the deletion of the DNA-binding domain (aa 1–115), which covers the NLS, could not be degraded by TRIM26 (S5B Fig.).
Deletion of the NES resulted in the nuclear accumulation of IRF3. To further confirm IRF3 was degraded in the nucleus, IRF3 NES mutant IL139/140MM was used. A large amount of IRF3 IL139/140MM was accumulated in the nucleus especially after SeV infection (Fig. 6B, lane 5 and 7). Western blot analysis of cytoplasmic and nuclear fractions showed that IRF3 IL139/140MM was efficiently degraded by TRIM26 in nucleus (Fig. 6B, lane 6 and 8), but not in the cytoplasm (Fig. 6B, lane 2 and 4). Furthermore, TRIM26 was found to degrade IRF3 5D-IL139/140MM, but not the 5A-IL139/140MM (S5C Fig.). Consistently, TRIM26 was found to interact with IRF3 WT IL139/140MM and 5D-IL139/140MM, but not with 5A-IL139/140MM (S5A Fig.). These data demonstrated that IRF3 nuclear translocation is required for TRIM26-induced degradation.
TRIM26 was translocated into nucleus after viral infection or TLR stimulation (Fig. 1). A recent study also demonstrated that TRIM26 was recruited to histone modifier Jmjd3 to mediate PHF20 ubiquitination and degradation [30]. These data indicate a functional NLS may be present in TRIM26 to facilitate its nuclear translocation. Indeed, PSORT program predicated a putative NLS rkkfwvgkpiarvvkkk between aa 265 and aa 281 in TRIM26. To investigate whether this NLS is responsible for TRIM26 nuclear translocation, the conserved amino acids (RKK and KKK) were mutated to arginine to give rise TRIM26 NLS mutant TRIM26-M-NLS. Fluorescent microscopy showed that TRIM26-M-NLS lost the ability to translocate into the nucleus after SeV and VSV infection (Fig. 6C). Similar to Flag-TRIM26, GFP-TRIM26 overexpression efficiently promoted the degradation of WT IRF3 and 5D (Fig. 6C, lane 1 and 6), but, not of 5A (Fig. 6D, lane 4). Whereas, GFP-TRIM26-M-NLS lost the ability to degrade both WT IRF3 and 5D (Fig. 6D, lane 8 and 12), indicating TRIM26 nuclear localization is required for IRF3 degradation.
To directly confirm nuclear translocation of IRF3 and TRIM26 is required for IRF3 degradation, nuclear import inhibitor Ivermectin was used. Ivermectin treatment prevented the nuclear translocation of IRF3 after SeV infection (S5D Fig.). While, IRF3 phosphorylation was not impaired by Ivermectin treatment (S5E Fig.). Notably, inhibition of IRF3 nuclear translocation by Ivermectin abolished TRIM26-induced IRF3 degradation in SeV-infected cells and uninfected cells (Fig. 6D, lane 4 Vs. lane 2 and lane 8 Vs. lane 5). Taken together, these results suggested that TRIM26 mediates ubiquitination and degradation of active IRF3 in the nucleus.
To investigate the physiological function of TRIM26, TRIM26-transgenic mice (TRIM26-Tg mice) was established. The transgenic mice were identified by PCR assays of genomic DNA from tails of transgenic mice (S6A Fig.). TRIM26-Tg mice are viable, normal in size and without gross physiological or behavioral abnormalities (data now shown). The expression of TRIM26 mRNA and protein in the thymus from the TRIM26-Tg mice was higher than that of the WT littermate (S6B Fig.). Similarly, the expression of TRIM26 mRNA and protein in peritoneal macrophages from TRIM26-Tg mice was higher than that from WT mice before and after SeV infection (S6C Fig.). Consistent with the data of TRIM26-mediated IRF3 degradation, SeV-induced IRF3 phsophorylation was reduced in macrophages from TRIM26-Tg mice, compared to that from WT mice (S6D Fig.). Notably, the level of total IRF3 was also decreased in macrophages from TRIM26-Tg mice. Consistent with function of TRIM26 to mediate IRF3 ubiquitination, more IRF3 ubiquitination was detected in macrophages from TRIM26-Tg mice after SeV infection compared to that from WT mice (S6E Fig.).
Primary peritoneal macrophages from TRIM26-Tg and WT mice were prepared and stimulated with LPS and poly(I:C) or infected with SeV and VSV, the expression of IFN-β mRNA and secretion of IFN-β protein was measured by quantitative RT-PCR and ELISA, respectively. After stimulation with LPS and poly(I:C) or infection with SeV and VSV, macrophages from TRIM26-Tg mice showed less IFN-β expression and secretion, compared to the macrophages from WT mice (Fig. 7A and B). VSV replication in macrophages from TRIM26-Tg mice was greatly increased compared to that from WT mice (Fig. 7C). To test the importance of TRIM26 in vivo, TRIM26-Tg mice and control WT littermates were infected with VSV, and the antiviral immune responses were examined. The amount of IFN-β protein induced by VSV infection was much less in sera, lung and liver of TRIM26-Tg mice than that of VSV-infected WT mice (Fig. 7D). In accordance with reduced IFN-β production, VSV replication in the lungs and livers of TRIM26-Tg mice was higher than WT controls (Fig. 7E). HE staining showed that severe infiltration of immune cells and injury were observed in the lungs of TRIM26-Tg mice, compared to that of WT mice after virus infection (Fig. 7F). Moreover, TRIM26-Tg mice were more susceptible to VSV infection than WT mice (Fig. 7G). These data suggested that TRIM26 is an important negative regulator of IFN-β production and antiviral immune responses, therefore TRIM26-transgenic mice have impaired antiviral response.
Several studies have demonstrated that phosphorylated IRF3 underwent ubiquitination and proteasomal degradation after infection with Sendai virus [23], [25]. However, the identity of the E3 ubiquitin ligase that is responsible for the ubiquitination of nuclear IRF3 is not defined. Here we have provided evidence to show that TRIM26 is an E3 ubiquitin ligase to promote the ubiquitination and degradation of nuclear IRF3. 1) TRIM26 interacted with IRF3 in the nucleus after TLR4 activation and SeV infection in macrophages. Furthermore, TRIM26 was found to interact with WT IRF3 and constitutive phosphorylation active mutant 5D, but not with phosphorylation deficient mutant 5A in co-transfection assays. 2) TRIM26 knockdown by siRNA mainly attenuated SeV-induced IRF3 ubiquitination in nucleus, but, has little effect for the cytoplasmic IRF3 ubiquitination. 3) TRIM26 was found to degrade WT IRF3 and 5D, but not 5A. Further analysis of cytoplasmic and nuclear fractions demonstrated that WT IRF3 and IRF3 5D were degraded by TRIM26 in nucleus, not in the cytoplasm. 4) IRF3 mutants KR77/78NG and 5D-KR77/78NG with mutation in the NLS could not interact with and be degraded by TRIM26. 5) TRIM26 mutant in the NLS, which did not move into nucleus, lost the ability to degrade IRF3. 6) Chemical inhibitor to inhibit IRF3 and TRIM26 nuclear translocation abolished TRIM26-induced IRF3 degradation. All together, these data indicated that active IRF3 was ubiquitinated and degraded by TRIM26 in nucleus.
Normally, IRF3 shuttles between nucleus and cytoplasm and a small portion of IRF3 translocates into nucleus [29]. This small portion of IRF3 may be targeted by TRIM26 to prevent the IFN-β production in normal conditions. This may explain why WT IRF3 was degraded without virus infection and the level of total IRF3 was decreased in TRIM26-Tg mice. In supporting this claim, we found the expression level of IRF3 mutant KR77/78NG in the NLS was much higher than the WT IRF3 in the presence of TRIM26. IRF3 WT and IRF3 5A have the same constitutive shuttling ability between cytoplasm and nucleus, while TRIM26 only promoted WT IRF3 degradation, but not 5A. Clearly, besides the constitutive shuttling between cytoplasm and nucleus, other factors also contribute to TRIM26-mediated IRF3 degradation. We found a considerable proportion of WT IRF3 was translocated into nucleus upon overexpression, which was similar to IRF3 5D. While this was not the case for IRF3 5A, a very small portion of IRF3 5A was transfected into nucleus. The different ability of nuclear translocation between IRF3 WT and 5A may be caused by IRF3 phosphorylation because phosphorylation is the prerequisite for IRF3 nuclear translocation. In fact, overexpression of IRF3 WT could induce IFN-β production and activation of IFN-β promoter in our system, indicating partial of WT IRF3 was phosphorylated and translocated into nucleus upon overexpression. Thus, IRF3 WT was targeted by TRIM26 for degradation because of its ability to translocate into nucleus, while 5A could not be targeted by TRIM26 because of the inability to move into nucleus due to the deficiency of phosphorylation. All together, our data indicated IRF3 phosphorylation and nuclear translocation are required for TRIM26-mediated degradation.
Several E3 ligases have been demonstrated to promote IRF3 ubiquitination and degradation. For example, RAUL, a HECT domain-containing E3 ligase, has been shown to promote ubiquitination and proteasomal degradation of IRF3 and IRF7 [31]. Although RAUL could degrade IRF3 and IRF7 in both nucleus and cytoplasm, RAUL promotes the degradation of the phosphorylation deficient mutant of IRF3 and IRF7, indicating RAUL may mainly promote the degradation of un-active form IRF3 and IRF7. E3 ubiquitin ligase RBCK1 has also shown to promote the degradation and ubiquitination of IRF3 [32]. Interestingly, RBCK1 specifically destabilizes IRF3, but not IRF7, which is different from TRIM26. We found IRF7 was degraded by TRIM26 overexpression (S7A Fig.). IRF9 was also degraded by TRIM26 (S7B Fig.). The E3 ligase Ro52 also known as TRIM21, has reported to target IRF3 for polyubiquitination and proteasomal-mediated degradation [33]. Although TRIM21 interacted with IRF3 after poly(I:C) stimulation in HEK293 cells, where IRF3 was ubiquitinated and whether the active form of IRF3 was targeted are not investigated in this report. We found TRIM21 promoted the degradation of IRF3 WT, 5A and 5D (S8A Fig.). Cullin 1-based ubiquitin ligase has been reported to promote the ubiquitination and degradation of phosphorylated IRF3 [34]. However, siRNA knockdown of Cullin 1 expression did not affect TRIM26-mediated IRF3 degradation (S8B Fig.). These data suggested that there is no competition between Cullin 1 and TRIM26. Therefore, TRIM26 is an essential E3 ubiquitin ligase specifically targeting active IRF3 for degradation in nucleus.
There are more than 70 members in the TRIM protein family. Accumulating results have confirmed that TRIM proteins play an essential role in the regulation of innate immune response by regulating the signal transduction mediated by PRRs [35]-[38]. Using TRIM26 transgenic mice, we demonstrated that overexpression of TRIM26 greatly attenuated IFN-β production in primary peritoneal macrophages after virus infection and TLR activation. Expression of IFN-β also greatly decreased in TRIM26 transgenic mice after virus infection in vivo. As a consequence, TRIM26 transgenic mice showed an impaired ability to inhibit virus replication and were more susceptible to virus infection. These in vitro and in vivo data strongly suggested that TRIM26 is a negative regulator for IFN-β production and antiviral immune responses. However, whether TRIM26 has redundancy to regulate IFN-β production with other TRIM family members requires the construction of TRIM26 deficient mice.
Ubiquitination-mediated degradation of transcription factors is now recognized as an efficient way to regulate transcription and terminate cellular signaling [39], [40]. Here we reported that TRIM26 negatively regulates activation of IRF3 transcription factor downstream of various signaling pathway including TLR signaling, RLR signaling and DNA-mediated signaling. Thus, our results identified a novel pathway to limit virus-induced signaling, inflammation and tissue injury after infection. At the same time, we demonstrated that virus infection increased TRIM26 expression and nuclear translocation, which promoted the ubiquitination and degradation of nuclear IRF3 leading to decreased IFN-β production and antiviral responses. Therefore, virus may induce TRIM26 expression to modulate the IRF3 activation and IFN-β production to facilitate their evasion of the innate immune system.
In summary, we identified a novel function for TRIM26 to negatively regulate IFN-β production and antiviral responses by targets nuclear IRF3 for ubiquitination and degradation. Given the pathological role of IFN-β in SLE and other autoimmune diseases, TRIM26 may be used as a therapeutic target to limit IFN-β overproduction to prevent and cure these diseases. Moreover, TRIM26 may also be used as a target for drug design to prevent the viral invasion of the innate immune responses.
C57BL/6J mice for preparation of peritoneal macrophages were obtained from Joint Ventures Sipper BK Experimental Animal (Shanghai, China). Mouse macrophage cell line RAW264.7, human HEK293 and Hela cells were obtained from American Type Culture Collection (Manassas, VA). HEK293/TLR4 and TLR3 cell lines were obtained from Invivogen (San Diego, CA, USA). Mouse primary peritoneal macrophages were prepared as described [41]. The cells were cultured at 37°C under 5% CO2 in DMEM supplemented with 10% FCS (Invitrogen-Gibco), 100 U/ml penicillin, and 100 μg/ml streptomycin. MG132, Chloroquine, JSH-23, and LPS (Escherichia coli, 055:B5) were purchased from Sigma (St. Louis, MO). poly(I:C), poly(dA:dT) and ISD were purchased from Invivogen (San Diego, CA, USA). LPS, poly(I:C), poly(dA:dT) and ISD were used at a final concentration of 100 ng/ml, 10 μg/ml, 2μg/ml and 10μg/ml, respectively. Ivermectin were purchased from Sigma and used at final concentration of 20μM. IFN-β was from PeproTech (Rocky Hill, New Jersey, USA) and used at a final concentration of 1000 U/ml. The antibodies specific for HA, Ub, TRIM26, STAT1, STAT2, β-actin, GAPDH and protein G agarose used for immunoprecipitation were from Santa Cruz Biotechnology (Santa Cruz, CA). Antibody against mice IFNR1 was purchased from Leinco Technologies (St. Louis, MO). The antibodies specific to Myc, IRF3, TBK1, phospho-IRF3 and PCNA were from Cell Signaling Technology (Beverly, MA). Antibody specific to Cullin 1 was from OriGene Technologies (Rockville, MD). The antibody for Flag and VSV G protein were from Sigma (St. Louis, MO). Their respective horseradish peroxidase-conjugated secondary antibodies were purchased from Santa Cruz Biotechnology (Santa Cruz, CA). Sendai virus was purchased from China Center for Type Culture Collection (Wuhan University, China). Vesicular stomatitis virus (VSV) was provided by Professor Hong Meng (Institute of Basic Medicine, Shandong Academy of Medical Sciences, China).
pCMV6-Flag-TRIM26 expression plasmid was purchased from OriGene (Rockville, MD). GFP-TRIM26 was generated by subcloning of TRIM26 coding sequence into pEGFP-N1 vector (CloneTech, CA). Expression plasmid for cGAS was constructed by subcloning of the coding sequence into corresponding vectors. Mutant plasmids for TRIM26 and IRF3 including TRIM26 C16A, C16/36A, GFP-TRIM26-NLS, IRF3 5D, 5A, KR77/78NG, IL139/140MM, ΔDBD and K70/87R were generated using the KOD-Plus-Mutagenesis kit (Toyobo, Osaka, Japan). All constructs were confirmed by DNA sequencing. HA-TRIM21 expression plasmid was provided by Dr. Chen Wang (Shanghai Institutes for Biological Sciences, Shanghai, China). Flag-IRF7 and IRF9 plasmids were provided by Dr. Alexander Espinosa (Department of Medicine, Weill Cornell Medical College). IFN-β reporter plasmid was provided by Prof. Xuetao Cao (Secondary Military Medical University, China) [42]. The ISRE reporter plasmid was provided by Prof. Hong-bing Shu (Wuhan University, China). Other plasmids used in this study were described previously [43]. For transient transfection of plasmids into RAW264.7 cells, jetPEI reagents were used (Polyplus-transfection). For transient transfection of plasmids into HEK293 cells, lipofectamine 2000 reagents were used (Invitrogen). For transient silencing, duplexes of small interfering RNA were transfected into cells with the Geneporter 2 Transfection Reagent (GTS, San Diego) according to the standard protocol. Target sequences for transient silencing were 5’-CCAAGGACUUCGCCAACAA-3’(siRNA 1) and 5’-GAAGUUCUGGAUUGGGAAA-3’ (siRNA 2) for TRIM26, ‘scrambled’ control sequences were 5’-UUCUCCGAACGUGUCACGU-3’. STAT1 siRNA and STAT2 siRNA were obtained from Santa Cruz Biotechnology (Santa Cruz, CA). Cullin 1 siRNA were from OriGene Technologies (Rockville, MD).
The concentrations of IFN-β in culture supernatants, sera, liver and lung were measured by ELISA Kits (R&D Systems, Minneapolis, MN).
Total RNA was extracted with TRIzol reagent according to the manufacturer’s instructions (Invitrogen). Specific primers used for RT-PCR assays were 5’- ATTCTGAACCACTTGAACACCC-3’, 5’- ATTCCGCCACAATGTACTGC-3’ for mTRIM26, 5’-AGTTACACTGCCTTTGCC-3’, 5’-GTTGAGGACATCTCCCAC-3’ for mIFN-β, and 5’-CAACAAGTGTCTCCTCCAAAT-3’, 5’-TCTCCTCAGGGATGTCAAAG-3’ for hIFN-β. For immunoblot analysis, cells or tissues were lysed with M-PER Protein Extraction Reagent (Pierce, Rockford, IL) supplemented with a protease inhibitor ‘cocktail’. Nuclear proteins and cytoplasmic proteins were extracted by NE-PER Protein Extraction Reagent (Pierce) according to the manufacturer’s instructions. Protein concentrations in the extracts were measured with a bicinchoninic acid assay (Pierce, Rockford, IL) and were made equal with extraction reagent. For immunoprecipitation (IP), whole-cell extracts were collected 36 h after transfection and were lysed in IP buffer containing 1.0% (vol/vol) Nonidet P 40, 50 mM Tris-HCl, pH 7.4, 50 Mm EDTA, 150 mM NaCl, and a protease inhibitor ‘cocktail’ (Merck). After centrifugation for 10 min at 14,000g, supernatants were collected and incubated with protein G Plus-Agrose Immunoprecipitation reagent (Santa Cruz) together with 1 μg corresponding antibodies. After 6 h of incubation, beads were washed five times with IP buffer. Immunoprecipitates were eluted by boiling with 1% (wt/vol) SDS sample buffer. For western blot analysis, immunoprecipitates or whole-cell lysates were loaded and subjected to SDS-PAGE, transferred onto nitrocellulose membranes, and then blotted with specific antibodies. The levels of phosphorylated IRF3 were quantified by measuring the band densitometry, which are normalized to the band densitometry of Actin.
Luciferase activity was measured with the Dual-Luciferase Reporter Assay system according to the manufacturer’s instructions (Promega) as described (43). Data were normalized for transfection efficiency by division of firefly luciferase activity with that of renilla luciferase.
For analysis of the ubiquitination of IRF3 in HEK293 cells, HEK293 cells were transfected with Myc-IRF3, HA-Ub (WT) or HA-Ub mutants and Flag-TRIM26 WT or mutants, and then whole-cell extracts were immunoprecipitated with anti-Myc and analyzed by immunoblot with anti-HA antibody. For analysis of the ubiquitination of IRF3 in macrophages, macrophages were infected with SeV, then whole-cell extracts or cytoplasmic and nuclear fractions were immunoprecipitated with anti-IRF3 and analyzed by immunoblot with anti-Ub antibody.
IRF3, TRIM26 WT and C16/36A mutant proteins were expressed with a TNT Quick Coupled Transcription/Translation System (Promega) according to the instructions of the manufacturer. Binding assays were performed by mixing TRIM26 and IRF3 together, followed by IP with TRIM26 antibody and WB with IRF3 antibody. Ubiquitination was analyzed with an ubiquitination kit (Boston Biochem) following protocols recommended by the manufacturer.
VSV plaque assay was performed as described [43]. Hela cells or macrophages (2×105) were transfected with the indicated plasmids or TRIM26 siRNA for 36 h prior to VSV infection (MOI of 0.1). At 1 h after infection, cells were washed with PBS for three times and then medium was added. The supernatants were harvested at 24 h after washing. The supernatants were diluted and then used to infect confluent HEK293 cells cultured on 24-well plates. At 1 h postinfection, the supernatant was removed, and 3% methylcellulose was overlayed. At 3-days postinfection, overlay was removed; cells were fixed with 4% formaldehyde for 20 min, and stained with 0.2% crystal violetin. Plaques were counted, averaged, and multiplied by the dilution factor to determine viral titer as Pfu/ml. Total cellular RNA was extracted and VSV RNA replicates were examined by Quantitative RT-PCR as described (43). Primers for VSV were as follows: 5’-ACGGCGTACTTCCAGATGG-3’ (sense) and 5’-CTCGGTTCAAGATCCAGGT-3’ (antisense).
HEK293 cells transiently transfected with plasmids encoding GFP-TRIM26 or Flag-TRIM26 were cultured on coverslips for 48 hours. Then the cells were infected with SeV or VSV or stimulated with IFN-β or LPS. For the cells transfected with GFP-TRIM26, cells were examined directly with an Olympus IX71 fluorescence microscope (Olympus Co., Tokyo, Japan). For HEK293/TLR4 cells transfected with Flag-TRIM26 or macrophages, cells were sequentially immunostained first with antibody against Flag or TRIM26 antibody, and then with proper Alexa Fluor 568-conjugated secondary Antibody (Molecular Probes, Invitrogen). DAPI (4’, 6’-diamidino-2-phenylindole hydrochloride; Molecular Probes, Invitrogen) was used to stain nuclei.
Transgenic founder mice expressing TRIM26 gene (TRIM26-Tg mice) were generated by Cyagen Biosciences Inc. (Guangzhou, Guangdong, China) in the FVB background and mated with WT FVB mice to produce mice used in all the experiments. Expression of TRIM26 was under the control of EF1a promoter (elongation factor 1a promoter). TRIM26-Tg mice produced viable offspring. The transgenic mice were identified by PCR of genomic DNA from tails with the following primers: 5’-ACGTAAACGGCCACAAGTTC-3’ (sense), 5’-GATCTTGAAGTTCACCTTGATGC-3’ (antisense). All mice used are 6 to 8 months of age.
TRIM26-Tg or WT mice (female, 6–8 weeks old) were intravenously infected with VSV (5×107 pfu per mouse) as described [44]. The virus titres in lung and liver were determined by standard plaque assays and by measurement of VSV V protein with VSV-G antibody. For the survival experiments, mice were monitored for survival after VSV infection. Lungs from control or virus-infected mice were dissected, fixed in 10% phosphate-buffered formalin, embedded into paraffin, sectioned, stained with hematoxylin-eosin solution and examined by light microscopy for histologic changes.
All data are presented as mean ± S.D. of three or more experiments. Statistical significance was determined with the two-tailed Student’s t-test, with a P value of less than 0.05 considered statistically significant.
All animal experiments were undertaken in accordance with the National Institute of Health Guide for the Care and Use of Laboratory Animals, with the approval of the Scientific Investigation Board of Medical School of Shandong University, Jinan, Shandong Province, China (Permit number: 201401039).
The accession numbers in the UniProtKB/SwissProt database for the proteins in the manuscript are followed: TRIM26, Q12899; IRF3, Q14653; cGAS, Q8N884; IFN-β, P01574; RIG-I, O95786; MAVS, Q7Z434; TRIF, Q8IUC6; TBK1, Q9UHD2; STING, Q86WV6; STAT1, P42225; STAT2, Q6P1X8; VSV-G, P04882.
|
10.1371/journal.pntd.0007411 | Deprivation of dietary fiber enhances susceptibility of mice to cryptosporidiosis | Based on our initial observations showing that mice consuming a probiotic product develop more severe cryptosporidiosis, we investigated the impact of other dietary interventions on the intracellular proliferation of Cryptosporidium parvum and C. tyzzeri in the mouse. Mice were orally infected with oocysts and parasite multiplication measured by quantifying fecal oocyst output. High-throughput sequencing of 16S ribosomal RNA amplicons was used to correlate oocyst output with diet and with the composition of the intestinal microbiota. On average, mice fed a diet without fiber (cellulose, pectin and inulin) developed more severe infections. As expected, a diet without fibers also significantly altered the fecal microbiota. Consistent with these observations, mice fed a prebiotic product sold for human consumption excreted significantly fewer oocysts. The fecal microbiota of mice consuming no plant polysaccharides was characterized by a lower relative abundance of Bacteroidetes bacteria. Since bacterial metabolites play an important role in the physiology of intestinal enterocytes, we hypothesize based on these observations that the impact of diet on parasite proliferation is mediated primarily by the metabolic activity of the anaerobic microbiota, specifically by the effect of certain metabolites on the host. This model is consistent with the metabolic dependence of intracellular stages of the parasite on the host cell. These observations underscore the potential of dietary interventions to alleviate the impact of cryptosporidiosis, particularly in infants at risk of recurrent enteric infections.
| The infection with Cryptosporidium parasite, a condition known as cryptosporidiosis, is a common cause of infant diarrhea in developing countries. We have previously shown that mice infected with C. parvum, one of the main cause of human cryptosporidiosis, develop a more severe infection if given probiotics. To investigate the mechanism of this effect, we fed mice prebiotics and diet lacking plant fiber. We found that fermentable fiber, whether administered as a prebiotic supplement or as part of the diet, has a protective effect against cryptosporidiosis in mice. We also observed a significant association between the severity of infection and the composition of the gut microbiota. A significant inverse correlation was found between severity of cryptosporidiosis and the ratio between the abundance of bacteria belonging to the phylum Bacteroidetes and the abundance of Firmicutes bacteria. This ratio is frequently viewed as a marker of a healthy microbiota. These results raise the possibility that dietary interventions could be used to alleviate the impact of cryptosporidiosis.
| Protozoa of the genus Cryptosporidium are important pathogens causing diarrhea in humans, ruminants and other species of animals worldwide [1]. Various Cryptosporidium species are recognized as opportunistic pathogens in patients with AIDS, where cryptosporidiosis can lead to protracted diarrhea and wasting. Although immunocompetent patients heal spontaneously within a few weeks, recent studies in developing nations have pointed to Cryptosporidium as the second leading cause of infant diarrhea [2, 3].
The resistance of Cryptosporidium parasites to anti-protozoal drugs [4], and the lack of alternative therapeutic options, led us to investigate the interaction between the gut microbiota and the parasite. The previously reported unexpected observation that a probiotic product can aggravate the course of cryptosporidiosis in mice [5] supports the hypothesis that parasite proliferation is impacted by diet and possibly by the effect of diet on the gut microbiota. This observation is significant because it could lead to the development of simple dietary supplements for mitigating cryptosporidiosis and perhaps other enteric infections in vulnerable infants.
The benefits to intestinal health of diets rich in plant fibers are well known [6]. It has been suggested that consumption of fiber below nutritional recommendations [7, 8] may lead to dysbiosis. A decrease in the Bacteroidetes/Firmicutes ratio has often been linked to a poor intestinal health index and to obesity [9]. Dysbiosis may also deplete the intestinal mucosal layer [10]. To what extent mucus depletion may play a role in susceptibility to cryptosporidiosis has not been investigated. Several mechanisms linking diet, microbiota and enteric infections have been proposed [11]. Bacterial metabolites, particularly those originating from the fermentation of certain plant polysaccharides, have been shown to play and important role in modulating the resistance to enteric bacterial infections [12]. Research on the interaction between the microbiota and the intestinal epithelium has shown the importance of bacterial metabolites, such as short-chain fatty acids originating from the anaerobic breakdown of plant polysaccharides [10]. The role of the intestinal microbiota in regulating the immune response and preventing inflammation has also been investigated [11, 13]. With respect to enteric infections, much research has focused the protective role of the microbiota, a phenomenon often referred to as "colonization resistance" [14, 15]. In contrast to what is known about the effect of diet and bacterial metabolites on the intestinal physiology, less research has focused on mechanisms linking diet and enteric infections. This limitation is particularly true for enteric protozoa [16].
With respect to cryptosporidiosis, research with germ-free severe combined immunodeficient (SCID) mice and SCID mice colonized with intestinal microbes conducted by Harp and co-workers showed that a normal intestinal microbiota delayed the onset of C. parvum oocyst excretion by several weeks [17, 18]. A protective role of the gut microbiota against cryptosporidiosis was also observed in neonatal mice [19, 20] A protein-deficient diet was also found to increase susceptibility of mice to C. parvum [21]. This phenotype was attributed to a reduced epithelial cell turnover. The effect of probiotics on the course of cryptosporidiosis was also observed by others [22, 23]. This research uncovered a beneficial effect of Enterococcus faecalis administration to mice infected with C. parvum. None of these studies have investigated potential mechanisms mediating the observed effect on the development of C. parvum.
Here we describe experiments with a mouse model of cryptosporidiosis aimed at investigating changes in the bacterial microbiome caused by dietary fiber and at relating these changes to the severity of cryptosporidiosis. The results show that relatively small changes in diet, or the administration of a prebiotic formulation, can reduce the severity of cryptosporidiosis.
C. parvum strain TU114 oocysts [24] was used in experiment 1 and 4 whereas C. tyzzeri oocysts were used in experiments 2, 3 and 5. C. parvum strain TU114 belongs to the anthroponotic subgroup characterized by a GP60 surface glycoprotein genotype IIc [25, 26]. C. tyzzeri is a species commonly found in domestic mice of the species Mus musculus [27]. Oocysts for the experimental infections were purified from feces of mice on Nycodenz (Alere Technologies, Oslo, Norway) step gradients as previously described [28]. The age of the oocysts was 65, 37, 22, 38 and 13 days for experiments 1, 2, 3, 4 and 5, respectively (Table 1).
To test the effect of dietary fiber, three experiments were performed using no-fiber diet and matched control diet ("medium-fiber diet") (Supplementary Table 1). In experiment 1, 8 female CD-1 mice aged ~5 weeks were randomly divided into two groups and immunosuppressed by adding disodium dexamethasone 21-phosphate (Sigma, cat. D1169) to drinking water at a concentration of 16 mg/L [29]. The immunosuppressive treatment was initiated on the day -5 post-infection (PI), where day 0 is the day of infection. In experiment 2, we used 8 female C57BL/6 mice, also divided into two groups of 4 mice. In experiment 3, 12 female C57BL/6 mice were divided into four groups of 3 mice. In all experiments, mice were provided ad libidum with autoclaved water. In experiments 1 and 2, each group was fed one type of diet and in experiment number 3, two groups ingested medium-fiber diet and two groups no-fiber diet. The diet was given starting on day -5 PI, i.e., 5 days before the animals were infected with Cryptosporidium oocysts. To test the effect of prebiotics on the microbiome and on the excretion of Cryptosporidium oocysts, we performed two experiments. In experiment 4, 16 CD-1 mice, randomly divided into 4 groups of 4 mice, were given normal diet and were immunosuppressed by the addition of dexamethasone to drinking water at a concentration of 16 mg/L. In addition to immunosuppression with dexamethasone, vancomycin and streptomycin were added to drinking water at a concentrations of 500 mg/L and 5 g/L, respectively, starting on day -6 PI. Metronidazole at the dose of 20 mg/kg was given daily by gavage, starting at day 6 PI. Antibiotic treatment was terminated on day 2 PI. The goal of the antibiotic treatment was to deplete the native intestinal microbiome [30], and replicate the treatment used in a previous series of experiments with probiotics [5]. From day -1 PI, the drinking water was supplemented with prebiotic (Supplementary Table 1) at a concentration of 2.8 g/L. Lastly, in experiment 5, 12 immunocompetent C57BL/6 mice divided into four groups were used, two were given prebiotic in the drinking water starting on day -5 PI, and the other two groups drank unsupplemented water. In this experiment all groups ingested medium-fiber diet. Experiments typically lasted 3 weeks.
Upon arrival, each mouse was individually tagged and randomly assigned to a treatment groups (Table 1). Mice were orally infected on day 0 PI with approximately 2 x 104 oocysts of C. parvum strain TU114 (experiment 1 and experiment 4) or C. tyzzeri (experiment 2, 3 and 5). To obtain fecal pellets for intestinal microbiota analysis, mice were individually transferred to a 1-L plastic cup and fecal pellets collected immediately after defecation. The pellets were stored at -20°C. To collect feces for oocyst enumeration using flow cytometry, mice were individually transferred overnight to collection cages fitted with a wire bottom. Feces collected overnight were stored at 4°C. Following overnight fecal collection, mice were returned to regular cages with their original cage mates. On days when feces were collected for oocyst enumeration, mice were individually housed for 14–16 h and spent the remaining time in regular cages with their respective cage mates.
Prior to processing for flow cytometry (FCM), fecal pellets were suspended in water and homogenizing to a slurry. The water volume was adjusted according to the volume of feces and varied between 1.5 ml and 4 ml. A previously described procedure [5] was used to immuno-label oocysts. The only modification consisted in filtering the fecal slurries through a 38-μm opening Nylon mesh, (Component Supply, Sparta, Tennessee, cat. 06725–01) before FCM. For each experiment, 3 samples were randomly selected for replication. Replication involved the processing and labeling of 5 separate aliquots originating from a fecal sample. The labeled samples were analyzed by FCM using a Becton Dickinson Accuri C6 cytometer. Oocyst counts for each mouse were converted to number of oocysts excreted per overnight collection event based on the sample volume analyzed by FCM and sample dilution. These values were summed over the experiment. The cumulative values obtained in this manner are designated "cumulative" oocyst counts. This values represents, for each mouse, the number of oocysts excreted over all collection periods. Feces were collected 6 times per experiment. To test the effect of each dietary treatment (dietary fiber and prebiotics) on oocyst output, cumulative oocyst output for each mouse was normalized against the mean cumulative oocyst output of the control mice. Specifically, the mean cumulative oocyst output of the control mice in each experiment was set equal 100%. Finally, normalized cumulative values were averaged over experiment 1–3 to test for the effect of dietary fiber, and over experiment 4 and 5 to assess the effect of prebiotics. In addition, the effect of treatment was also tested based on the individual FCM data obtained for each mouse and collection event.
To ensure that oocyst counts are not impacted by diet, 3 fecal samples from negative mice fed regular diet and from the same number of mice fed diet without fiber were spiked with the same dose of oocysts and processed for oocyst enumeration. The results showed no significant effect of diet on FCM counts (p = 0.14).
The procedures for DNA extraction, amplicon library construction and bioinformatics were previously described [5, 31]. Briefly, fecal DNA was PCR amplified to prepare amplicons of the V1V2 variable region of the bacterial 16S rRNA gene [32, 33]. The multiplexed amplicon library was size-selected on a Pippin HT system (Sage BioScience, Beverly, Massachusetts) and sequenced in an Illumina MiSeq sequencer at the Tufts University genomics core facility (tucf.org) using single-end 300 nucleotide strategy. To control for technical variation introduced during PCR, library preparation and sequencing, each library included two replicates of two randomly selected samples. Replication involved the separate processing of duplicated fecal samples and tagging each amplicon with a different barcode.
FASTQ formatted sequences were processed using programs found in mothur [34] essentially as described [5, 35]. Briefly, random subsamples of 5000 sequences per sample were processed. Pairwise UniFrac phylogenetic distances [36] between samples were calculated in mothur. Analysis of Similarity (ANOSIM) [37] was used to test the significance of clustering by treatment. Program anosim was run in mothur using a weighted UniFrac distance matrix as input. Operational Taxonomic Units (OTUs) were obtained using program cluster, using the OptiClust clustering method [38]. A distance cut-off of 3% was applied.
Redundancy Analysis (RDA) was used to test the significance of association between OTU profile and oocyst concentration. The program was run in CANOCO [39]. The pseudo-F statistic was calculated by Monte Carlo with 1000 permutations of samples between treatment groups. OTU abundance values for the 150 most abundant OTUs served as dependent variables. Oocyst concentration determined by flow cytometry as described above served as independent variable. Where two experiments were pooled, i.e., experiments 2 and 3, any effect of the experiment was excluded by defining the experiment as covariate. GenAlEx [40] was used to draw Principal Coordinate Analysis (PCoA) plots using weighted UniFrac distance matrices as input. Linear Discriminant Analysis as implemented in program LEfSe [41] was used to identify statistically significant differences in OTU abundance profiles between two groups of samples defined by the dietary treatment.
Sequence data from experiments 1–5 were deposited in the European Nucleotide Archive under study accession numbers PRJEB31954, PRJEB31955, PRJEB31958, PRJEB31959 and PRJEB31960, respectively.
The animal experiments adhered to the National Institutes of Health’s Public Health Service Policy on Humane Care and Use of Laboratory Animals. The animal experiments were approved by the Tufts University Institutional Animal Care and Usage Committee (IACUC). The IACUC approved the experiments described above and as described in document G2016-40.
To test whether no-fiber diet affects the severity of cryptosporidiosis, immunosuppressed and competent mice were infected with C. parvum and C. tyzzeri oocysts, respectively. The intensity of the infection in mice fed medium-fiber diet or regular diet was measured by quantifying oocyst output by FCM 6 times over the duration of each experiment, where each collection event lasted approximately 16 h. In the 3 experiments designed to compare the effect of dietary fiber, mice fed a diet lacking fiber excreted 3.12, 1.97 and 1.64 times more oocysts than the control mice. The effect of dietary treatment tested over the 3 experiments was statistically significant (Mann-Whitney Rank Sum test, U = 20, n = 13, p = 0.001; Table 2). Similarly, the alternative analysis based on individual fecal samples also revealed a significant effect of diet for all three experiments (U = 56, n = 15, p = 0.02; U = 118, n = 24, p = 0.001; U = 392, n = 36, p = 0.004). Fig 1 shows overnight oocyst output over time for the three experiments. We performed analogous experiments to test the effect of prebiotics, which are in essence fermentable fibers (S1 Table). The impact of the dietary supplement on the severity of Cryptosporidium infection was also significant (U = 38, n = 14, p = 0.006; Table 2). This effect was apparent when prebiotics were given to mice fed no-fiber diet (experiment 5, U = 542, n = 47, p = 0.001) or regular diet (experiment 6, U = 305, n = 30, p = 0.033). Fig 2 shows the pattern of oocyst production over time for prebiotic experiments 4 and 5. Body weights were recorded multiple times during each experiment. In none of the experiments was a statistically significant effect of the treatment on the final weight identified.
Based on the results described above and on previously published observations [5], we investigated whether the effect of diet and dietary supplements on cryptosporidiosis could be mediated by the intestinal microbiota. To evaluate this model, the fecal bacterial microbiota was analyzed using 16S amplicon sequencing. Weighted UniFrac distances [36] between pairs of microbiota from each experiment were visualized on PCoA plots (Figs 3 and 4). In experiments 1, 2 and 3, (no-fiber vs. medium-fiber diet), fecal sample collection was initiated on the fifth day after the onset of dietary intake, the day the mice were infected, and continued until day 23 of treatment (day 18 PI). Feces were collected four times during this interval. Demonstrating an effect of diet on the intestinal microbiota, this analysis revealed a non-overlapping distribution of data points according to dietary treatment. ANOSIM R-values between diet groups for the three experiments testing the effect of diet are statistically significant (Table 2). Significant clustering according to prebiotic treatment was observed in experiments 4 and 5 based on 50 and 49 samples, respectively, collected between day 0 PI and day 15 PI. Consistent with a significant effect of the prebiotics, the ANOSIM R-value in experiment 4 was 0.059, (p = 0.046) and in experiment 5 0.202 (p < 0.0001). (Table 2).
Having detected an association between dietary fiber and cumulative oocyst output, and between dietary fiber and fecal microbiota profile, we focused on the microbiome on day 0 PI. We reasoned that if the gut microbiota impacts the severity of cryptosporidiosis, the microbiota on day 0 PI would be the most relevant to examine. Since colonization of the gut epithelium by Cryptosporidium is known to impact the microbiota [35], the analysis of the microbiota on day 0 PI eliminates the effect of cryptosporidiosis on the microbiota and enables detecting any effect of the microbiota on cryptosporidiosis. The effect of no-fiber diet on the fecal microbiota was already detectable after 5 days of treatment (day 0 PI) in experiments 2 and 3 (ANOSIM R = 0.82, p = 0.03; R = 0.74, p = 0.001, respectively). In experiment 1, the effect was not significant (ANOSIM R = 0.17, p = 0.22). Administration of prebiotics in experiments 4 and 5 did not significantly change the microbiota composition according to ANOSIM (R = 0.10, p = 0.06, n = 16; R = 0.06, p = 0.29, n = 12, respectively). To examine to what extent the day 0 microbiota composition correlates with total oocyst output over the course of the infection, we merged experiments 2 and 3, which are exact replicates, to increase the power of the analysis. Experiment 1 was excluded from this analysis because the mice were immunosuppressed, because the microbiota on day 0 did not show any impact of diet and because of mortality only 31 samples were available. For the remaining 4 experiments, we analyzed the correlation between cumulative oocyst output for each mouse and the microbiota OTU profile using RDA. Of the 20 days 0 microbiota samples from pooled experiments 2 and 3, 8 originated from experiment 2 and 12 from experiment 3. Defining the experiment as covariate, a Monte Carlo permutation test indicated a significant correlation between cumulative oocysts output and the OTU profile (pseudo-F = 2.1, p = 0.0354). As expected from the lack of prebiotic effect on the microbiota on day 0, RDA of experiment 4 and 5 showed a non-significant association between day 0 microbiota profile and cumulative oocyst output (pseudo-F = 0.7, p = 0.4688; pseudo-F = 1.3, p = 0.172, respectively).
As expected from the different treatments used in the 5 mouse experiments, the taxonomy of the fecal bacterial microbiota differed extensively between experiments. S1A Fig illustrates the magnitude of the effect of diet and antibiotics pretreatment on the microbiota. As expected, pre-treating mice with antibiotics in experiment 4 profoundly modified the microbiota when compared with microbiota from untreated mice. Removing the data points from experiment 4 from the PCoA reveals the impact of dexamethasone treatment and/or Cryptosporidium species on the microbiota (S1B Fig). Since the experiments were not designed to investigate the effect of these variables, we cannot infer the relative effect of each of these 2 variables on the microbiota. This is because in experiments 1 and 4 mice were immunosuppressed before infecting them with C. parvum, whereas infection with C. tyzzeri in experiments 2, 3 and 5 did not require immunosuppression. The position of experiment 1 data points in S1B Fig also indicates that immunosuppression and/or parasite species has a large effect on the microbiota as compared to diet. Without a direct comparison, it is difficult to infer the effect of untested variables on the microbiota.
The combined samples collected on day 0 PI from experiments 2 and 3 were the primary focus of a taxonomic analysis because of the relatively large sample size (n = 20 mice). Combining these two experiments is consistent with them being exact replicates (Table 1). LDA, as implemented in program LefSe, was used to identify bacterial taxa significantly associated with dietary treatment. This analysis identified 95 taxa significantly more abundant in the no-fiber microbiota and 92 in the medium-fiber microbiota (S2 Table). Of the 95 taxa in the former group, only 24 (25%) belonged in the phylum Bacteroidetes, which compares to 42 (45%) Bacteroidetes taxa in the medium-fiber group. A Chi-square test confirms that Bacteroidetes taxa were significantly enriched in mice consuming medium-fiber diet (χ2 = 8.5, p = 0.003). Given the wide interest in the Bacteroidetes/Firmicutes ratio as a marker of a healthy gut microbiota [42–44], we calculated day 0 Bacteroidetes/Firmicutes from experiment 2/3. As shown in Fig 5, cumulative oocyst output was negatively correlated with Bacteroidetes/Firmicutes (Pearson r = -0.47, p = 0.04; Spearman rs = -0.46, p = 0.04). As expected from the metabolic function of the Bacteroidetes microbiota, mean Bacteroidetes/Firmicutes on day 0 PI was also significantly correlated with diet (mean no-fiber diet = 2.213, mean medium-fiber diet = 3.950; Mann-Whitney U = 305, p = 0.001). In contrast, Firmicutes relative abundance on day 0 was not significantly correlated with cumulative oocyst output (S3 Table). In the other experiments this correlation was not observed on day 0 PI, but calculating the Bacteroidetes/Firmicutes ratio for the entire experiment (all time points), revealed a significant effect of diet in experiment 1 (n = 31, U = 160, p = 0.025) and in experiment 5 (n = 49, U = 181, p = 0.018) As indicated above, this outcome could however be related to the effect of the infection of the microbiota. In experiment 4, we did not observe a significant difference between treatment groups (n = 50, U = 180, p = 0.9). This observation is consistent with the fact that in this experiment the prebiotic supplement was given to mice fed medium-fiber diet. In addition, pretreatment of mice with antibiotics in this experiment profoundly impacted the microbiota (S1 Fig).
Consistent with previously published observations [5], the results presented here show that in the mouse a diet low in fermentable fiber impacts the intestinal microbiota and aggravates the infection with C. parvum and C. tyzzeri. Significantly, this effect was observed in two models, immunosuppressed mice infected with the human pathogen C. parvum and immunocompetent mice infected with the rodent parasite C. tyzzeri. In three experiments performed with customized diet, a statistically significant increase in the elimination of Cryptosporidium oocysts was observed in mice deprived of dietary fiber. The observation that the effect of the treatment on oocyst output did not impact the weight of the mice is likely explained by the fact that diet has a quantitative impact on the infection as opposed to a curative effect.
The benefits to intestinal health of diets rich in plant fibers are well known [12, 45]. Research on the interaction between the microbiota and the intestinal epithelium has revealed the importance of bacterial metabolites, such as SCFAs originating from the breakdown of plant polysaccharides [6]. Elucidating to what extent this interaction can impact the proliferation of an enteric pathogen could lead to the development of simple "nutraceuticals" capable of mitigating the infection. Dietary supplements would have significant advantages over drugs and vaccines, because they are cheap and do not require refrigeration, a significant advantage for distributing to vulnerable populations such as infants in developing countries. Diet could play a role for controlling cryptosporidiosis as no effective anti-cryptosporidial drugs nor vaccine is available. Moreover, such treatments are unlikely to generate resistant parasites. Although statistically significant, the effect of dietary treatments tested to date on the course of cryptosporidiosis is modest. Clearly, more effective treatments are desirable. Eradication of the infection, however, is not necessarily the most desirable outcome. An intervention which prevents diarrhea, while enabling the host to develop immunity, may be as effective at preventing the long-term consequences of recurrent infant diarrhea [46] than a complete cure. Conceivably, dietary treatments could one day be used to enhance the effect of a drug, when it becomes available, and as prophylactics.
A similar study with the enteric protozoan Giardia lamblia concluded that gerbils fed a low-fiber diet were significantly more likely to become infected than animals fed a high-fiber diet [47]. This observation suggests that diet may act directly on the parasite, as Giardia multiplies extracellulary in the intestinal lumen. The observed beneficial effect on the course of giardiasis, suggests that dietary treatments may affect multiple enteric pathogens.
To maximize the beneficial effect of dietary interventions on cryptosporidiosis, a better understanding of the mechanisms linking diet and parasite proliferation in the intestinal epithelium is needed. The increased severity of certain enteric infections in individuals who eat low-fiber diets can be explained by different mechanisms. A low-fiber diet may increase the abundance of bacteria that degrade the intestinal mucus layer. According to this model [10], infection of enterocytes by enteric pathogens could be facilitated by a depleted mucus layer, thought to be one of the main innate defense mechanisms against such pathogens [48, 49]. The effect of diet on parasite proliferation could also be linked to the production of SCFAs or other bacterial metabolites [50–54]. An example of a metabolite which may have such an effect was uncovered in research with human volunteers. Chappell and co-workers detected a significant association between luminal concentration of the bacterial metabolite indole and susceptibility to cryptosporidiosis [55]. Given the metabolic dependence of the parasite on host cell metabolites inferred from the annotation of several Cryptosporidium genomes [56, 57], it is also conceivable that bacterial metabolites could affect the parasite's intracellular proliferation by limiting or increasing the availability of essential molecules in the enterocyte. Metabolomics analyses will be needed to study the actual mechanism linking diet and parasite. The importance of microbial metabolites for epithelial integrity, function and immune function has been demonstrated [58]. Such mechanisms could be relevant to understanding the link between diet, microbiome and Cryptosporidium proliferation.
We previously showed that administration of a probiotic product can aggravate cryptosporidiosis [5]. The prebiotics used here in experiment 4 and 5 are also found in probiotics we already tested, but combined with 14 strains of probiotic bacteria belonging to the genera Lactobacillus, Bifidobacterium and Streptococcus (S1 Table). The observed mitigating effect of the prebiotics in the absence of probiotic bacteria indicates that the aggravating effect of the probiotics product may be caused by probiotic bacteria. Experiments to test the effect of probiotic bacteria, given individually or in different combinations, on the course of cryptosporidiosis may contribute to elucidating the mechanisms of interaction between the gut environment and Cryptosporidium parasites. Such experiments should combine the analysis of the microbiota and metabolites to identify mechanisms linking diet with parasite proliferation. It is interesting to note that a reduction in the severity of the infection in response to prebiotics occurred regardless of the type of diet consumed. Although an effect on cryptosporidiosis was observed in experiments 4 and 5, the impact on the microbiome was more accentuated in experiment 5. This is likely explained by the fact that both experiment 4 groups already ingested fibers with the diet.
To study the link between diet, intestinal microbiota and the course of cryptosporidiosis, fecal transplant experiments into germ-free mice will be needed. Dietary treatments found here to be effective at reducing the severity of cryptosporidiosis in the mouse should also be tested in another model, like the pig [59], to assess the extent to which diarrhea and increased gut motility impacts the effectiveness of the dietary treatment.
|
10.1371/journal.pcbi.1002792 | Experimental Studies and Dynamics Modeling Analysis of the Swimming and Diving of Whirligig Beetles (Coleoptera: Gyrinidae) | Whirligig beetles (Coleoptera, Gyrinidae) can fly through the air, swiftly swim on the surface of water, and quickly dive across the air-water interface. The propulsive efficiency of the species is believed to be one of the highest measured for a thrust generating apparatus within the animal kingdom. The goals of this research were to understand the distinctive biological mechanisms that allow the beetles to swim and dive, while searching for potential bio-inspired robotics applications. Through static and dynamic measurements obtained using a combination of microscopy and high-speed imaging, parameters associated with the morphology and beating kinematics of the whirligig beetle's legs in swimming and diving were obtained. Using data obtained from these experiments, dynamics models of both swimming and diving were developed. Through analysis of simulations conducted using these models it was possible to determine several key principles associated with the swimming and diving processes. First, we determined that curved swimming trajectories were more energy efficient than linear trajectories, which explains why they are more often observed in nature. Second, we concluded that the hind legs were able to propel the beetle farther than the middle legs, and also that the hind legs were able to generate a larger angular velocity than the middle legs. However, analysis of circular swimming trajectories showed that the middle legs were important in maintaining stable trajectories, and thus were necessary for steering. Finally, we discovered that in order for the beetle to transition from swimming to diving, the legs must change the plane in which they beat, which provides the force required to alter the tilt angle of the body necessary to break the surface tension of water. We have further examined how the principles learned from this study may be applied to the design of bio-inspired swimming/diving robots.
| The whirligig beetles belong to the family Gyrinidae, consisting of over 700 species of water beetles. They are characterized by a divided eye, ellipsoidal body, and rapidly swim in circles when alarmed. Perhaps the most interesting characteristic of whirligig beetles is their ability to rapidly swim on the surface of water, and also to quickly transition to diving beneath the surface. In this study, we have measured the key physical parameters that allow whirligig beetles to swim and dive, and have used these values to develop dynamics models of the swimming and diving processes. Based on these models, we were able to analyze how the beetle is capable of making sharp turns, the efficiency of varying leg beating patterns, and the key parameters involved in swimming, as well as diving. We were then able to identify fundamental principles used by the beetle to transition from swimming to diving, and examine how the morphology and “design” of the beetle leads to its ability to rapidly swim and maneuver. Based on the results obtained, we further identified principles and components of the beetle design that could be translated into the development of bio-inspired robotics.
| Few organisms maintain the ability to freely crawl on land, swim in water, and fly through the air; however, the whirligig beetle (Coleoptera Gyrinidae) is able to efficiently maneuver in all three environments [1]. The whirligig beetle also has the fastest measured speed for a swimming insect, while still maintaining the ability to produce very sharp turns. In this study, we will focus on investigating how the whirligig beetle uses its legs to swim on the surface of water, and how it transitions from surface swimming to diving. Ultimately, we will use mathematical models combined with experimental data to quantitatively characterize the detailed kinematics and dynamics for the swimming and diving processes.
The morphology of whirligig beetles is highly adapted for the environment in which they live. As shown in Figure 1A–B, they have divided compound eyes for simultaneously looking above and below the water's surface, a pronounced pair of anterior appendages for grasping prey and climbing, and two pairs of paddle-like legs for swimming [2]. While many studies have attempted to understand the highly efficient swimming motion of whirligig beetles [2], [3], [4], few studies have investigated the swimming mechanism and the transition to diving. The insect's ability to swiftly transition between swimming and diving is particularly interesting for bio-inspiration of swimming/diving robots. In this paper, parameters related to swimming and diving of the whirligig beetle were characterized through experimental analysis, and further used to conduct simulations to answer questions that could not be experimentally verified.
Water-walking arthropods have received much attention in recent years [5], [6], [7], [8], [9], [10], [11], [12], [13], however, less research has been conducted on swimming/diving insects. Despite this fact, whirligig beetles have long fascinated researchers, owing to some of the astounding features of their movement. Previous studies have concluded that whirligig beetles can swim at speeds up to 44.5 body lengths/s with a maximum turning rate of 4428°/s and a maximum centripetal acceleration of 2.86 g [14]. In addition to the incredible speed these insects are able to achieve, the turning radius can be as small as 24% of the body length, and typically 84% of the energy devoted to swimming can be transformed into forward propulsion [1], [15]. This propulsive efficiency is believed to be the highest measured for a thrust generating apparatus within the animal kingdom [3]. Additionally, the swimming mechanics of the whirligig beetle has been studied in terms of wave management and turning performance [1], [4], [16], [17]. Findings from these studies have also shown that whirligig beetles are able to attain high swimming speeds while reducing or avoiding hydroplaning and maximum drag due to their unique leg kinematics and structures. Results from studying the management of fluid and wave resistances produced by the beetles, has also led to a better understanding of the efficiency of their propulsive mechanisms, and what enables the insect to maintain such high speeds. In order to understand the ability of the insect to rapidly and efficiently maneuver, it is necessary to investigate how it uses the unique morphology of its propulsive structures to swim and dive.
One of the morphological adaptations that allow whirligig beetles to rapidly swim on the surface of water is the streamlined ellipsoidal body shape, which minimizes the fluid resistance [14]. The forward propulsive efficiency is further increased by the prevention of lateral movement due to the rigid body of the beetle. Although lateral forces do not make direct contributions to forward propulsion, they assist in this process by increasing stability and maneuverability [18]. To maintain this low drag body shape, the forelegs remain folded underneath the body, preventing drag that would be generated, if they were extended [1]. To further reduce the drag when moving on the surface, the beetle has been reported to have a waxy covering that prevents wetting of the body [19], [20]. But perhaps the most important morphological characteristic of the whirligig beetle in relation to its propulsive efficiency is the design of the swimming legs. Unlike the forelegs, the middle and hind legs have evolved into highly efficient swimming paddles with specialized morphology [3]. As shown in Figure 1C–F, both pairs of the swimming legs, termed middle and hind legs, have a large number of swimming “hairs” that increase the effective contact area generating a larger propulsive force [21]. During the power phase, the middle and hind legs have a contact area about 40 times greater than during the recovery phase [3], [15]. A previous report indicated that the middle legs can also paddle at a frequency up to 25 Hz, with the hind legs beating twice as fast [22]. When the beetle swims in a straight line, the left and right swimming legs beat together with the hind and middle legs beating in an alternating fashion. However, the left and right legs paddle asymmetrically during turning [1].
Another feature of the whirligig beetles' unique motion is its ability to rapidly transition from swimming on the water surface to diving below the surface. While the diving behavior has been observed as a necessary trait for predator avoidance and egg laying, the dynamics associated with the transition from swimming to diving has not been well understood. In fact, the effects of swimming acceleration and body size on the mechanical energy consumptions of diving have only been investigated in a few organisms, such as ducks [23] and marine mammals [24], [25]. In general, the rigid exoskeleton of insects limits the efficiency of diving. In this paper, we will examine how the whirligig beetle can overcome this limitation.
This research combines both experimental analysis of the swimming and diving of the whirligig beetles, and the development of dynamics models to better understand these behaviors. By using parameters obtained from the systematic analysis of high-speed video and microscopic images, dynamics models for both the swimming and diving patterns were developed. Based on simulations from these models, we were able to understand several phenomena that could not be directly observed through experimental studies and further inspire principles that may be used in the design of swimming or diving robots.
In order to build dynamics models for both swimming and diving, it was necessary to determine the parameters related to the dimensions of whirligig beetles. The average beetles' mass (M) was determined to be 10±2 mg by blotting the beetles with filter paper and weighing them on a precision balance. To determine the morphology of the beetles, images were captured while they were floating on the water's surface. Measurements obtained from feature traces were conducted in ImageJ. From the acquired traces, we identified that the beetles had a characteristic body length (Lb) of 5.23±0.31 mm and a width (Wb) of 2.2±0.23 mm, as shown in Figure 2A. Additionally, by tracing the outline of the beetle's body, the contact line length (C) was determined to be 10.93±0.54 mm, and the contact area (Sb) was 7.32±0.14 mm2. Using a similar tracing approach, the depth of the submerged portion of the body (h) was determined to be 0.74±0.14 mm. The average frontal area (Ay) and average side area (Ax) of the beetles in contact with water was calculated by tracing the submerged portion of the body in the y–z and x–z planes as shown in Figure 2B–C, and was determined to be 1.31±0.4 mm2 and 2.65±0.2 mm2, respectively. For the convenience of analysis, we define the x–y–z coordinates as the lateral direction (x-axis), the longitudinal direction (y-axis, the forward direction), and the vertical direction (z-axis) (Figures S1 & S2). Due to the small size of the legs, it was necessary to use a higher magnification imaging system to determine the accurate measurements of their morphology.
As shown in Figure 1C–F, light micrographs of dissected middle and rear legs were analyzed to obtain dimensions for these structures. From the micrographs, the length of the middle legs (Lm) was 2.08±0.08 mm, and the length of the hind legs was (Lh) 2.67±0.03 mm. By conducting polygonal traces of the hind legs when they were outstretched, we have obtained the area of these legs without the swimming laminae extended (Sh−) as 1.08±0.03 mm2. Similarly, analysis of the middle legs showed a reduced area (Sm−) of only 0.66±0.01 mm2. Considering that the true effective area of the swimming legs during the power stroke is dependent on the extension of “swimming laminae” used to increase the surface area [26], it was necessary to measure the increase in area with these structures extended. Due to the small size of the laminae, scanning electron microscopy (SEM) was used to measure the size of these structures, as shown in Figure 3. Based on the SEM data, the average length of the laminae (Llaminae) was 366.98±31.3 µm with an average width (Wlaminae) of 30.84±0.03 µm. It appeared that the laminae on the exterior portion of the leg were longer than those on the interior portion of the legs. Previous studies of Gyrinus indicated that 74 laminae were present on the hind legs and 47 on the middle legs [26], [27]. Using this value for the number of laminae, the effective area of the hind legs with the laminae extended (Sh+) was estimated to be 1.92 mm2. This represents a 77% increase in propulsive area compared to the hind leg without the laminae extended. The area for the middle legs with the laminae extended (Sm+) was 1.19 mm2, representing an 80% increase in propulsive area, when compared to the folded state. Since the laminae are folded in the recovery phase of the beat and only extended in the power phase, this leads to a reduced-drag recovery stroke aiding in the propulsive efficiency. On the other hand, the legs may also be oriented at different angles, so that the maximum area is not perpendicular to the direction in which the beetle is moving. All parameter values obtained from the analysis of the imaging data are summarized in Table 1.
After completing analysis of the static images, experimental measurements were obtained from the high-speed camera system in order to obtain parameters related to the dynamic motion of the beetle. A typical hind leg stroke is shown in Figure 4 and Video S1. Based on our experimental studies, we observed a peak leg speed in the forward direction (Up) of 0.67 m/s for rapid swimming. In addition, the average forward velocity (Uy) of the beetles observed in this study was 0.0936±0.0226 m/s. The maximum forward velocity (Umax) was 0.8 m/s, which is astonishing for such a small organism. The maximum forward speed (Umax) measured is in agreement with that reported by Voise [1]. Important parameters related to the development of the dynamics model for swimming were the values related to the beating motion of the swimming legs, and their positions relative to the center of mass of the beetle. The center of mass of the beetle was assumed to be the center point of the beetle on the x–y axis. Previous studies have shown that the legs during one beating cycle have a maximum sweep of ∼120° around the point of attachment of the leg [3]. To determine the maximum and minimum angles between the negative y axis and the straight line from the center of mass to the acting point of force on the legs, the maximum and minimum angles of both the hind (Øhmax,, Øhmin) and middle (Ømmax,, Ømmin) legs were measured relative to the center of mass using the angle tool in ImageJ. For both pairs of legs, at the minimum angle (Øhmin,, Ømmin), the point at which the legs completed the power stroke, was 0°. This means that upon the completion of the cycle, the legs were parallel to the longitudinal, y axis, of the beetle. The maximum angles relative to the center of mass during beating, however, were 79.5±10.32° for the hind legs (Øhmax) and 120.4±8.14° for the middle legs (Ømmax). Previous studies have determined that the acting point of drag on the leg, essentially the position along the length of the leg where 50% of the torque is generated, occurred at a point approximately 68% from the point of attachment for the leg [26]. This means that the acting point of force along the middle legs was 1.41 mm, and 1.81 mm for the hind legs based on the leg measurements. Due to the error that could occur in manually tracing this position, a relationship was established to allow for a more precise calculation of the distance between the acting point of force on the leg and the center of mass. First, it was necessary to determine the position of attachment of the legs relative to the origin, which was achieved by measuring the distance from the origin to the attachment point of the legs using micrographs of the underside of the beetle (Figure 1B). From these micrographs, the point of attachment for the middle legs (Pm) was 0.44 mm anterior to the origin, whereas the point of attachment of the hind legs (Ph) was 0.53 mm posterior to the origin. Next, the distance from the center of mass to the acting point of drag on the legs at Øhmin and Ømmin, rhmin and rmmin as shown in Figure 2, could be easily calculated as 2.34 mm and 0.97 mm, respectively. Since the distance from the center of mass to the acting point of drag on the legs at Øhmax and Ømmax, rhmax and rmmax, were not linear, calculation of these variables was achieved by using the triangle formed from these angles, the length of the legs at their acting points of force, and the distance from the attachment point to the center of mass of the body. Using these relationships, the triangle could be solved, giving a value of 1.8 mm for the distance from the center of mass to the acting point of drag on the legs at Øhmax (rhmax) and 1.6 mm for the distance from the center of mass to the acting point of drag on the legs at Ømmax (rmmax). All parameter values obtained from the high-speed video analysis are summarized in Table 2.
Weight support plays an important role in diving and swimming of whirligig beetles. Since whirligig beetles have a density greater than water, the body weight must be supported by the combination of the buoyancy force (Fb) and the curvature force (Fc), i.e, Mg = Fb+Fc, where M is the mass of the whirligig beetle, and g is the gravitational constant. The buoyancy force can be calculated by integrating the hydrostatic pressure over the effective body surface area Sb in contact with the water, and is equal to the weight of fluid displaced above the body and inside the contact line C, as shown in Figure 2. From this definition, Fb = ρgVb, where ρ is the density of water at 20°C, 998.2 kg/m3, g is the gravitational constant, 9.8 m/s2, and Vb is the volume of water displaced by the body. Vb is equal to the total submerged volume of the body, or Sbh. Using the measurements obtained from the experiments, the buoyancy force on the beetles was 52 µN. Similarly, the weight of the beetles (Mg) was calculated as 98 µN. From the relationship outlined above, the curvature force due to the surface tension of water, Fc, can be calculated as 46 µN and contributes 46.9% of the total weight support. The percentage of weight supported by the curvature force is much lower than water walking insects due to the submersion of the ventral portion of the beetle, whereas a much smaller total volume is submerged in water walking insects. In fact, in an analysis of the weight distribution of a variety of water striders, >90% of the weight was supported by the curvature force in the majority of species tested [28].
In general, the motion of the legs of aquatic insects is characterized by a high Reynolds number (Re), Re>>1. Using the equation Re = Uw/υ, where U is the peak speed of the object, w is the characteristic leg width, and υ is the kinematic viscosity of water, which is 9.79×10−7 m2/s [7], . Based on the experimental measurements obtained from this study, the Re of the hind legs was 444.8 and the Re of the body was 2042.9. Since the Re of both the hind legs and body were much greater than 1, the inertial forces dominate the flow, allowing us to neglect viscous forces when modeling the dynamics of the beetle.
Similar to the approach taken to study the swimming kinematics of the whirligig beetle on the surface of water, analysis of the diving process of whirligig beetles was conducted using a high-speed video camera. Video S2 shows a typical diving motion from which the parameters related to the diving were obtained. The complete diving process was further divided into a pre-diving stage and a diving stage. The pre-diving stage occurred over the first few leg beats, and was characterized by an oscillation in tilt angle. For the diving stage, the tilt angle constantly increased from the maximum observed in the pre-diving stage. The tilt angle (γ) was defined as the degree of body rotation, where a negative value indicates that the head of the beetle dips toward the water, and a positive value indicates that the beetle's head is raised above the water surface. The maximum change in tilt angle observed during the pre-diving process (γmax) was 10.2°, which can be defined by the difference in the instantaneous maximum and minimum tilt angles produced during one oscillation. After achieving this maximum oscillatory change in tilt angle, the average tilt angle over the remaining oscillations steadily decreased as the beetle further dove, serving as the signal for the initiation of the diving process. During the pre-diving process, the average velocity (Upre) was relatively slow, 0.1 m/s. This average pre-diving speed was generated by four leg beats, leading to a leg beating frequency (fpre) of 52 Hz. From the direct observation of the beating motion of the legs in both the pre-diving and diving motion, we found that the legs beat primarily in the y–z plane, as opposed to beating primarily in the x–y plane, which was a characteristic of swimming. In other words, during the swimming process, the legs beat more along the side of the body, whereas in diving, the legs beat further underneath the body. Not surprisingly, this change in beating direction leads to a reduced sweep range and a slower velocity, while also providing the force necessary for angular rotation around the x axis. This slower speed, combined with an oscillating γ, also allows for a larger wave resistance leading to the formation of a large wave in front of the beetle [1], which will further increase the clockwise rotation by applying a downward force on the anterior portion of the beetle's body. The maximum angular velocity measured in the pre-diving process (ωpre) was 1090°/s. Upon completion of the pre-diving process, the beetle rapidly dove underneath the water surface by increasing its average leg beating frequency (fdiving) to 100 Hz, and realized a maximum leg beating frequency (fmax) of 142 Hz. This represented a 1.92 fold increase in average leg beating frequency from the pre-diving to diving stage. The maximum velocity (Umax) of 0.56 m/s, was attained during the first leg beat in the diving process, while the average velocity during the diving process (Udiving) was much lower, 0.17 m/s, due to the decrease in velocity associated with increasing fluid resistance. In total, 9 full leg beating cycles were completed during the 89 millisecond diving process, compared to 4 full leg beating cycles over 83 ms for the pre-diving process, as shown in Figure 5.
The final set of parameters that could be obtained from the high-speed video of the diving process, were the parameters related to the motion of the legs. Similar to the analysis conducted from the swimming videos, it was necessary to obtain the maximum angle (Ømax) and minimum angle (Ømin) between the negative y axis and the straight line from the center of mass to the acting point of drag on the legs. In essence, this will be a measure of the sweep range of the legs. Whereas this angle was calculated in the x–y plane for swimming, as discussed earlier, the motion of the legs during the diving process necessitate the measurement of these angles in the y–z plane. Using the same procedure described for swimming, the average maximum angle of the hind legs over all leg beats used in diving (Øhmax) was found to be 93.9±10.37°, while the minimum angle (Øhmin) was 26.1±6.94°. This means that the average sweep of the middle legs during the diving process was 61.8° in the y–z plane. Similarly, the maximum angle for the middle legs (Ømmax) was 124.9±12.7°, and the minimum angle (Ømmin) was 43±5.26°. Again, the sweep angle was relative to the origin, center of mass, and thus not the sweep angle from the attachment point of the leg to the body. Using the known distance from the attachment point of the leg to the origin, the distance from the attachment point of the leg to the acting point of the force on the leg, and the angle of the straight line formed from the origin to the acting point of force on the leg, it was possible to solve for the distance between the distance from the center of mass to the acting point of drag on the legs (r), as shown in Figure 2B–C. It was assumed that the acting point of drag on the legs was 1.41 mm for the middle legs, and 1.81 mm for the hind legs based on the values calculated for swimming. Despite the change in the plane of beating from occurring primarily in x–y during swimming to primarily in y–z during diving, this acting point of drag was assumed to remain the same, since the maximum area of both pairs of legs was beating in the y–z plane, since the maximum area of the legs in both planes are assumed to remain equal with the only change being the angle at which the legs strike and not the orientation of the legs as they generate forward propulsion. The point of attachment of both the hind legs and middle legs to the origin was the same as measured for swimming, 0.53 mm posterior for the hind legs and 0.44 mm anterior for the middle legs. Using these known values, the distance from the center of mass to the acting point of drag on the legs at the maximum and minimum angle, Ømax and Ømin, was calculated as, 2.3 mm for rhmin, 1.7 mm for rhmax, 1.1 mm for rmmin, and 1.6 mm for rmmax. The diving parameters obtained from the experimental analysis of the high-speed capture of the diving process are summarized in Table 3.
In order to analyze the swimming and diving processes, simulations were conducted based on the models described above, and the parameters obtained from the experimental analysis listed in Tables 1, 2, and 3. Two types of trajectories were analyzed from the swimming simulations: trajectories that displayed a net forward motion and trajectories that generated a repeating circular shape. These trajectories were chosen since they were the most relevant, based on the typical swimming patterns observed in nature. Similarly, these types of trajectories were the most important for inspiration of robotics principles, as they directly relate to steering and forward propulsion. For simplification, the leg beating patterns used in the swimming simulations are annotated as follows: m or h indicates the beating of the middle or hind leg, and the subscript r or l indicates either the right or left leg. Further, if the legs beat simultaneously, then the notation will be a summation. For example, if the right and left middle legs beat in unison, the notation would be (mr+ml). If the leg beats are followed by one another, then there will be a “,” separating the beats. For the case of a middle right leg beat followed by a hind right leg beat, the notation will be (mr, hr). For all swimming simulations, the duration of the simulation was 2 seconds, and the initial values for velocity (Ux, Uy), angular velocity (ω), and the turning angle of the body (β) were all set to zero.
Out of all simulations conducted, six beating patterns that generated a net forward motion were analyzed. Most trajectories simulated showed a high degree of repetitive motions that led to complex patterns, or no net forward motion. The six forward beating patterns analyzed were (mr+ml), (hr+hl), (hr+hl, mr+ml), (mr, ml), (hr, hl), and (hr+hl, mr, hr+hl, ml), as indicated using the notation described above. The trajectories for these motions are shown in Figure 6 and the values calculated from these trajectories are shown in Table 4. Based on the results obtained from these simulations, the highest forward velocity, 0.5268 m/s, was achieved for (hr+hl, mr+ml), the simultaneous beating of the hind legs followed by the simultaneous beating of the middle legs, while the lowest maximum speed, 0.2473 m/s, was attained for (mr, ml), the alternate beating of the middle legs. Similarly, the greatest net forward motion, 0.8284 m, was observed for (hr+hl, mr+ml) and the least, 0.3445 m, was observed for (mr, ml). However, if we consider these beating motions in terms of efficiency, then the most efficient strategy would be the one that results in the largest net forward motion per leg beat. From a biological perspective, the energy expenditure relative to distance traveled is important, since excessive leg beating will lead to exhaustion. Using this definition of efficiency, the most efficient strategy of the Whirligig beetle was calculated to be (hr+hl), the simultaneous beating of the hind legs, at 17.1 mm/beat, followed by (hr, hl), the alternate beating of the hind legs, at 14.5 mm/beat. What can clearly be concluded from this efficiency value is the importance of the hind legs in forward propulsion. When comparing (hr+hl) to (mr+ml), the average speed and net forward distance traveled using the hind legs is 1.55 times larger. In terms of the strategy that was the most efficient per beat in total distance traveled, (hr, hl), 17.7 mm/beat, was the most efficient followed by (hr+hl), 17.1 mm/beat. In general, over the total distance traveled, strategies using the hind legs only (hr+hl) were 1.55 times more efficient than those with the middle legs only (mr+ml). When considering the total distance traveled per beat as a measure of efficiency, (hr+hl, mr+ml) was the worst strategy 10.3 mm/beat, despite moving the greatest overall distance. It should be noted that this pattern of beating was not observed in the experimental study, and may be the result of the low efficiency of this beating pattern.
The results obtained from the simulations can be interpreted from a biological perspective to understand reasons of the beating patterns, and trajectories observed in nature. One of the hallmarks of the movement of whirligig beetles is their overall rounded trajectories, and the common observation of S-shaped trajectories. Based on the analysis from the simulations, S-shaped trajectories generated by (hr, hl) represent the most energetically favorable strategy for covering a large distance with an efficiency of 17.7 mm/leg beat (Table 4). While this may seem counterintuitive, the anatomical structure of the beetle may dictate the efficiency of this strategy over a linear trajectory. In terms of predator avoidance and escape, an effective “flight” response would allow the beetle to rapidly move away from the perceived threat. In addition, to maximize the distance between the beetle and the threat, a strategy must be chosen that would balance the energetic costs of escape, in terms of both speed and distance traveled. A short beating duration using (hr+hl, mr+ml) would result in a maximum burst speed, but would most likely lead to a tiring of the beetle due to the higher energetic cost. This may explain why linear trajectories have rarely been observed when studying the whirligig beetles. The most efficient strategy that would allow the beetle to cover a large distance with minimal energy expenditure would be the S-shaped trajectory generated by the alternate beating of the hind legs. This strategy would allow the beetle to outdistance the predator while moving along a less predictable trajectory, without leading to exhaustion, and may explain why S-shaped trajectories are commonly observed in nature.
Another result from the analysis of the leg patterns used in swimming was the obvious propulsive advantage of using the hind legs over the middle legs. This may explain why several researchers have noted that the hind legs are often observed beating twice as fast as the middle legs [14]. The results from this study suggest that, rather than a physiological constraint that allows the hind legs to beat twice as fast, the hind legs beat twice as much due to the benefit in overall propulsion efficiency. The middle legs would then be expected to contribute to stability control to maintain a given path and prevent a loss of control during swimming. This will be analyzed in greater detail in the following section.
Five of the simulations displayed an overall circular trajectory as shown in Figure 7. These beating patterns demonstrating an overall circular trajectory were (mr), (hr), (mr, hr), (mr+ml, hr), and (mr, hr+hl), and the values associated with their analysis are shown in Table 5. From these trajectories, we can conclude that the most efficient, in terms of total distance traveled per leg beat was (hr), 30.9 mm/beat, which was also the most efficient out of all the leg beating patterns analyzed in this study. In fact, this was 1.74 times more efficient than the most efficient beating pattern observed from the net forward trajectories (hr, hl) of the beetle. Comparison of the efficiency of the motion generated by patterns (hr) and (mr), the use of only one hind leg or middle leg, led to a 1.74 fold increase in the overall distance traveled per leg beat, similar to the results observed from the forward trajectories. As expected, motion generated by (hr) had a much higher average angular velocity, 4336.52°/s, compared to (mr), 1958.74°/s. This is a 121% fold increase from the hind leg only to middle leg only beating, which is likely due to the attachment point of the rear legs being 0.53 mm posterior to the origin of the body. In addition, the longer length of the hind legs and their larger propulsive area, further have a significant impact on turning. The beating pattern that had the greatest average angular velocity, 4742.54°/s, was the middle right leg followed by the hind right leg (mr, hr). This was in accordance with several studies that have pointed out that the most common leg beating pattern in the circling behavior of whirligig beetles is the beating of the outboard legs, which in the case of the simulation was (mr, hr). Similarly, the values for angular velocity obtained from the simulations were very close to the value of 4428°/s obtained in previous studies. The beating patterns that were most likely to reproduce a truly circular trajectory, were (mr+ml, hr) and (mr, hr+hl), both with nearly the same efficiency, distance traveled, and number of beats. There was however a significant difference in the average angular velocity among these beating patterns, where (mr+ml, hr) was 2.15 times faster than (mr, hr+hl), resulting in a difference in the radius of curvature.
From the analysis of the circling trajectories, we see again that the hind legs were more effective per leg beat, and also were able to produce a larger angular velocity when compared to the middle legs. Unlike the forward trajectories, however, as shown in Figure 7, the trajectory from only the right hind leg beating, (hr), was more chaotic due to the larger increase in the angular velocity. The most common trends observed in nature were the circling trajectories produced by (mr+ml, hr), the simultaneous beating of the middle legs followed by the beating of the right hind leg, and (mr, hr+hl), a middle right leg beat followed by the simultaneous beating of the hind legs. By using the middle legs to balance out the large angular velocity of the hind legs, it is possible to get a much more stable path. The exact reason that the beetles prefer rapid circular trajectories remains unknown; however, from this study, it can be concluded that these patterns appear to be more energy efficient than other motions. Biologically, the energy conserved by using these motions may dictate their use for both predator avoidance, and prey capture.
To determine the effect of errors associated with the measurement of the swimming parameters obtained experimentally, parametric analysis was conducted to examine how perturbation of these terms affected the swimming simulations. In order to maintain realistic variation among these terms, they were perturbed by the percent change associated with the standard deviations reported in Table 1. The complete results from this parametric analysis are included in the Text S1.
A numerical study of the diving process was also conducted to validate the dynamics model for the diving process. The values of morphological parameters were given in Table 1. The parameters obtained for diving are listed in Table 3. The initial values of the forward velocity, the angular velocity, and the turning angle of the body were chosen to be 0.17 m/s, −333°/s, and −7°, which were consistent with the parameter values obtained at the start of the diving process, based on the experimental study. The initial depth of the body under the free surface of water was set to be 0.74 mm, with the assumption that the center of mass was in the plane of free surface at t = 0. The 89 ms diving process was simulated, as shown in Figure 5, using the body coordinate system. For diving, the hind legs beat together with a speed of 0.18 m/s, followed by simultaneous middle leg striking at 0.14 m/s. The leg beating pattern that was observed in the diving video, and used in the simulation, contained 6 hind legs beats and 3 middle legs beats. Additionally, the timing of the leg beats used for the simulation was the same as that observed in the diving video. Despite careful analysis of the diving videos, determination of the dynamic changes in the buoyancy and curvature force during diving proved difficult. Since these forces could not be obtained experimentally, it was necessary to establish a coefficient to explain the dynamic changes in these forces. As such, the combination of the buoyancy and curvature force was estimated to change nonlinearly according to the segmented function This segmented function was established to explain the slow change in tilt angle over the first 60 ms, followed by the rapid change in tilt angle during the final 29 ms, as shown in Figure 5. The change in tilt angle will have a large impact on the value for the curvature force term, since the contact line length will change correspondingly, and the buoyancy force will more gradually increase to a maximum when the beetle is completely submerged. Using this approximation as the change in both curvature and buoyancy forces, a simulation result was attained that closely followed the actual diving trajectory observed in the experimental study, Figure 8. This simulation used the initial values from the experimental study, as indicated above. To determine if any of these parameters had a crucial effect on the beetle's ability to dive, further simulations were conducted by varying the initial values of the forward velocity, angular velocity, turning angle of the body, hind leg speed, and middle leg speed.
The initial conditions of forward velocity, angular velocity, tilt angle of the body, hind leg speed, and middle leg speed were 0.17 m/s, −333°/s, −7°, 0.18 m/s, and 0.14 m/s, respectively. For each parameter, the value was varied ±30% of the initial value, while the other values were kept constant. A 30% increase in any of the following velocity terms, forward velocity, hind leg speed, and middle leg speed, led to an increase in the distance traveled in the y-axis and the depth of the dive in the z-axis. Of these parameters, the middle leg speed had the smallest effect on the overall diving trajectory. This was due to the leg beating pattern modeled, where only 3 middle leg beats occurred, compared to 6 for the hind legs. The lower propulsive area of the middle legs also led to only a minimal change in the observed diving trajectories for ±30%. The parameter that had the largest effect on the diving trajectory was the hind leg speed, since these legs have a large effect on propulsion, and there are twice as many beats in the diving simulation. For the larger hind leg speed (the solid magenta line), the beetle will dive farther in y 13.57% and deeper 37.6%, compared to the simulated initial trajectory. Unlike the previous velocity terms, increases in the angular terms, tilt angle and angular velocity, led to diving trajectories with decreased distance traveled in y, 4.53% and 6.37%, respectively. However, increases in these terms still led to an increase in the depth of the dive. Biologically this can be explained, since the increase in the angular terms equates to the beetle more rapidly entering the water, leading to a deeper dive with a decrease in the length of the dive (distance in y). Overall, the simulation data showed that changes in any of the initial terms used for the simulation played only a minor role compared to dynamic changes in the buoyancy and curvature forces. With 2D data, it was not possible to determine the absolute dynamic changes in these terms, thus, a 3D setup must be used to obtain experimental values for these forces.
As discussed above, the high speed of swimming and diving, large angular velocity of the body in swimming, and unique strategy for maximizing the effective area during a propulsive stroke, are key features for bio-inspiration of robot design. Based on the simulations, we were able to determine that with the morphology present in the whirligig beetles, it was energetically more efficient to use an alternating beating of the hind legs, even though this beating pattern does not move in a straight line. This is consistent with experimental observations of curved trajectories being the most common and can be used for potential propulsion system design in small swimming robots under similar conditions. The S-shaped swimming motion, commonly observed in the wild, thus represents a more efficient strategy than linear motion, assuming a robot with the same size and morphology as the beetle. This is a very interesting phenomenon and could be used for small swimming robot path planning to improve energy efficiency. Similarly, the attachment point of the legs relative to the origin, allows the beetle to attain an incredible angular velocity by paddling the outboard legs in an alternating fashion. Another important source of inspiration for robotic design is the morphology of the hind leg, which is crucial in both propulsion and turning. By minimizing the drag in the recovery stroke, and maximizing the effective propulsive area in the power stroke, the whirligig beetle is able to achieve rapid speeds with a highly efficient motion. By designing an oar, with similar morphology to the hind legs of the whirligig beetle, it would be expected that the swimming device would achieve much more efficient propulsion. Many studies have sought to develop biomimetic robots to achieve a goal similar to that of their biological counterparts. These robots include water-walking robots based on the water strider [29], [30], [31], snake inspired robots [32], [33], and even wall climbing gecko robots [34], [35]. Based on the highly efficient design of the whirligig beetle's legs, we envision a bio-inspired robot that can mimic the design of both the hind and middle legs of the beetle.
Similarly, by using comparable body geometry, the robot would be able to achieve a high angular velocity by beating the rear paddles, while adjusting its trajectory, in effect, steering with the middle paddles. In addition, the motion of the legs during both swimming and diving provides a source of inspiration. In swimming, the legs beat predominantly in the x–y plane, whereas in diving, the legs beat predominantly in the y–z plane. By changing the plane of the beating motion, the beetle is able to achieve an angular rotation in the y–z plane relative to the surface of the water, creating an angle that will allow it to break the surface tension of the water and dive. This seemingly subtle change has potential applications in robot design. For example, by designing a robot with similar leg structures, the robot would be able to swim or dive using the same propulsive structures, which are dependent on the plane of beating. The ability to adjust this angle of the leg beating plane during swimming would also enable a much larger set of trajectories offering more precise and finer control over the desired swimming path. The pattern of the leg beating used by the beetle to alter its angle of the leg beating plane, and eventually break the surface tension, may also represent an ideal pattern that could be implemented into a diving robot.
Yet another source of inspiration comes from the ability of the beetle to maintain an overall ellipsoid shape when the legs are not beating. As shown in Figures 1 and 3, the beetles could fold the legs underneath the body, thus reducing drag. In terms of the body, the point of attachment of the legs allows them to swing out away from the body during beating, but to return underneath the body when not beating. This allows the beetle to effectively coast after each beat, which conserves energy. This phenomenon was observed in nearly all beetle studies, where after a beat, there is a period of time where the beetle will decelerate and “coast” prior to initiating another beat. This was also typically observed in turning, where either a right or left leg would beat, and the beetle would continue to rotate without further beating. Previous robotics studies have sought to mimic the morphology of other biological organisms to design more advanced and efficient robots [36], [37], [38]. By designing a swimming robot that can effectively coast and reduce drag when not using its propelling structures, it would be possible to reduce the energetic costs required to move the robot over an equal distance, leading to a more efficient strategy and utilization of its energy.
By integrating experimental studies and theoretical analysis, this research has made several contributions to the study of dynamics and kinematics involved in the swimming and diving of whirligig beetles with potential applications in bio-inspired robotics. First, it was discovered that the whirligig beetle dives by altering the plane in which the legs are beating, from x–y in swimming, to y–z in diving. The dynamics model developed in this study further supports this claim. Second, results from the swimming dynamics model demonstrated that the most efficient strategy for net forward motion was the beating of the hind legs simultaneously (hl+hr). However, the most efficient beating over the total distance traveled was observed in S-shaped swimming (hr, hl), alternating beating of the hind legs, and circling (mr, hr), alternating beating of the middle right and hind right legs, (mr), the beating of the middle right leg only, and (hr), the beating of the hind right leg only. This finding explains why these swimming trajectories are most commonly observed in nature. Third, analysis of the beating patterns used in the generation of circular trajectories showed that the largest average angular velocity was attained by the beating of a middle leg followed by the beating of a hind leg on the outboard side of the turn. This was consistent with the experimental observations of circular swimming seen in the wild, where this was the most common beating pattern. Comparison of simulations using either the hind or middle legs showed that the hind legs were able to generate larger forward propulsion, and also a greater turning angle when compared to the middle legs. This led to the conclusion that the middle legs serve mainly to control the stability of the beetle, and for path correction. This was confirmed by the generation of stable circular trajectories with middle leg beating, compared to unstable trajectories observed as shown Figure 5 without the presence of the middle legs.
Based on the results obtained from this study several key points of inspiration were identified related to the design of swimming/diving robots. The unique morphology of the legs, allowing for greater increase in area during the power stroke, through the use of collapsible laminae, may lead to the design of more advanced paddles or oars. Next, the ability of the legs to fold underneath the body and maintain an ellipsoidal body shape, reduces the drag on the beetle and allows it to effectively coast, preventing the need for constant beating. Finally, by changing the plane of beating of the legs, an angular rotation can be created that provides the angle necessary for penetration below the surface, essentially diving. By combining these principles, it may be possible to build a more efficient bio-inspired swimming/diving robot.
The whirligig beetles were collected from the Tennessee River and maintained in an aquarium at room temperature. A system consisting of several components was assembled to generate a platform for high-contrast imaging of the beating legs. A Powerview HS-650 (TSI, Inc., Shoreview, MN) particle tracking camera with a Sigma 18–200 lens was used to capture the leg beating pattern and swimming motion at more than 800 fps. Using this setup, we were able to track the leg beating pattern throughout its entire movement. Other components of the system included Camware (PCO AG) and ImageJ (NIH), a motion analysis software package. Camware allowed for the playback of videos at different controllable rates. The speed of each leg in both the swimming and diving processes was analyzed by conducting traces of the movement over individual frames using ImageJ. An SEM study was also conducted on the model LEO 1525 from Carl Zeiss equipped with a ‘Gemini’ column.
The swimming of whirligig beetles has been well-studied, with all previous studies operating under the assumption that the body is rigid, with no flexibility, and that the legs behave as “rigid paddles” or “swimming blades” during a swimming stroke [1], [2], [4], [5], [8]. While the legs are actually separated into 3 distinct segments (femur, tibia, and tarsi), we define a swimming stroke as starting, when the leg is completely unfolded and extended from underneath the body, and then terminating when the leg begins to fold back underneath the body and returns to its starting position. Using this definition, the angular sweep of a single stroke can be calculated as an arc, as illustrated in [4], and Figure 2 of the manuscript. Using this definition of a stroke is consistent with the previous works [4], [8], the flexibility of the legs during a swimming stoke can be neglected. Similarly, the swimming laminae may exhibit some minor flexion, but for the purposes of this study, this flexibility is negligible.
In this work two models were developed to study the swimming and diving of Whirligig beetles. Both of the models were used to obtain simulations of these processes. The complete description of the models and code are contained in the Text S1; however, the key factors involved in the development of the models are identified in this section. In the swimming model, the key hydrodynamic forces involved in the model are the fluid resistive force in x and y (frx and fry), and the drag force of the legs in x and y (fmx, fmy, fhx, fhy). Since the swimming model is 2D, and the beetle is assumed to always be on the surface of the water, the buoyancy and curvature forces, Fb and Fc, are neglected because they do not oppose the direction of motion. Unlike Fb and Fc, however, wave resistance can significantly affect the motion of the beetle. Previous studies have shown that wave resistance and fluid resistance are of similar magnitude for whirligig beetles, however, all current models for calculating wave resistance assume an absence of contact between the object and the water surface, quantification of wave resistance is not currently possible based on this assumption [1]. Following the model for wave resistance provided by [39], and assuming that the beetle has no size, but a defined weight, the magnitude of wave resistance would be discontinuous producing a maximum value at 23 cm/s, a value of 0 at speeds <23 cm/s, and decrease exponentially at speeds >23 cm/s [1].
Considering that the diving model functions in the y–z plane, as opposed to the x–y plane of the swimming model, the hydrodynamic forces considered in this model differ from those defined in the swimming model. The fluid resistive forces in y and z (fry and frz), and the drag force of the legs in y and z (fmy, fmz, fhy, fhz) are considered, similar to the swimming model. Again, wave resistance was neglected from the diving model for the same reasons as the swimming model, described above. However, since diving acts in opposition to both buoyancy and curvature forces, these forces must be considered in the diving model. As described above in the analysis of the diving simulations, due to limitations in experimentally determining the dynamic changes in the buoyancy and curvature forces, a segment function (fseg) was created to account for these forces. This segment function accounted for the slow change in tilt angle over the first 60 ms of the diving process, compared to the rapid change in tilt angle during the final 29 ms of the diving process. The change in tilt angle is related to the curvature force, by changing the contact line length, and thus this approximation was used in simulations of the diving process.
In both models, the forces generated by the creation of vortices from the movement of the legs and body have been neglected due to the use of an experimentally measured drag coefficient. The drag coefficient of the beetle, Cdb, used in this study was obtained from [26], where this parameter was calculated for a variety of Whirligig beetles using both wind tunnel and water channel experiments. In these experimental studies, the value for Cdb accounts for the vortices created by the beetle. To further confirm that the coefficient of drag accounted for the formation of the resulting vortices, we calculated the force from the body vortices of the Whirligig beetles used in this study as outlined in the equations provided by [40]. The force from the formation of vortices by the body was found to be 18 µN, while our calculation of the drag force using the coefficient of drag at the same velocity was 24 µN. Considering that other factors were involved in the drag calculation used in this study, the values indicate that the force from the formation of vortices has been accounted for in our drag calculation. Further, other studies have used the drag coefficient from [26] to calculate the drag from Whirligig beetles, and obtained values similar to those obtained in this study [1].
|
10.1371/journal.pgen.1001270 | Friedreich's Ataxia (GAA)n•(TTC)n Repeats Strongly Stimulate Mitotic Crossovers in Saccharomyces cerevisae | Expansions of trinucleotide GAA•TTC tracts are associated with the human disease Friedreich's ataxia, and long GAA•TTC tracts elevate genome instability in yeast. We show that tracts of (GAA)230•(TTC)230 stimulate mitotic crossovers in yeast about 10,000-fold relative to a “normal” DNA sequence; (GAA)n•(TTC)n tracts, however, do not significantly elevate meiotic recombination. Most of the mitotic crossovers are associated with a region of non-reciprocal transfer of information (gene conversion). The major class of recombination events stimulated by (GAA)n•(TTC)n tracts is a tract-associated double-strand break (DSB) that occurs in unreplicated chromosomes, likely in G1 of the cell cycle. These findings indicate that (GAA)n•(TTC)n tracts can be a potent source of loss of heterozygosity in yeast.
| Although meiotic recombination has been much more studied than mitotic recombination, mitotic recombination is a universal property. Meiotic recombination rates are quite variable within the genome, with some chromosomal regions (hotspots) having much higher levels of exchange than other regions (coldspots). For mitotic recombination, although some types of DNA sequences are known to be associated with elevated recombination rates (highly-transcribed genes, inverted repeated sequences), relatively few hotspots have been described. In this report, we show that a 690 base pair region consisting of 230 copies of the (GAA)n•(TTC)n trinucleotide repeat stimulates mitotic crossovers in yeast 10,000-fold more strongly than an “average” yeast sequence. This sequence is a preferred site for chromosome breakage in stationary phase yeast cells. Our findings may be relevant to understanding the expansions of the (GAA)n•(TTC)n trinucleotide repeat tracts that are associated with the human disease Friedreich's ataxia.
| Several inherited human diseases are a consequence of the expansion of trinucleotide tracts [1], [2]. Although the mechanism by which tract expansions are generated is not yet understood, most of the trinucleotide tracts prone to expansion can form secondary structures such as “hairpin-like” DNA (intrastrand pairing) or triplexes (intramolecular pairing events involving complexes with three paired strands). Friedreich's ataxia is caused by expansion of tracts of the trinucleotide GAA•TTC, a sequence that is associated with triplex formation [3].
In the yeast Saccharomyces cerevisiae, (GAA)n•(TTC)n tracts greater than 40 repeats in length result in an orientation-dependent stall of the replication fork [4], [5]. The stall of the replication fork is observed when the (GAA)n sequence is located on the lagging strand template. Long (GAA)n•(TTC)n tracts have high frequencies of contractions and expansions (primarily contractions) in both orientations, although these alterations are somewhat more frequent when the (GAA)n sequences are on the lagging strand template; in our subsequent discussion, we will refer to tracts in this orientation as (GAA)n tracts and the same sequence in the opposite orientation as (TTC)n tracts. The poly(GAA) tracts are also associated with a high rate of double-stranded DNA breaks (DSBs) and a high rate of terminal chromosome deletions [5]. In addition, (GAA)230 tracts stimulate ectopic recombination between lys2 heteroalles 200-fold more than (TTC)230 tracts [5]. In contrast to the strong orientation-dependence observed in studies of replication fork stalling, DSB formation, and ectopic recombination, the frequency of large-scale expansions of the long (GAA)n•(TTC)n tracts is affected only slightly by tract orientation [6].
In addition to studies done in yeast, the properties of (GAA)n•(TTC)n repeats were also examined in bacterial and mammalian systems. In E. coli, (GAA)n•(TTC)n tracts stimulate plasmid-plasmid recombination by a mechanism that is dependent on both the orientation and length of the repetitive tract [7]. In mammalian cells, length-dependent expansions of (GAA)n•(TTC)n and (CTG)n•(CAG)n tracts are observed; these expansions are stimulated by transcription, and are observed in non-dividing cells, indicating that they are not initiated by stalled replication forks [8]–[10].
The yeast studies of (GAA)n•(TTC)n tracts described above were done in haploid strains. In the analysis described below, we examined the properties of long (230 repeats) and short (20 repeats) tracts on reciprocal mitotic crossovers (RCOs) between homologous chromosomes in diploids. The diploid strains described in the Results section allow the selection and mapping of mitotic crossovers. In addition, crossovers are often associated with gene conversion events, the local non-reciprocal transfer of information near the site of the crossover [11], [12]. Most meiotic gene conversion events reflect heteroduplex formation between allelic sequences, followed by repair of the resulting mismatch [11], [13]. During meiotic recombination in yeast, the length of a gene conversion tract is usually about 1–2 kb [14], although mitotic conversion tracts are often much longer with a median length of 7 kb [15]. In our study, both crossovers and conversion events were mapped.
We find a strong stimulation of RCOs for long (230-repeat), but not short (20-repeat) tracts. This hotspot activity is observed in strains heterozygous, as well as homozygous, for the long tracts, and this stimulation is not substantially affected by the orientation of the tract relative to the replication origin. Analysis of the recombination events suggests that the recombinogenic property of the long tracts is a consequence of a double-strand DNA break (DSB) formed within an unreplicated chromosome.
The method allowing the selection and mapping of crossovers and associated gene conversion events is shown in Figure 1 [15]–[17]. A G2-associated RCO can generate two daughter cells that are homozygous for markers that were heterozygous in the starting diploid strain. On one copy of chromosome V, the diploid has the can1-100 allele, an ochre-suppressible mutation in a gene regulating sensitivity to canavanine; yeast strains with the wild-type CAN1 allele are killed by this drug. On the other copy of chromosome V, the CAN1 gene has been deleted and replaced by SUP4-o, a tRNA gene encoding an ochre suppressor. In addition, the diploid is homozygous for ade2-1, also an ochre mutation. In the absence of an ochre suppressor, ade2-1 strains are adenine auxotrophs and form red colonies as a consequence of accumulation of a pigmented precursor to adenine [18]. The starting diploid strain is canavanine-sensitive (CanS), and forms white colonies. A RCO can be selected as a red/white sectored canavine-resistant colony.
In Figure 1, we show only one of the two possible segregation patterns, the one in which the recombined chromosomes segregate with the unrecombined chromosomes. If the two recombined chromosomes segregate into one daughter cell and the two unrecombined chromosomes segregate into the other, no canavanine-resistant sectored colony will be observed. In S. cerevisiae, these two segregation patterns are equally frequent [19]. Thus, the rate of RCOs is equivalent to twice the frequency of CanR sectored colonies in the 120 kb CEN5-can1-100/SUP4-o interval [16].
By constructing diploid strains from haploids with diverged sequences, Lee et al. [15] used single-nucleotide polymorphisms (SNPs) located on chromosome V to map recombination events. Thirty-four polymorphisms that altered restriction enzyme recognition sites were used. Genomic DNA from each sector of a red/white CanR colony was purified and used as a template to generate PCR products containing the SNPs. By treating these fragments with diagnostic restriction enzymes, followed by gel electrophoresis, Lee et al. [15], [17] could determine whether the sector was homozygous or heterozygous for the polymorphism.
As described in the Introduction, crossovers are frequently associated with gene conversion events. For example, in Figure 2A, we show conversion of one of the polymorphic sites adjacent to the RCO, resulting in the converted allele being found in three of the four chromosomes involved in the initial exchange; this type of event is termed a “3∶1” conversion. These events can be detected by examining the markers in both sectors of a sectored colony. In addition to 3∶1 conversion tracts (Figure 2A), in analyzing spontaneous mitotic crossovers, Lee et al. [15] also found two other types of conversion tracts: 4∶0 tracts (Figure 2B) and 3∶1/4∶0 hybrid tracts (Figure 2C). These events are likely to reflect a DSB in one homologue in G1 of the cell cycle, followed by replication of the broken chromosome, and repair of two broken chromatids in G2. Replication of a chromosome broken in G1 is an expected outcome, since single DSBs formed in G1 do not activate the DNA damage checkpoint machinery [20] and are inefficiently processed to recombination intermediates [21], [22]. If the conversion tracts associated with repair of both DSBs include the same markers, a 4∶0 event is generated. If one conversion tract is more extensive than the other, a hybrid 3∶1/4∶0 event would be observed. This explanation of the spontaneous mitotic RCOs and associated conversions is supported by the observation that the RCOs resulting from gamma-radiation of G1-synchronized yeast cells have 4∶0 and 3∶1/4∶0 hybrid tracts, whereas cells irradiated in G2 do not [17].
An alternative explanation of the 4∶0 and 3∶1/4∶0 hybrid tracts is that they represent two independent repair events of DSBs generated in G2. The rate of RCOs in WXT46 is 8.5×10−5/division (Table 1). Of the 29 conversion events associated with the RCOs, 8 were 3∶1 events and 21 were 4∶0 or 3∶1/4∶0 hybrid tracts. If the 3∶1 events are interpreted as the frequency of single repair events in G2, we calculate that the frequency of single events is about 2.3×10−5 ([8/29] × [8.5×10−5]). The expected frequency of independent double events would be (2.3×10−5)2 or about 5.3×10−10. The observed frequency of “double events” (conversion events of the 4∶0 or 3∶1/4∶0 classes) was 5.3×10−5. We conclude, therefore, that the 4∶0 and 3∶1/4∶0 hybrid tracts do not reflect two independent cycles of DSB formation and DSB repair.
Previously, we used the system shown in Figure 1 and Figure 2 to measure the frequency and location of spontaneous or gamma-ray-induced recombination events in the 120 kb interval between CEN5 and the can1-100/SUP4-o markers on chromosome V. In the current study, we constructed yeast strains with insertions of (GAA)n•(TTC)n tracts of two different sizes (230 and 20 repeats) in two different orientations near the URA3 gene on chromosome V (details of the constructions in Text S1). In the strains used in our study, the (GAA)n•(TTC)n tracts are embedded within lys2 sequences inserted in the intergenic region between GEA2 and URA3. This position is about 22 kb centromere-proximal to ARS508 and about 31 kb centromere-distal to ARS510; both of these ARS elements are active origins [23].
In previous studies [4], [5], it was shown that long (>100-repeat) (GAA)n tracts on the lagging strand template result in a replication fork block whereas long TTC tracts on the lagging strand template do not. To determine how replication forks were blocked for strains with the (GAA)n•(TTC)n tracts inserted on chromosome V, we constructed two isogenic haploid strains in which a (GAA)230•(TTC)230 tract was inserted in two orientations. In the haploid MD512, the tract was oriented such that the GAA sequence was on the “Watson” strand as designated in Saccharomyces Genome Database, and the haploid MD510 had the tract in the opposite orientation. By two-dimensional gel electrophoresis, we found a blocked replication fork in MD510 but not in MD512 (Figure 3). Since a replication fork initiated at ARS510 would encounter the GAA tract on the lagging strand in MD510, this result suggests that tracts are replicated primarily by a replication fork initiated at ARS510 rather than ARS508, although we have not directly examined fork movement. In our subsequent discussion of yeast strains, tracts oriented in the same direction as MD510 will be termed “(GAA)n” tracts and those with the opposite orientation will be termed “(TTC)n” tracts; this nomenclature is consistent with previous studies [5]. It should be noted that, in other genetic backgrounds, the chromosomal region in which we inserted the (GAA)n•(TTC)n tracts is replicated using forks that move in the opposite direction from the one observed in our genetic background [24].
We first performed a pilot experiment to examine the recombinogenic effects of (GAA)230 and (TTC)230 tracts in diploids heterozygous for insertion near URA3. As described above, the rate of RCOs in the CEN5-can1-100/SUP4-o interval can be calculated from the frequency of CanR red/white sectored colonies. The rates of RCOs in MD506 (heterozygous for the [GAA]230 tract) and MD508 (heterozygous for the [TTC]230 tract) were 13×10−5/division (±3×10−5) and 6.2×10−5/division (±2×10−5), respectively; 95% confidence limits are shown in parentheses. The rate of RCOs in an isogenic diploid without the tract insertion is 5.8×10−6/division [15]. Thus, the heterozygous tract insertions stimulated RCOs in the CEN5 to can1-100/SUP4-o interval by about 10- to 20-fold and tracts in both orientations were recombinogenic.
The diploids MD506 and MD508 did not have the polymorphisms required to map the recombination events (details of their genotypes in Text S1 and Table S1). Consequently, we constructed six other diploids that were heterozygous for polymorphisms that allowed mapping of RCOs and associated conversions. The strain names, and their tract sizes and orientations are: WXTMD42, (GAA)20/(GAA)20; WXTMD46, (GAA)230/(GAA)20; WXTMD43, (GAA)230/(GAA)230; WXTMD40, (TTC)20/(TTC)20; WXTMD45, (TTC)20/(TTC)230; WXTMD41, (TTC)230/(TTC)230. In all strains, the (GAA)n•(TTC)n tracts were inserted at the same position on chromosome V near the URA3 gene (green rectangles in Figure 4). These diploid strains were constructed from two haploid parents (PSL2 and PSL5) with numerous sequence polymorphisms allowing mapping of the positions of the crossovers as described further below.
The rates of RCOs with 95% confidence limits, based on an average of the number of sectored colonies in at least 20 cultures, are shown in Table 1. Strains homozygous for (GAA)20 or (TTC)20 tracts (WXTMD42 and WXTMD40) had rates of RCOs of about 4×10−6/division. These rates are very similar to that observed in the isogenic PSL101 strain (6×10−6) that had no GAA•TTC tracts [15]. The strains homozygous for either the (GAA)230 or (TTC)230 tracts (WXTMD43 and WXTMD41, respectively) had RCO rates of about 2×10−4/division. Thus, the addition of a GAA•TTC tract that is only 690 base pairs in length elevated the rate of RCOs in a 120 kb interval by more than 30-fold. The strains heterozygous for the long tracts (WXTMD46 and WXTMD45) also had substantially (20-fold) elevated rates of RCOs; the rates of RCOs in the heterozygous strain were about half those observed in the homozygous strains, indicating the GAA•TTC sequences on the two homologues functioned independently. As found previously for the MD506 and MD508 strains, the orientation of the GAA•TTC tract has no strong effect on its recombinogenic properties. It should be noted that Break-Induced Replication (BIR) [12] and local gene conversion events can generate unsectored canavanine-resistant colonies; however, these colonies cannot be unambiguously distinguished from RCOs that occur prior to plating cells on canavanine-containing medium [16].
From the results described above, one obvious possibility is that (GAA)230•(TTC)230 tracts are preferred sites for formation of a DSB or some other type of recombinogenic DNA lesion. By this model, one would expect most of the tract-stimulated recombination events to map at or near the position of the tract. In addition, in meiotic and mitotic recombination events in yeast analyzed previously, if a diploid is heterozygous for a preferred site of DSB formation, the chromosome with the preferred site is the recipient of genetic information in a gene conversion event [12]. We examined the positions of crossovers and associated gene conversion events in two strains: WXTMD46 (a diploid heterozygous for a (GAA)230 tract) and WXTMD42 (a diploid homozygous for [GAA]20 tracts).
The positions of the crossovers and gene conversion events were mapped by the methods described previously [15]. In brief, using PCR and restriction analysis, for both sectors of a CanR red/white sectored colony, we determined whether polymorphic sites on chromosome V were homozygous for the PSL2 form of the polymorphism (shown in red in Figure 4), the PSL5 form of the polymorphism (shown in black in Figure 4), or were heterozygous. In the previous studies of spontaneous or gamma-ray-induced mitotic crossovers, four types of sectored colonies were commonly observed: 1) RCOs unassociated with an adjacent gene conversion tract, 2) RCOs associated with an adjacent 3∶1 tract (as defined in Figure 2), 3) RCOs associated with an adjacent 4∶0 tract, and 4) RCOs associated with a hybrid 3∶1/4∶0 tract. Spontaneous recombination events are distributed throughout the 120 kb interval with a minor “hotspot” located near the can1-100/SUP4-o marker and a minor “coldspot” near CEN5 [15].
A summary of the mapping of crossovers and associated conversions in WXTMD46 is shown in Figure 4A. All markers proximal to the crossover are heterozygous in both red and white sectors, and homozygous distal to the crossover in both sectors (as illustrated in Figure 2). As observed for spontaneous events previously, most of the crossovers (29 of 33) were associated with conversion tracts of various sizes. 3∶1 and 4∶0 conversion tracts (as defined in the Introduction) are indicated by thin and thick vertical lines in Figure 4, respectively. 3∶1/4∶0 hybrid tracts are shown by adjacent thick and thin lines. Conversion tracts shown in black indicate that genetic information was transferred from the PSL5-related homologue and red tracts show transfer of information from the PSL2-related homologue.
Almost all of the conversion events in WXTMD46 included one or both of the markers flanking the GAA•TTC tract, as expected if the recombination event initiated within the tract. All four of the crossovers unassociated with conversion (shown as green Xs) occurred in the region containing the tract. In Figure 5, we compare the distribution of conversion events in WXTMD46 and PSL101 (an isogenic diploid without a GAA•TTC tract; data from Lee et al. [15]). The difference in the distributions of conversion events in the two strains is evident. In addition, in WXTMD46, the conversion tracts were strongly biased in the direction that represents transfer of information from the PSL5-related homologue. This result is consistent with the recombinogenic lesion occurring on the PSL2-related homologue that contains the (GAA)230 tract rather than the chromosome with the (GAA)20 tract.
Several other features of the conversion events are important. First, most of the conversion tracts were either 4∶0 tracts or hybrid 3∶1/4∶0 tracts. As discussed previously, such tracts are most simply interpreted as representing repair in G2 of a DSB formed in G1 (Figure 2B and 2C). This issue will be discussed in more detail below. Second, although some of the observed conversion events extended symmetrically to both sides of the tract, others were asymmetric. Thus, conversion events can extend either unidirectionally or bidirectionally from the initiating DNA lesion. Third, as observed with spontaneous recombination events and events induced by gamma rays in G1 [15], [17], the conversion tracts were long compared to those observed in meiosis. We estimated tract length by averaging the minimal tract length (the distance between the markers included in the tract) and the maximal tract length (the distance between the closest flanking markers not included in the tract). The median length of the tracts was 20.3 kb (95% confidence limits of 12.5–23.4 kb), somewhat larger than the length observed in spontaneous events without the (GAA)n•(TTC)n tracts (6.5 kb; [15]). The median size of meiotic conversion tract lengths is about 2 kb [14]. Fourth, as in previous studies, we found a number of examples of crossovers within a conversion tract; these events are indicated by asterisks in Figure 4. As discussed in Lee et al. [15], most of these events are explicable as representing the independent repair of two broken chromatids. An example of this class of conversion event is shown in Figure S1.
We also mapped a small number of RCOs in WXTMD42, the strain homozygous for the (GAA)20 tracts (Figure 4B). As expected, these events were distributed throughout the CEN5 to can1-100/SUP4-o interval. In addition, the conversion events involved transfer of information from both homologues with approximately the same frequency. The median conversion tract length in WXTMD42 is 11.6 kb (95% confidence limits of 3.7–22.3 kb).
The genetic evidence predicts the existence of a tract-associated DSB in G1 diploid cells. To look for such DSBs directly, we prepared DNA samples from stationary phase cells (>95% unbudded cells) of two isogenic haploid strains, WXT10 with a (TTC)20 tract and WXT11 with a (TTC)230 tract. Intact chromosomal DNA was isolated from cells suspended in agarose plugs to prevent shearing and the resulting samples were analyzed by contour-clamped homogeneous electric field gel electrophoresis (CHEF gels; [25]). The separated chromosomal DNA molecules were transferred to nylon membranes and hybridized to URA3-specific probe. We observed a chromosomal fragment at the position expected for a DSB within the tract (Figure 6A) in WXT11, but not in WXT10 (Figure 6B). The fraction of broken chromosomal molecules observed in three independent experiments was about 0.013 (average of 0.013, 0.017, and 0.01). Although this frequency of DSBs is considerably higher than the observed frequency of RCOs (about 10−4), it is likely that many of the DSBs are repaired by pathways, such as BIR and gene conversion unassociated with RCOs, that do not generate RCOs [12].
In yeast, long (CTG)n•(CAG)n tracts are preferred sites for DSB formation in mitosis [26]. In meiosis, long (greater than 75 repeats) (CTG)n•(CAG)n tracts were hotspots of recombination in one study [27], but were not in another [28]. Short (10-repeat) (CTG)n•(CAG)n tracts were not meiotic recombination hotspots [29].
As shown above long (GAA)n•(TTC)n promote DSB formation in mitosis. It was reasonable to ask, therefore, whether long GAA•TTC tracts stimulate meiotic recombination, as well. To address this question, we performed tetrad analysis, measuring meiotic recombination distances in three intervals on chromosome V: CEN5-ura3; ura3-can1-100/SUP4-o (the interval containing the tracts), and can1-100/SUP4-o to V9229::HYG. The heterozygous HYG gene (encoding a protein that results in resistance to hygromycin) was inserted approximately 20 kb centromere distal to the can1-100 gene. This analysis was done in WXTMD46 (which contains (GAA)230 on one homologue and (GAA)20 on the other) and PSL101 (which lacks (GAA)n•(TTC)n tract insertions). No significant differences were observed in map distances for any of the intervals (details of the analysis in Table S4). The map distance for the interval containing the insertion was 36 cM in WXTMD46 and 37 cM in PSL101 (total of about 100 tetrads examined in each strain).
Strong meiotic recombination hotspots are associated with high rates of gene conversion and crossovers [30]. The (GAA)230 tract in WXTMD46 is located about 1 kb from the mutant ura3 allele and the (GAA)20 tract is located the same distance from the wild-type URA3 allele. If the (GAA)230 tract is a preferred site for meiotic DSB formation, we would expect an elevation in gene conversion events of the 3 Ura+:1 Ura− class, since the chromosome that receives the DSB acts as a recipient for information derived from the uncut chromosome [12]. This effect should be detectable since the strong meiotic recombination HIS4 hotspot stimulates meiotic conversion events at sites located 2.7 kb from the hotspot [31], a distance longer than that between the (GAA)230 tract and URA3. In PSL101, we observed two conversion events, both 1 Ura+: 3 Ura− tetrads, in a total of 118 tetrads. In 105 tetrads derived from WXTMD46, we found no gene conversions of the 3+:1− or 1+:3− classes, but four tetrads that had 4 Ura+: 0 Ura− spores. This 4∶0 type of conversion is consistent with a mitotic gene conversion occurring within a sub-population of the WXTMD46 cells prior to sporulation [11]. Consistent with this hypothesis, in two of the tetrads with 4 Ura+ spores, all four spores had the SUP4-o marker and were HygS. These segregation patterns are consistent with a mitotic gene conversion at the ura3 locus associated with a mitotic crossover.
We also examined the meiotic stability of the (GAA)n•(TTC)n tracts by PCR analysis of spore DNA in 20 tetrads. Three patterns were observed. In 10 tetrads, two of the tracts were about 20 repeats in length and two were about 230 repeats in length. In 5 tetrads, two of the tracts were 20 repeats in length and two were of equal size but shorter than 230 repeats; this class is consistent with a sub-population of WXTMD46 cells in which the 230-repeat tract had undergone a mitotic deletion. In the third class (5 tetrads), two spores had 20-repeat tracts, one had a 230-repeat tract, and one had a tract of intermediate size; this class is consistent with a meiotic deletion event in one of the two 230-repeat tracts. Taken together with the mapping and gene conversion data, these results argue that the long (GAA)n•(TTC)n tracts are somewhat meiotically unstable, but the DSBs formed within the tract do not strongly stimulate meiotic recombination between the homologous chromosomes. This issue will be discussed further below.
The main conclusions from our study are: 1) (GAA)230•(TTC)230 tracts in both orientations strongly stimulate recombination between homologous chromosomes in mitosis, but not in meiosis, 2) the recombinogenic properties of the (GAA)n•(TTC)n tracts suggest that most of the events are initiated by a DSB formed in G1 of the cell cycle (a conclusion supported by a physical analysis of tract-associated DSBs), 3) the gene conversion events associated with the (GAA)n•(TTC)n repeats resemble those associated with spontaneous mitotic crossovers, and 4) single conversion events can be propagated from the location of the (GAA)n•(TTC)n insertion either unidirectionally (in either direction) or bidirectionally. Each of these conclusions will be discussed in detail below.
Although the tendency of certain trinucleotide tracts to expand in size was first demonstrated in humans, much of the experimental research concerning the effects of genome-destabilizing effects of these sequences has been done in bacteria and the yeast Saccharomyces cerevisiae [1], [2], [32]. In yeast, three types of repetitive trinucleotide tracts, (CTG)n•(CAG)n, (CGG)n•(CCG)n, and (GAA)n•(TTC)n, have been examined in detail. All three types of tracts undergo frequent size alterations with the frequencies of alterations increasing as a function of the number of repeats [32]. The frequency of these alterations is also affected by the orientation of the repetitive tract with respect to the replication origin. All three tracts are capable of forming secondary structures in vitro with one strand forming a more stable secondary structure than the other [1]. The orientation in which the strand with the most stable secondary structure is on the lagging strand for replication has the highest frequency of tract alterations. This orientation is also associated with replication fork pausing [1]. For the (GAA)n•(TTC)n repeats, as discussed above, replication fork pausing is observed when the (GAA)n repeats are on the lagging strand [4]. Somewhat unexpectedly, large-scale expansions of (GAA)n•(TTC)n tracts occur with approximately the same frequency regardless of the orientation of the tract [6].
Since DSBs are recombinogenic [12] and since DSBs are observed at the sites of long (CTG)n•(CAG)n and (GAA)n•(TTC)n tracts [5], [26], one would expect that such tracts would be hotspots for recombination. Long (CTG)n•(CAG)n tracts stimulate intrachromosomal recombination between repeats and sister-chromatid exchanges [26], [33]; long (GAA)n•(TTC)n tracts elevate the frequency of recombination between repeats on non-homologous chromosomes in yeast [5] and plasmid-plasmid recombination in E. coli [7]. In these assays, it was unclear whether the recombination events were reciprocal (producing two recombined DNA molecules) or non-reciprocal. The assay used in our current study selects for reciprocal events.
We found that the 230-repeat tract elevates the rate of RCOs in 120 kb interval from about 5×10−6/division (strains with no tract or a 20-repeat tract) to about 2×10−4/division. We calculate that the rate of RCOs/kb in the strains without the tract is about 4×10−8/kb/division. The 690 bp tract has a rate of RCOs of about 3×10−4/kb/division. Consequently, the (GAA)230•(TTC)230 tract is about 104-fold more recombinogenic than an average yeast sequence.
In contrast to the strong recombinogenic effects of the tract on mitotic recombination, no strong stimulation was observed for meiotic exchange. Since we observed meiosis-specific alterations in tract length in about 25% of the tetrads that were analyzed, it is likely that the long (GAA)n•(TTC)n tracts are substrates for DSB formation in meiosis. The lack of a detectable effect of the tracts on meiotic recombination can be explained in two ways. First, it is possible that tract-associated DSBs are repaired by intrachromosomal interactions (Synthesis-Dependent Strand Annealing, SDSA) or sister-chromatid exchanges [12]; neither of these events would be detected by standard tetrad analysis. Meiosis-specific intra-allelic changes in the lengths of minisatellites consistent with SDSA events have been observed previously in humans [34] and yeast [35], [36]. Second, it is possible that the effects of a weak tract-associated hotspot would be obscured by the very high frequency of meiosis-specific DSBs catalyzed by Spo11p. We note, however, that a strong tract-associated hotspot would have been detected by our analysis. The strong HIS4 recombination hotspot, for example, increases the map length in the LEU2-HIS4 interval from 20 cM to 36 cM [37].
Previously, we showed that about 40% of spontaneous RCOs were associated with 4∶0 or 3∶1/4∶0 hybrid conversion tracts [15]. We suggested that such events were a consequence of DSB formation on an unreplicated chromosome, followed by replication of the broken chromosome, and repair of the two resulting broken chromatids (Figure 2B and 2C). Since most of the RCOs stimulated by the (GAA)n•(TTC)n repeats are associated with 3∶1/4∶0 tracts, it is likely that the recombinogenic DSBs are formed in G1. This conclusion, based on genetic analysis, is also supported by the physical analysis demonstrating tract-associated DSBs in stationary phase cells (Figure 6). Since the DSBs occur in G1/G0, the observation that the tract-associated stimulation of RCOs is independent of the orientation of the tract is expected.
It should be emphasized that our results do not show that tract-associated DSBs occur only in G1/G0. We observed previously that (GAA)n•(TTC)n tracts stimulate ectopic recombination between repeats on non-homologous chromosomes in an orientation-dependent mechanism [5]. We suggest that these events are likely to be non-reciprocal and, therefore, regulated differently than the RCOs that are the subject of the present study.
In summary, our studies of the properties of (GAA)n•(TTC)n tracts indicate that they promote genetic instability by several different mechanisms. One mechanism is dependent on the orientation of the repeats and is likely to reflect breakage of replication forks [5]; this mechanism is also associated with small tract contractions/expansion and ectopic recombination events [4], [5]. A second mechanism is the orientation-independent large expansion of (GAA)n•(TTC)n tracts that may involve strand-switching events in which the leading strand copies an Okazaki fragment [6]. The third mechanism is also independent of the orientation and likely reflects DSB formation in G1 to yield RCOs.
Although we have not determined the source of the G1-induced DSBs, they may reflect the action of DNA repair enzymes and/or topoisomerases interacting with secondary structures formed by the tracts. Replication-independent instability has been observed in mammalian cells for both (GAA)n•(TTC)n and (CTG)n•(CAG)n tracts [8], . This instability appears to be related to DNA repair events associated with transcription [9], [10]. In most of the strains examined in our study, the (GAA)n•(TTC)n tracts were embedded in a promoter-less fragment of the LYS2 gene.
The most obvious difference in the patterns of spontaneous RCOs observed previously [15] and those seen in the current study is the location of the events. All of the events observed in the current study are at or near the site of the (GAA)n•(TTC)n tracts, presumably because all events are initiated at or near the tracts. Although the distribution of spontaneous events observed by Lee et al. is not completely random, it is clear that the events can be initiated at many sites within the 120 kb interval (Figure 5). Although the properties of DNA sequences that regulate the probability of initiating a mitotic recombination events have not yet been completely established, mitotic recombination is promoted by closely-spaced inverted repeats [38] and by high rates of transcription [39], [40].
The median length of the tract-stimulated conversion events in WXTMD46 (20.3 kb) is longer than those observed for spontaneous events in the absence of the repetitive sequence (6.5 kb) and conversion events generated in G1-arrested cells by gamma radiation (7.3 kb; [17]). The median tract length is much longer than the median length observed associated with RCOs induced by gamma radiation in G2-arrested cells (2.7 kb; [17]).
Most of the conversion tracts are 3∶1/4∶0 hybrid tracts (Figure 4A). As discussed in the Introduction, such tracts can be explained by independent repair of two DSBs. If the DSBs occur within the GAA•TTC insertions, we expect that the 4∶0 region of the hybrid tract should include one or both of the markers flanking the tract, and this expectation is met (Figure 4A). If processing of the broken DNA ends is bidirectional and symmetric from the site of the DSB, most tracts should have a 4∶0 region flanked by 3∶1 regions. Although we observe this pattern for some of the conversion events, for other events, the 4∶0 region is at one end of the hybrid tract. Thus, we infer that the mechanism that generates the gene conversion in mitosis can be asymmetric. In addition, single conversion events can be propagated from the initiation site either toward the centromere or toward the telomere. Meiotic gene conversion tracts share these properties [41], [42].
Two different mechanisms can result in a gene conversion event. During meiotic recombination, most conversion events reflect heteroduplex formation followed by repair of any resulting mismatches. One key early intermediate in this process is a broken end that has been “processed” by 5′ to 3′ degradation on one of the two strands [13]. It is possible that mitotic conversion events involve much more extensive processing than meiotic events or extensive branch migration of the Holliday junction(s) associated with the strand invasion. An alternative possibility is that the conversion events involve the repair of a double-stranded gap [43]. Although there is strong evidence that mitotic events that generate relatively short conversion tracts are a consequence of heteroduplex formation followed by mismatch repair [44], [45], it is currently unclear whether the very long tracts are a consequence of mismatch repair or gap repair [15].
In summary, we have demonstrated that (GAA)230•(TTC)230 tracts strongly stimulate RCOs and our analysis indicates that these events are initiated by a DSB in unreplicated DNA. These results have several implications relevant to the genetic instability observed in patients with Friedreich's ataxia. First, a G1-associated DSB may be an intermediate in the expansion process in at least a sub-set of the expansion events. Second, since we find that the (GAA)230•(TTC)230 tracts are highly recombinogenic by a mechanism that is independent of DNA replication, our findings may be relevant to the observation that the FRDA-associated tracts are unstable in post-mitotic (non-dividing) cells and these expansions contribute to pathogenesis. For example, the highest rate of somatic instability is observed in dorsal root ganglia, which is the most damaged tissue in FRDA patients [46]. In addition, expanded (GAA)n•(TTC)n tracts may elevate the frequency of loss of heterozygosity (LOH) on the chromosome containing the expanded tract, allowing heterozygous mutations to become homozygous. Since there are other (GAA)n•(TTC)n runs within mammalian genomes that are prone to expansions [47], such tracts may also promote LOH on other chromosomes. It would be of interest to examine tissues of FRDA patients or cell lines derived from patients for tract-associated DSBs (using ligation-mediated PCR) or LOH of single-nucleotide polymorphisms located centromere-distal to the expanded tracts.
Most of the experiments involve diploids generated by crosses of haploids with diverged DNA sequences. The haploid strain PSL5 [15] is derived from the YJM789 genetic background whereas PSL2 [15] is derived from W303a [39]. The details of the constructions and genotypes of the haploid and diploid strains are given in Text S1 and Tables S1, S2, S3. The diploids strains used to measure the effect of GAA•TTC tracts on RCOs were homozygous for the ade2-1 mutation, and heterozygous on chromosome V for can1-100 and an allelically-placed copy of SUP4-o. As described in the text, this system allows the selection of RCOs as CanR red/white sectored colonies.
Standard yeast procedures were used for transformations, mating, sporulation, and tetrad dissection [48]. Media were prepared as described previously [15], [16]. The two-dimensional gel analysis of replication forks was done as described previously [5]. DNA samples for the gel analysis were treated with the AflII restriction enzyme, and the Southern blot was hybridized to the 3.9 kb LYS2-specific AflII fragment isolated from pFL39LYS2 (described in Text S1). To analyze tract-associated DSBs, we grew haploid strains to stationary phase (three days of growth in rich growth medium [YPD] at 30°C), and then prepared DNA by methods described previously [25]; in the stationary-phase cultures, >95% of the cells were unbudded as expected for cells in G1/G0. Chromosomal DNA molecules were separated using the Bio-Rad CHEF Mapper XA. The Southern analysis was done using a URA3-specific probe that was prepared by PCR amplification of genomic DNA with the primers: URA3-f (5′ GGTTCTGGCGAGGTATTGGATAGTTCC) and URA3-r (5′ GCCCAGTATTCTTAACCCAACTGCAC). The hybridization signals were detected and quantitated using a PhosphorImager.
The methods used to quantitate RCOs in various strains were identical to those described previously [15]. In brief, individual colonies formed on rich growth medium were suspended in water, and plated on non-selective medium (omission medium lacking arginine [SD-arg]) or on medium containing canavanine (SD-arg with 120 micrograms/ml canavanine). Plates were incubated at room temperature for four days, followed by storage for one day at 4°C (which accentuates the red color of sectors). The rate of RCOs for each strain was determined by averaging the frequency of crossovers observed in at least 20 independent cultures (colonies).
Red and white CanR strains were purified from each half of the sectored colonies. DNA was isolated by standard procedures [48]. As we have done previously, we mapped crossovers by examining 34 single-nucleotide polymorphisms (SNPs) located in the 120 kb interval between CEN5 and the can1-100/SUP4-o markers. For each SNP, the DNA from one of the haploid parents contained a diagnostic restriction enzyme recognition that was altered for the other parent. For each SNP, we amplified genomic DNA using primers flanking the heterozygous marker, treated the fragment with the diagnostic restriction enzyme, and examined the products by gel electrophoresis. From this analysis, we could determine whether the sectored colony was homozygous for the YJM789 form of the SNP, homozygous for the W303a form of the SNP, or heterozygous for the polymorphism. The sequence of the primers and restriction enzymes used in the analysis are given in Lee et al. [15].
Statistical analyses were done using the VassarStats Website (http://faculty.vassar.edu/lowry/VassarStats.html). Most of the comparisons involved the Fisher exact test. 95% confidence limits on the rates of RCOs were calculated by determining the 95% confidence limits on the proportions (number of sectored colonies/number of colonies on non-selective plates) using the Wilson procedure with a correction for continuity. Calculations of median conversion tract lengths and 95% confidence limits on the median were done as described previously [17].
|
10.1371/journal.pntd.0004830 | Cytokine Profile of Children Hospitalized with Virologically-Confirmed Dengue during Two Phase III Vaccine Efficacy Trials | Two large-scale efficacy studies with the recombinant yellow fever-17D–dengue virus, live-attenuated, tetravalent dengue vaccine (CYD-TDV) candidate undertaken in Asia (NCT01373281) and Latin America (NCT01374516) demonstrated significant protection against dengue disease during two years’ active surveillance (active phase). Long-term follow up of participants for breakthrough disease leading to hospitalization is currently ongoing (hospital phase).
We assessed the cytokine profile in acute sera from selected participants hospitalized (including during the active phase) up to the beginning of the second year of long-term follow up for both studies. The serum concentrations of 38 cytokines were measured in duplicate using the Milliplex Human Cytokine MAGNETIC BEAD Premixed 38 Plex commercial kit (Millipore, Billerica, MA, USA). Partial least squares discriminant analyses did not reveal any difference in the overall cytokine profile of CYD-TDV and placebo recipients hospitalized for breakthrough dengue regardless of stratification used. In addition, there was no difference in the cytokine profile for breakthrough dengue among those aged <9 years versus those aged ≥ 9 years.
These exploratory findings show that CYD-TDV does not induce a particular immune profile versus placebo, corroborating the clinical profile observed.
| A live-attenuated, tetravalent dengue vaccine (CYD-TDV) has been shown to provide protection against dengue disease in two large-scale, placebo-controlled, phase III efficacy studies. Continued surveillance of study participants was subsequently undertaken to better define longer term vaccine efficacy and safety. A yet unexplained higher incidence of hospitalization for dengue disease was observed among children aged <9 years in year 3 of follow up. While the clinical outcome of the hospitalized cases was similar between CYD-TDV and placebo recipients, it was important to further investigate whether the immune profile induced by breakthrough infection differed between the two study groups. We compared the profile of 38 cytokines, chemokines and growth factors in acute phase sera collected from participants with breakthrough disease in the two groups. No difference in overall profile was observed between CYD-TDV and placebo recipients. Similarly, no difference in the cytokine profile for breakthrough dengue was observed between those aged <9 years and those aged ≥ 9 years. Based on these analyzed factors, our study shows that CYD-TDV does not induce an overall altered immunological profile with breakthrough disease compared with placebo, in agreement with the similar clinical pictures and viremia observed in the two groups.
| Dengue virus (DENV) is the most important mosquito-borne pathogen threatening approximately half of the world’s population, mostly in tropical and subtropical areas including Latin America and Southeast Asia [1,2]. Infection with any of the four DENV serotypes can be asymptomatic or cause a spectrum of clinical symptoms from mild fever (dengue fever/DF) to more severe, potentially life-threatening disease including dengue hemorrhagic fever and shock syndrome (DHF/DSS) [3]. Severe dengue is most often observed in previously infected subjects undergoing secondary infection with a different DENV serotype [4]. DHF/DSS is associated with excessive immune activation, or a ‘cytokine storm’, which may contribute to increased vascular permeability with extensive plasma leakage and resultant signs of shock [5,6]. Differences in the cytokine profile in severe versus non-severe disease have been demonstrated [7–10].
Innate and adaptive immune responses have been proposed to contribute to the cytokine profile and disease outcome [11]. Firstly, in the humoral response to DENV infection, antibody-dependent enhancement of infection (ADE) by preexisting “enhancing” antibodies may play a role in the development of severe disease with second or subsequent infection with heterotypic dengue serotypes [12]. “Intrinsic ADE” could also occur through production of the immunosuppressive IL-10 and proinflammatory mediators, including IL-6 and TNF-α, responsible in part for increased vascular permeability [13,14]. Secondly, in the cell-mediated immune response, several studies suggest involvement of T-cells in the ‘original antigenic sin’ phenomenon in severe dengue. According to this hypothesis, inappropriate low avidity cross-reactive T-cells induced by a secondary heterotypic infection may produce an altered cytokine profile, including increased production of inflammatory cytokines, such as TNF-α and CCL4/MIP1β, and decreased production of IFN-γ and IL-2 [11].
There are currently no vaccines or antiviral therapies available for the management of dengue disease. Several vaccine formulations are in development [15], and one has recently obtained marketing authorization in several countries, Dengvaxia, the recombinant yellow fever-17D–dengue virus, live-attenuated, tetravalent dengue vaccine (CYD-TDV). Two large-scale phase III efficacy studies have recently been conducted with CYD-TDV in Asia and Latin America, demonstrating significant protection against dengue disease [16,17]. Continued surveillance of participants from these two studies for breakthrough disease leading to hospitalization is ongoing to better define long-term vaccine efficacy and safety. After 3 years, the cumulative risk of hospitalization among children aged 2–16 years was still lower in the CYD-TDV group than the control group; however, an unexplained higher incidence of hospitalization for dengue disease was observed among children aged <9 years in year 3 [18]. While the clinical outcome of these hospitalized/severe cases was similar between CYD-TDV and placebo recipients, with all cases fully recovered following supportive medical care, it was important to investigate further if the immune profile induced after infection could differ between the two treatment groups. We assessed 38 cytokines in sera collected during the acute phase of breakthrough disease among CYD-TDV and placebo recipients from the two phase III trials.
Details of the clinical protocols (ClinicalTrials.gov numbers: NCT01374516 and NCT01373281) and study population are described elsewhere [16–18]. Briefly, eligible participants aged 2–14 years in the Latin American study (CYD15) and 9–16 years in the Asian study (CYD14) were randomized (2:1) to receive CYD-TDV or placebo (0.9% NaCl). After the initial active surveillance phase (active phase) for breakthrough dengue (up to 25 months after the first injection), a four-year long-term follow-up phase was initiated to monitor breakthrough disease leading to hospitalization (hospital phase). Participants attended yearly clinic visits, with regular contact with study personnel between visits (at least every 3 months). Hospitalization for acute fever was recorded during study contacts and visits, and by self-reporting and surveillance of non-study hospitals. As a post-hoc exploratory analysis, we tested blood samples from selected participants at one month post-dose 3 for baseline cytokine assessment, and during the acute phase of febrile illness from participants who required hospitalization for virological confirmation of dengue infection in the active phase, and during the first and beginning of the second year of long-term follow up. The cut-off for the current analysis was 28 November 2014. During the acute phase, apart from a few cases, samples were obtained between day 1 (D1) and D7 after fever onset, and no difference was observed in the daily distribution of samples between the CYD-TDV and placebo groups in this interval of time (p = 0.99, Kuiper test).
The trials were undertaken in compliance with good clinical practice guidelines and the principles of the Declaration of Helsinki. Ethics review committees approved the protocol and subsequent amendments as well as the consent and assent forms. Parents or legal guardians provided written informed consent for all children, and older children also signed informed-assent forms, in compliance with the regulations of each country. For the exploratory study reported herein, the laboratory was blinded to treatment group during sample assessments and unblinded in order to perform the statistical analysis.
We evaluated 38 cytokines: sCD40L (sCD154), EGF, eotaxin/CCL11, FGF-2/FGF-basic, Flt3 Ligand, Fractalkine/CX3CL1, G-CSF, GM-CSF, GRO (isoforms GROα (CXCL1), GROβ (CXCL2), and GROγ (CXCL3)), IFN-α2, IFN-γ, IL-1α, IL-1β, IL-1Ra, IL-2, IL-3, IL-4, IL-5, IL-6, IL-7, IL-8/CXCL8, IL-9, IL-10, IL-12 (p40), IL-12 (p70), IL-13, IL-15, IL-17A/CTLA8, IP-10/CXCL10, MCP-1/CCL2, MCP-3/CCL7, MDC/CCL22, MIP-1α/CCL3, MIP-1β/CCL4, TGFα, TNF-α, TNF-β, and VEGF-A. Serum samples were tested undiluted and in duplicate (baseline samples were tested once). Serum cytokine levels were analyzed using the Milliplex Human Cytokine MAGNETIC BEAD Premixed 38 Plex commercial kit according to the manufacturer’s instructions. The fluorescent signals were detected using the multiplex array reader Bio-Plex 200 System. Raw data were initially measured as relative fluorescence intensity and then converted to cytokine concentrations in pg/mL based on the standard curve generated from the reference concentrations supplied with the kit. Means of the duplicates were reported.
Because the levels of two factors—CXCL10/IP-10 and sCD40L —were above the upper limit of the test in a significant number of samples, these factors were measured again separately using dedicated kits for 199 out of 207 samples (for the remaining 8 samples there was not enough volume left to re-run them: 4 in the CYD-TDV (sampled at day 3 (D3), D4, D5 and D7) and 4 in the placebo group (all sampled at D1). Samples were used at 1/10 dilution, except when those values were still above the upper limit, in which case 1/100 dilution was used. The Milliplex Human Cytokine MAGNETIC BEAD 2 Plex custom kit was used according to the manufacturer’s instructions. Cytokine concentrations were measured by the same analytical method as that described above.
For each cytokine, the mean concentration in pg/mL was log10 transformed. All comparisons were performed using a Student test if the normality hypothesis of both groups was accepted (Shapiro Wilks test or Kolmogorov test, depending on the sample size). If the normality hypothesis of one or both groups was not accepted, the Wilcoxon non-parametric test was used. In accordance with data distribution, geometric means were considered when comparisons were performed using a Student test, and medians were considered with the Wilcoxon non-parametric test.
The comparisons between the CYD-TDV and placebo groups were stratified as follows: i) all hospitalized cases, irrespective of phase or severity; ii) severe or non-severe cases only; iii) active phase or hospital phase only; iv) cases aged <9 years or aged ≥9 years. Additionally, the profile of overall cases (CYD-TDV and placebo groups combined) was compared between severe and non-severe cases. For cytokines with a significant difference between the CYD-TDV and placebo groups, an analysis of variance was conducted to evaluate whether the difference was study- or group-dependent.
The distribution of samples obtained between day 1 and day 7 after fever onset were compared between CYD-TDV and placebo groups using a Kuiper test.
Partial least squares discriminant analyses (PLS-DA) were performed for cytokines that appeared to be differentially expressed between baseline and acute phase of dengue illness to analyze the overall immune profile between: i) the CYD-TDV and placebo groups, ii) CYD14 and CYD15 studies, and iii) the two age groups (<9 years or ≥9 years). The goal was to provide a dimension reduction to relate a binary response variable (the 3 situations described above) to a set of predictor variables (the cytokines). The PLS-DA results give percentage discrimination between the two tested groups, ranging from 0% (no discrimination) to 100% (perfect discrimination).
All comparisons were performed using SAS v9.2 software. The PLS-DA was done with SIMCA-P 9. Margins of error were 5% for effects of the main factors and 1% for normality tests.
Of the 265 breakthrough cases occurring up to 28 November 2014, 207 serum samples (with sufficient volume for analysis) were available for immune profile assessment (S1 Table). Of these, 24 and 74 in the CYD-TDV group were from cases classified as severe and non-severe, respectively, as defined by the independent data monitoring committee, and 28 and 80, respectively, in the placebo group. Information regarding severity of illness was missing for one sample. An additional 40 samples were taken one month post-dose 3 (17 and 23 in the CYD-TDV and placebo groups, respectively) to assess baseline cytokine levels.
Of the 40 baseline samples, 33 could be paired with acute-phase samples. At baseline, only 7 cytokines had median levels above the lower limit of quantification (LLOQ): IP-10, MCP-1, MDC, MIP-1β, EGF, GRO and IL-8. In the acute phase, an additional 8 cytokines had median levels above the LLOQ: IFN-γ, TNF-α, FGF-2, IL-1α, IL-1Ra, IL-10, eotaxin, and sCD40L. A paired comparison of these 15 cytokines (irrespective of the treatment group) showed that 12 (eotaxin, FGF-2, IFN-γ, IL-10, IL-1Ra, IL-1α, IL-8, IP-10, MCP-1, MDC, TNF-α and sCD40L) were differentially expressed between baseline and acute phase of dengue illness (Fig 1).
A two dimension PLS-DA of all acute samples (n = 207) was performed on 12 cytokines that were differentially expressed between baseline and the acute phase. The 12 cytokines were plotted to compare samples between the CYD-TDV and placebo groups (S1A Fig), and between the CYD14 and CYD15 studies (S1B Fig). The percentage discrimination between CYD-TDV and placebo groups was 4.5% and between the CYD14 and CYD15 studies, 23.6%. These results suggest that there is no difference in the global cytokine profile between the two study groups or between the two studies. The data from the two studies were thus pooled hereafter.
Median levels for the 15 cytokines (described earlier) in the CYD-TDV (n = 108) and placebo (n = 99) groups are summarized in Fig 2 for all hospitalized cases. Differences in IP-10 levels in acute samples could not be analyzed since the majority of samples from both CYD-TDV and placebo recipients were levels above the upper limit of the assay (>10,000 pg/mL). Therefore, samples were re-run for this chemokine in a second experiment; as one could not directly compare data obtained in these new analyses to the ones obtained in the first experiment, levels and comparisons based on the new quantifications are presented in S2 Table. In summary, no difference was observed between the two groups for hospitalized cases.
Of the remaining cytokines, only two (IL-1Ra and MCP-1) were significantly different between CYD-TDV and placebo groups—both were higher in the placebo group (median levels for IL-1Ra, 32.30 pg/mL vs. 63.11 pg/mL, respectively [p = 0.038]; geometric means for MCP-1, 735.47 pg/mL vs. 957.94 pg/mL [p = 0.021]; Fig 3A). No differences were observed for these two cytokines between the two groups at baseline (p-values of 0.984 and 0.366 for MCP-1 and IL-1Ra respectively). There was no significant effect of study on IL-1Ra (p = 0.419) or MCP-1 levels (p = 0.064). Thus the difference between the CYD-TDV and placebo groups for these two cytokines was observed equally in both studies. There were no differences in IL-1Ra or MCP-1 levels between the two study groups with respect to the day after fever onset. Both cytokines peaked on the first day after fever onset and declined to baseline levels in the subsequent days (S2A and S2B Fig). The kinetics of sCD40L showed that the levels were similar between day 1 and day 3, declined slightly but stayed at elevated levels until day 7 (S2C Fig).
For the severe cases, IFN-α2, IL-15, IL-1Ra, IP-10 and MCP-1 levels were all significantly lower in the CYD-TDV group (n = 24) than in the placebo group (n = 28). Most participants had IP-10 levels above the upper limit of quantification of the assay preventing a meaningful comparison of data between the two study groups. As stated above, a new analysis was done for this factor, resulting in the same conclusion: significantly lower levels of IP-10 were observed in CYD-TDV than in the placebo group for severe cases (median levels 21648 pg/mL vs 40288 pg/mL, p = 0.033; see S2 Table). Median levels of IL-1Ra for severe cases were 25.4 pg/mL vs. 216.4 pg/mL for CYD-TDV and placebo groups, respectively (p = 0.008), and geometric means of MCP-1 were 600.8 pg/mL vs. 1325.2 pg/mL (p = 0.002) (Fig 3B). For IL-15, these values were 0.9 pg/mL vs. 5.8 pg/mL (p = 0.004) in the two groups, respectively, and for IFN-α2 were 2.4 pg/mL vs. 20.9 pg/mL (p = 0.021). However, IL-15 and IFN-α2 were not significantly different when all cases irrespective of severity were compared; nevertheless, the levels of these cytokines were generally low or undetectable, regardless of group (S3 Fig).
There was no difference in IL-1Ra and MCP-1 levels between non-severe cases in the two study groups (Fig 3C): the median levels of IL-1Ra were 33.5 pg/mL vs. 43.9 pg/mL (p = 0.530) for CYD-TDV and placebo groups, respectively, and the geometric mean levels of MCP-1 were 786.7 pg/mL vs. 855.1 pg/mL (p = 0.513). Since no differences in cytokine profile between the CYD-TDV and placebo groups were found in non-severe cases, the differences found previously for all hospitalized cases (Fig 3A) were due to higher IL-1Ra and MCP-1 levels in the placebo group for severe cases.
PLS-DA could not discriminate between the overall immune profile of severe and non-severe cases (discrimination index, 10.3%). However, when compared individually, EGF, IL-1Ra, and sCD40L were significantly different between severe and non-severe cases, with IL-1Ra the only cytokine higher in severe than in non-severe cases (median levels, 147.8 pg/mL vs 41.2 pg/mL, respectively, p = 0.015). Soluble CD40L and EGF were significantly higher in non-severe cases than in severe cases (sCD40L: 5460.6 pg/mL vs 2643.0 pg/mL, respectively, p<0.001; EGF: 80.1 pg/mL vs 49.5 pg/mL, respectively, p = 0.025) (Fig 4). As for IP-10, some samples presented values above the upper limit of quantification for sCD40L. This factor was then measured again separately, and comparisons are presented in S3 Table. This new evaluation resulted in the same conclusion, i.e. that higher levels of sCD40L were observed in non-severe than in severe dengue cases (median levels 2058.9 pg/mL vs 1144.1 pg/mL, p<0.001). On the other hand, IP-10 new quantification showed higher levels in severe than non-severe cases (median levels 27765.8 pg/mL vs 17493.2 pg/mL, p<0.001) (S2 Table).
No overall significant differences were found between the CYD-TDV and placebo groups when breakthrough cases in the active and hospitalization phase were analyzed separately (for each cytokine p-value > 0.05) except in the active phase, where sCD40L was significantly higher in CYD-TDV than in placebo recipients (10000 pg/mL [upper range] vs 5311.9 pg/mL, respectively, p = 0.017). In the CYD-TDV group, sCD40L levels dropped significantly in the hospital phase to 2381 pg/mL (p = 0.001), which was also the case in the placebo group dropping to 3203.8 pg/mL although this was non-significant (p = 0.0538). No significant differences were observed for sCD40L between CYD-TDV and placebo in the hospital phase, or between active and hospital phases in the placebo group. The new quantification confirmed above differences with higher levels in vaccinees than in placebos (2437.1 pg/mL vs 1863.6 pg/mL, p = 0.003; S3 Table).
PLS-DA could not discriminate between the overall immune profile of children aged <9 years and those ≥9 years in the CYD-TDV and placebo groups, with discrimination indices of 23.8% and 12%, respectively (S4A and S4B Fig). Moreover, no discrimination between the two age groups was pointed out (percentage discrimination of 12.7%) regardless of treatment groups. The same was true for the comparison between CYD-TDV and placebo recipients when considering each age group (discrimination indices of 18.9% and 3.8% for <9 years and for ≥9 years respectively).
Each cytokine was also compared individually between the two age groups: Table 1 summarizes those cytokines with significant differences between children aged < 9 years and those aged ≥9 years in the CYD-TDV or placebo groups. In the placebo group, IL-10 levels were significantly lower in children aged <9 years than those aged ≥9 years (p = 0.025). Among children aged <9 years, significant differences in IL-10 (Fig 5A), MCP-1 (Fig 5B) and IFN-α2 levels were observed between treatment groups (Table 1), with higher IL-10 levels (p = 0.037), and lower MCP-1 (p = 0.026) and IFN-α2 (p = 0.006) levels, among CYD-TDV recipients. No significant differences between treatment groups were found among the individual cytokine comparisons for those aged ≥9 years.
As mentioned in above sections, these two factors were above the upper limit of quantification in the first experiment. New tests were then performed to quantify them more accurately, using different kits from the initial test, thus preventing direct comparisons with the former results. Therefore, the different group comparisons were re-assessed as presented in S2 and S3 Tables. As stated previously, these new analyses confirmed and expanded previous conclusions. For each of these factors, two comparisons showed significant differences: i) regarding IP-10, levels were higher in severe cases irrespective of treatment group, and were also higher in placebos vs vaccinees when specifically considering severe cases; ii) in contrast, regarding sCD40L, levels were higher in non-severe cases, and were also higher in vaccinees vs placebos when specifically considering the active phase.
Thirty-eight cytokines were measured in serum samples from CYD14 and CYD15 study participants during their acute phase of dengue illness. No differences in overall immune profiles were observed between CYD-TDV and placebo recipients regardless of the stratification used for the PLS-DA. This corroborates the absence of any difference in symptomatology, viremia and blood parameters previously reported between the CYD-TDV and placebo groups [16–18]. Of the 15 cytokines with median levels above the LLOQ, only two, IL-1Ra and MCP-1, significantly differed between the treatment groups. Both cytokines were higher in the placebo group. Notably, MCP-1 was also significantly higher in placebo group than in the CYD-TDV group among participants aged <9 years. Moreover, these two cytokines appeared significantly higher in severe cases than non-severe cases in the placebo group.
IL-1Ra is an anti-inflammatory cytokine that can attenuate IL-1-induced fever (by neutralizing IL-1β). It is therefore thought to exert antipyretic action and may be produced as a feedback mechanism to counteract the early increase in IL-1β levels in DF [19]. In a prospective study of children with DSS (n = 50), IL-1Ra and IL-6 plasma concentrations were significantly higher in non-survivors than in survivors [8]. Moreover, IL-1Ra plasma concentrations at day of admission were significantly associated with mortality and were elevated in severe cases.
MCP-1 is a potent chemotactic factor for monocytes and macrophages, both are major sources of MCP-1. This cytokine may increase vascular permeability thereby leading to plasma leakage in dengue patients [20,21]. Higher MCP-1 levels in dengue-infected patients compared to controls have been reported [19,20]. Since IL-1Ra and MCP-1 levels were significantly higher in placebo recipients than CYD-TDV recipients in our study, and both are soluble mediators associated with increased dengue severity, this is a positive finding regarding a potential risk linked to the vaccine.
sCD40L levels were higher during the active phase in the CYD-TDV group, and in non-severe cases irrespective of the treatment group (S3 Table). Soluble CD40L is mainly released by activated platelets and contributes to promote coagulation, therefore higher levels could be beneficial in this regard [22]. A related hypothesis is also put forward by another study proposing that the levels of sCD40L may directly and more significantly reflect thrombocytopenia than platelet counts; in agreement with our findings, the same authors also demonstrating lower levels of sCD40L in severe dengue [23]. sCD40L may then be associated to platelet counts and help counteract hemorrhage in severe dengue cases, in-line with the higher sCD40L levels in non-severe than severe cases in our study. Interestingly, in the case of Ebola hemorrhagic fever, non-fatal cases had higher levels of sCD40L [24], which may be in agreement of the present observation in the case of non-severe dengue. Still in favor of a positive vaccine impact, the opposite finding was seen for IP-10 with levels higher in severe cases and lower in vaccinees than in placebos when specifically considering severe cases. Our findings are also in agreement with another study demonstrating higher levels of IP-10 in severe cases [25].
The lower EGF levels observed in severe cases are also consistent with previous observations of significantly lower mean EGF levels in patients with DHF than DF [26].
Intrinsic ADE was shown in some culture-based studies to suppress IL-12, IFN-γ and TNF-α expression, and to increase the expression of the anti-inflammatory cytokines IL-6 and IL-10 [13,27]. This differential expression of pro- and anti- inflammatory cytokines may play a role in the pathogenesis of severe dengue [28]. In our study, median levels of IL-6 were below the LLOQ in both the CYD-TDV and placebo groups, and the median levels of IL-10 did not differ between the two groups overall (46.8 pg/mL and 58.7 pg/mL, respectively [p>0.05]). In contrast to the overall trend, although IL-10 was significantly higher in children aged <9 years in the CYD-TDV group than the placebo group, this was due to an unexpectedly low level in the placebo group whereas the value observed in the CYD-TDV group was within the range expected. There was no significant difference in IL-10 levels in vaccinated children aged <9 years compared with those ≥9 years or those aged ≥9 years who received placebo. However, the IL-10 levels reported in the current study were lower than those commonly reported in the literature during symptomatic/severe dengue [7,29,30]. The small difference in IL-10 levels in our study is unlikely to explain the imbalance in hospitalization for breakthrough dengue among children aged <9 years between the two groups. In this regard, since IL-10 levels have been shown to be higher in secondary dengue infections than in primary infections or in healthy controls [31], it can be hypothesized that, in children aged <9 years, infections occurring in the CYD-TDV group are akin to secondary infections where vaccination may be considered as the primary infection/challenge and breakthrough disease as secondary infection. Whereas, in the placebo group, the breakthrough disease observed may represent predominantly primary infections [32]. This is supported by the observation that IFN-α2, elevated in primary infections [33], was higher in children aged <9 years in the placebo group.
IL-6 has been shown by several groups to be elevated during dengue illness [7–10] and associated with severe forms of the disease. In our study, only 30% of the participants in the two study groups had levels above the LLOQ, albeit at low to moderate levels. No differences were seen between the two study groups stratified by age, phase of study or severity of disease. In agreement with the similar overall immune profile between CYD-TDV and placebo recipients, no differences were seen in clinical symptoms and viremia levels in hospitalized/severe cases between the two groups [18]. Overall, only two cytokines (IL-1Ra and MCP-1) were significantly higher in the placebo than the CYD-TDV group. Notably, CYD-TDV recipients had higher sCD40L and lower IL-1Ra levels, a profile linked to a milder disease in the present study.
Numerous studies have focused on the identification of cytokines specifically expressed during dengue illness, which could predict the development of severe forms of the disease [34,35]. A limitation of these studies, including our present work, is the fact that samples are often collected at a single time point and at different times of the clinical course of infection (between 1 and 15 days after fever onset). This heterogeneity in sampling time could lead to misleading results due the transient expression of circulating cytokines. Although our study was not designed to control for this variable, sampling times do not appear to be different between the CYD-TDV and placebo groups (sampling occurred at an average of 3.82 and 3.36 days after fever onset, respectively [S1 Table]). Finally, the cytokines that showed significantly elevated levels in the acute samples compared to baseline (IFN-γ, IL-8, TNFα, IL-1α, IL-1Ra, IL-10, FGF-2, eotaxin, MCP-1, MDC, sCD40L, and IP-10) were those expected based on previous findings [7–11,34].
In conclusion, our study confirms that CYD-TDV does not induce an overall altered cytokine profile with breakthrough disease compared with placebo. This is in agreement with the similarities in symptomatology, viremia and blood parameters reported in the two groups.
|
10.1371/journal.pbio.1001890 | A Novel Nodal Enhancer Dependent on Pluripotency Factors and Smad2/3 Signaling Conditions a Regulatory Switch During Epiblast Maturation | During early development, modulations in the expression of Nodal, a TGFβ family member, determine the specification of embryonic and extra-embryonic cell identities. Nodal has been extensively studied in the mouse, but aspects of its early expression remain unaccounted for. We identified a conserved hotspot for the binding of pluripotency factors at the Nodal locus and called this sequence “highly bound element” (HBE). Luciferase-based assays, the analysis of fluorescent HBE reporter transgenes, and a conditional mutation of HBE allowed us to establish that HBE behaves as an enhancer, is activated ahead of other Nodal enhancers in the epiblast, and is essential to Nodal expression in embryonic stem cells (ESCs) and in the mouse embryo. We also showed that HBE enhancer activity is critically dependent on its interaction with the pluripotency factor Oct4 and on Activin/Nodal signaling. Use of an in vitro model of epiblast maturation, relying on the differentiation of ESCs into epiblast stem cells (EpiSCs), revealed that this process entails a shift in the regulation of Nodal expression from an HBE-driven phase to an ASE-driven phase, ASE being another autoregulatory Nodal enhancer. Deletion of HBE in ESCs or in EpiSCs allowed us to show that HBE, although not necessary for Nodal expression in EpiSCs, is required in differentiating ESCs to activate the differentiation-promoting ASE and therefore controls this regulatory shift. Our findings clarify how early Nodal expression is regulated and suggest how this regulation can promote the specification of extra-embryonic precusors without inducing premature differentiation of epiblast cells. More generally, they open new perspectives on how pluripotency factors achieve their function.
| In the early mouse embryo, Nodal, a member of the TGFbeta superfamily of signalling proteins, promotes the differentiation of extra-embryonic tissues, as well as tissues within the developing embryo itself. Characterising the regulation of Nodal gene expression is essential to understand how Nodal signals in diverse tissue types and at different stages of embryonic development. Four distinct enhancer sequences have been shown to regulate Nodal expression, although none could account for it in the preimplantation epiblast or in embryonic stem cells. We identified a novel enhancer, HBE, responsible for the earliest aspects of Nodal expression. We show that activation of HBE depends on its interaction with a well-known pluripotency factor called Oct4. HBE itself also controls the activation of at least one other Nodal enhancer. Our findings clarify how early Nodal expression is regulated and reveal how pluripotency factors may control the onset of differentiation in embryonic tissues.
| The gene Nodal encodes a TGFβ family member signaling via the Smad2/3-dependent Activin/Nodal pathway. Nodal is a key factor during early development, required for the specification of cell identities in embryonic and extra-embryonic lineages [1],[2]. Its re-expression in the adult has been associated with tumor progression and its signaling pathway is essential to the maintenance of human embryonic stem cells (ESCs) [3]–[5]. There is therefore a broad interest in understanding how its expression is initiated and regulated.
In the mouse, Nodal expression starts in the inner cell mass (ICM) of the E3.5 blastocyst [6],[7]. At E4.0, shortly before implantation, Nodal is detected in the two tissues that derive from the ICM: the epiblast, which will give rise to all fetal lineages, and the primitive endoderm (PrE), an extra-embryonic layer [6]. Nodal expression remains detectable in their postimplantation derivatives up to gastrulation stages but exhibits complex dynamics, foreshadowing the establishment of the anterior–posterior axis and the formation of the primitive streak [1]. Its re-expression in the node at E7.5 and in left lateral plate mesoderm at E8.0 contributes to the establishment of left–right asymmetry [1].
Nodal expression starts at E3.5, but the earliest molecular defects characterized in Nodal−/− embryos so far were detected after implantation. The epiblast of Nodal−/− embryos differentiates prematurely and their visceral endoderm, a derivative of the PrE, is not properly regionalized [8]–[10]. Pluripotent cell lines offer convenient in vitro models to study the role of Nodal and Activin/Nodal signaling during epiblast development. ESCs are derived from the nascent preimplantation epiblast [11]. They express Nodal and have an active Activin/Nodal signaling pathway, but this is not essential to their maintenance [3],[12]. In contrast, epiblast stem cells (EpiSCs) are derived from the postimplantation epiblast, and their capacity to self-renew depends critically on Activin/Nodal signaling [13],[14]. When exposed to Activin and FGF, ESCs can be converted into EpiSCs, a differentiation process dependent on Activin/Nodal signaling and described as a transition from a ground state of pluripotency to a primed state of pluripotency [11],[15]. This protocol is now commonly used to mimic events surrounding the maturation of the preimplantation epiblast into postimplantation epiblast.
Several studies showed that in ESCs Nodal expression is dependent on pluripotency factors or on Activin/Nodal signaling itself [16]–[19]. Four Nodal cis-regulatory elements are already known. None is controlled by pluripotency factors, and only one, ASE, is both dependent on Activin/Nodal signaling and known to be active before implantation [6],[20],[21]. ASE contains two functional FoxH1-Smad2,3 binding motifs and acts as an autoregulatory element allowing Nodal to amplify its own expression, notably in the postimplantation epiblast [20],[21]. The deletion of ASE results in a phenotype far less severe than that of Nodal−/− embryos and characterized by later patterning defects [20], indicating that it is not required to initiate Nodal expression. Our previous analysis of the expression profiles of fluorescent reporter transgenes for ASE showed that, although they could recapitulate some aspects of Nodal expression at preimplantation stages, they could not account for the timing of its onset in the ICM and its presence in nascent preimplantation epiblast cells [6]. This strongly suggested that these particular aspects of Nodal expression are dependent on cis-regulatory sequences other than ASE.
We sought to uncover how Nodal expression is initiated. We identified a novel Nodal enhancer, which we call HBE, that matches the expected profile. HBE is activated ahead of other Nodal enhancers in the ICM and in the preimplantation epiblast, and it is the predominant Nodal enhancer in ESCs. Furthermore, HBE is a hotspot for the binding of pluripotency factors and mediates the influence of Oct4, Klf4, and Activin/Nodal signaling on the expression of Nodal. The deletion of HBE by homologous recombination eliminates expression of the mutated allele in ESCs and in the early embryo. Strikingly, it also impairs its expression when ESCs are induced to differentiate, revealing an early requirement for HBE to trigger the activation of at least one other enhancer, the ASE, which drives Nodal expression in more differentiated cell types. We find also that the deletion of HBE in ESCs results in a region close to ASE accumulating the repressive histone mark H3K27me3, implying that it is via its implication in the recruitment of chromatin modifiers that HBE controls ASE. Our findings shed light on how enhancers regulated by the molecular machinery of pluripotency control gene expression and drive development forward.
One study identified Nodal as a tentative direct target of the pluripotency factors Oct4, Sox2, and Nanog in ESCs [19]. It showed that the expression of Nodal declined when the gene encoding Oct4 was knocked down, whereas it was upregulated when Nanog or Sox2 were supressed. We therefore searched relevant ChIP data, which revealed the existence of a hotspot for the binding of pluripotency factors, including Oct4, Nanog, Sox2, and Klf4, in a 2 kb region lying 1 kb upstream of the Nodal transcription start site (TSS) (Figure 1A) [22]–[26]. We called this region HBE, for highly bound element. This noncoding sequence is conserved in eutherian mammals, an indication that it may be involved in gene regulation (Figure 1A). In ESCs, this sequence scores positive for four criteria now used to identify active enhancers: low levels of the repressive histone mark H3K27me3, low levels of the active but promoter-associated histone mark H3K4me3, high levels of the active histone marks H3K4me1 and H3K27ac, and a binding peak of the acetyltransferase and transcriptional coactivator p300 [27]–[31] (Figure S1). In contrast, none of the known Nodal enhancers, PEE, NDE, AIE/LSE, or ASE [32]–[35], appeared to bear the hallmark of an active enhancer in ESCs (Figure S1). The ASE, however, although not bearing the active enhancer mark H3K4me1, presents marks suggestive of possible transcriptional activity: a binding peak for p300, high level of H3K27ac, and a peak of the active promoter-specific H3K4me3.
A luciferase-based assay was used to test HBE's capacity as an enhancer in ESCs and to compare it to that of ASE and PEE, the only Nodal enhancers known to be active at peri-implantation stages [6],[36]. This assay was done both with the minimal promoter E1b [36] and with the 940-bp-long stretch of sequence, termed NIS, for Nodal intervening sequence, which separates HBE from the ORF of the gene and contains the endogenous Nodal promoter. In both cases, HBE came out as the strongest enhancer (Figure 1B), whereas PEE and ASE showed minimal activity and NDE and AIE/LSE showed no activity whatsoever. We performed the same assay in EpiSCs. This time, although HBE still showed enhancer activity, the activity of ASE was higher while that of PEE, NDE, and AIE/LSE was unchanged (Figure 1C). The higher activity of ASE is consistent with it being dependent on Activin/Nodal signaling [6],[20],[34] and the presence of Activin in EpiSC culture medium. These results indicate that HBE is the predominant Nodal enhancer in ESCs and that it is still active in EpiSCs.
To find out when and where HBE is active during embryonic development, we generated transgenic lines where the expression of a nuclear version of Venus-YFP is placed under the control of HBE-NIS—that is, the 3 kb of genomic sequence directly upstream of the Nodal ORF. The two independent HBE-YFP mouse lines we obtained both showed the same reporter expression profile, thus precluding the influence of position and confirming its specificity (Figure 2). The fluorescence was first detected at E3.5 in one or two cells of the ICM (n = 12/15 embryos analyzed; Figure 2A–A″). By E4.5, more ICM cells were positive and the signal was stronger (Figure 2B–C″, 2E–E″). These cells all co-expressed the pluripotency factor Oct4 (Figure 2B–B″). Counts performed on E4.5 embryos stained for the PrE marker Gata-4 found that 93% of epiblast cells were YFP-positive. Most YFP-positive cells (98%) were also found to co-express the pluripotency factor Nanog (Figure 2C–C″). This is in marked contrast to the ASE-YFP transgene, which showed an expression profile broadly complementary to that of Nanog in the epiblast around the time of implantation [6], and suggests that HBE-YFP is expressed in epiblast cells earlier than ASE-YFP. However, at these early stages HBE-YFP expression is not restricted to the embryonic lineage. Co-expression with Gata-4 was detected in a subset of PrE cells in some embryos at E3.75 and E4.5 (n = 3/13 and n = 5/11, respectively; Figure 2D–D″, E–E″). There was no expression in extra-embryonic endoderm after this (unpublished data and Figure 2F–F″). After implantation, at E5.5, HBE-YFP was expressed in all epiblast cells, albeit with varying levels of intensity (n = 15/16; Figure 2F–F″). By E6.5, the expression of the transgene in the epiblast was clearly heterogeneous (n = 13/13; Figure 2G–G″), suggesting it was progressively downregulated in some cells whereas it was maintained in others. Between E6.5 and E7.5, HBE-YFP–positive cells could still be detected in the epiblast and in all epiblast derivatives, including the extraembryonic mesoderm (Figure 2G–G″ and unpublished data). However, they constituted a steadily declining fraction of these tissues. At E8.0, fluorescent nuclei were still detected in the node and in cells scattered in all three germ layers along the full length of the headfold stage embryo (Figure 2H,I). By E8.5, HBE-YFP expression was no longer detected (unpublished data).
Although HBE-YFP fluorescence just became detectable at E3.5, in situ hybridization with a YFP probe detected expression of the transgene in the ICM of all E3.5 transgenic embryos analyzed (n = 16/16; Figure 2J), whereas a similar analysis previously detected the ASE-YFP transgene in the ICM of no more than 50% of the embryos [6] (A.P.G. and J.C., unpublished data).
The expression profile of HBE-YFP does account for the early aspects of Nodal expression that were not fully recapitulated by the ASE transgene. It suggests HBE could be involved in the regulation of Nodal expression from its onset at E3.5 until late gastrulation stages.
The fact that HBE is a hotspot for the binding of pluripotency factors in ESCs suggests that this sequence is the interface enabling these factors to modulate Nodal expression. To test this hypothesis we first assessed the influence of Oct4 and Nanog on HBE enhancer activity, using genetically modified ESC lines. RCNβH ESCs contain a conditional allele of Nanog, which can be deleted by exposure to Tamoxifen—triggering GFP expression [37]. Luciferase assays showed that the enhancer activity of HBE was not affected by the resulting absence of Nanog (Figure 3A), indicating that it is not via HBE that Nanog represses Nodal expression [19]. Successful deletion of Nanog was confirmed by the up-regulation of GFP and the downregulation of Nanog itself (Figure S2A–B″), whereas Oct4 expression was maintained (Figure S2C–D″). In contrast, in ZHBTc4 ESCs, where Doxycyclin treatment induces a knockdown of Oct4 [38], Oct4 depletion drastically down-regulated the expression of HBE constructs (Figure 3B), suggesting that HBE mediates the influence of Oct4 on Nodal expression [19]. Successful down-regulation of Oct4 was confirmed by immunofluorescence (Figure S2E–F′). However, these experiments could not establish whether the activity of HBE required a direct interaction between this enhancer and Oct4.
A systematic analysis was then undertaken to determine how the major pluripotency factors known to bind HBE contribute to its transcriptional activity in ESCs. Sequence comparison among eutherian mammals had uncovered four conserved regions within HBE, which we called HBE1 to 4 (Figure 3C). We used the BiFa bioinformatic tool [6],[39] to identify putative binding sites for Oct4, Nanog, Sox2, and Klf4 over the entire HBE sequences (Figure 3C and Figure S3A). Putative binding sites for Nanog/Sox2 (2), Sox2 (1), Klf4 (10), and Oct4 (3) were found in HBE2 and 3. Only these two regions showed significant enhancer activity, which was drastically increased when these two sequences were combined (Figure 3D). Fragments of HBE23 of increasing lengths were then assayed to identify subregions that are critical for this activity. Significant increases in enhancer activity were seen when fragments HBE2d, which contains a cluster of putative Klf4 binding sites, and HBE3d, which contains putative Oct4 and Nanog/Sox2 binding sites, were added to the reporter construct (Figure 3E). The addition of HBE3d resulted in the most dramatic gain in enhancer activity.
To assess the relevance of these binding sites to HBE enhancer activity, they were all mutated in HBE23-E1b and HBE-NIS luciferase constructs. Point mutations were designed with the help of the BiFa algorithm so as to prevent binding of the relevant transcription factor to its putative target sequence, while minimizing effects on the binding of other transcription factors. The impact of each mutation on transcription was first assessed separately and then in combination with others. We found that putative binding sites for all four factors—Nanog, Sox2, Oct4, and Klf4—were contributing to HBE23 enhancer activity in ESCs (Figure 3F). Mutations in Klf4 and Oct4 binding sites were, however, far more detrimental to this activity than mutations in Nanog and Sox2 binding sites. In particular, the elimination of the first Oct4 binding site in HBE3d was the single mutation causing the most dramatic drop in luciferase activity (Figure 3F). Its combination with mutations in the two other putative Oct4 binding sites did not reduce this activity further (Figure S3B). We confirmed that this single mutation was able to prevent the binding of Oct4 in gel shift assays with ESC extracts (Figure S4A).
Mutations in Nanog and Sox2 putative binding sites only had a significant impact on Luciferase activity when they were all combined in an NIS-driven construct, and still the decrease was modest (Figure 3F). The BiFa algorithm identified all putative Nanog binding sites in HBE as putative, lower ranking, Sox2 binding sites. In gel shift assays, extract from Nanog-depleted RCNβH ESCs slowed the migration of the target sequence we tested, indicating that it was bound by one other factor at least (Figure S4B). The mutated version of the sequence, however, prevented this binding, indicating that although some factor, such as Sox2, could possibly compensate for the absence of Nanog in RCNβH ESCs, our mutation allowed the contribution of their common binding sites to HBE and Nodal regulation to be assessed. Together with the Oct4 result, this suggested that our approach to mutation design was effective. We found that the addition of all Nanog and Sox2 mutated binding sites to a construct already containing all Klf4 and Oct4 mutated binding sites did not reduce its transcriptional activity further (NSKO*; Figure 3F), suggesting that the contribution of Nanog and Sox2 to HBE enhancer activity is secondary to that of Oct4 and Klf4.
Notably, we found that the first Oct4 binding site in HBE3d, the one most critical to HBE enhancer activity, is the most conserved of all the putative binding sites we identified in HBE, as it is the only one present in all mammalian genomes tested so far (Figure 3G). Furthermore, this conserved stretch of DNA contains an extended version of the Oct4 binding site that recent evidence suggests can be bound by Oct4 alone and is critical to its reprogramming function [40],[41].
To confirm the relevance of our findings to the regulation of HBE in vivo, we electroporated eight-cell stage embryos with constructs in which a nuclear version of Venus-YFP is under the control of either native HBE or its KO* version, where all Klf4 and Oct4 putative binding sites are mutated. Electroporation efficiency was assessed by co-electroporating a construct expressing mCherry under the control of the strong promoter CAG. Electroporated embryos were cultured 30 h, allowing most of them to reach the blastocyst stage. A majority of the embryos that had been electroporated with the native HBE construct (n = 19/24) showed YFP expression in a few cells. In contrast, embryos that had been electroporated with the mutated HBE-KO* construct showed only very weak or undetectable expression of YFP (12/12; Figure 3H–I″).
These results indicate that both in ESCs and in preimplantation embryos, HBE is under the control of pluripotency transcription factors, notably Oct4 and Klf4, whose cognate binding sites are critical to its enhancer activity.
The fact that not all E3.5 to E4.5 Oct4-positive ICM cells expressed HBE-YFP in transgenic embryos suggested that some other factor was essential for the activation of HBE. Several studies have shown that Nodal expression in ESCs is dependent on Activin/Nodal signaling [16]–[18]. Furthermore, a recent genome-wide ChIP study showed that, in ESCs, pSmad3 co-occupies the genome with Oct4, with which it forms a complex, and that this correlated with sensitivity to TGFβ signaling for Oct4-bound genes [42]. Notably, this study showed that with respect to Nodal expression, Oct-4 depletion led to a 5-fold reduction in its response to Activin exposure. Two of the positions where both Oct4 and Smad3 were found to bind are within HBE (Figure 1A). Our own results showed that reporter constructs and reporter transgenes for the Activin/Nodal signaling-dependent ASE had very limited transcriptional activity in ESCs (this study, and N. Sasaki, A.B., and J.C., unpublished results). Together, these data strongly suggested that Activin/Nodal signaling might be the other signal required to elicit HBE activation in preimplantation epiblast. To test this hypothesis, we cultured E2.5 HBE-YFP embryos for 48 h in the presence of 40 µM SB-431542, a pharmacological inhibitor of the type I Activin receptors ALK4, 5, and 7 [43]. We found that SB-431542–treated embryos had a similar number of Oct4-positive cells as DMSO-treated control embryos, indicating that at this concentration the formation of the ICM is not significantly affected (Figure 4A–E). SB-431542 exposure nevertheless resulted in a drastic reduction of the percentage of YFP-positive embryos and of YFP-positive cells among Oct4-positive ones. In addition, cells that expressed the transgene in SB-431542–treated embryos did so at a lower level than their counterparts in DMSO-treated embryos (unpublished data). We conclude that HBE-YFP expression is dependent on Activin/Nodal signaling, presumably reflecting a similar requirement for the activation of the endogenous HBE.
Having established that HBE is an enhancer active in ESCs and in the mouse embryo, we assessed its contribution to Nodal expression. We generated a targeting construct in which HBE was floxed and the first 80 bp of Nodal ORF were replaced by the coding sequence for a destabilized nuclear Venus-YFP, so that the expression of the modified allele could be monitored (Figure S5A and Figure 5A). Successful targeting of the Nodal locus in ESCs was confirmed by PCR and Southern hybridization (Figure S5B–D). Almost all recombinant cells expressed the YFP, although at different levels (Figure 5B–B″). In contrast, Cre-mediated deletion of HBE resulted in most cells having completely lost YFP expression 2 d after transfection, indicating that HBE is essential to Nodal expression in ESCs (Figure 5C–C″). The few cells expressing YFP (∼7% of total) tended to be found at the periphery of colonies and to have low or no Oct4 expression, suggesting they corresponded to differentiating cells in which Nodal expression was driven by other enhancers.
To investigate this possibility, we analyzed the expression of the HBE-deleted allele in EpiSCs, where our luciferase-based assays had shown that ASE is the predominant Nodal enhancer. We thus induced ESCs carrying the conditional HBE allele NodalcondHBE-YFP to differentiate into EpiSCs. Real-time PCR (RT-PCR) analysis of the expression dynamics of four key markers—Klf4, Oct4, FgF5, and Bra—confirmed the successful conversion of the cells to an EpiSC identity (Figure S6A). RT-PCR analysis showed that the NodalcondHBE-YFP allele and the wild-type (WT) Nodal allele followed similar expression dynamics, indicating that the conditional allele is a fair reporter of WT Nodal expression (Figure S6B and unpublished data). EpiScs carrying the NodalcondHBE-YFP allele were then transfected with two constructs expressing either the Cre recombinase or the fluorescent marker mCherry. Widespread mCherry expression confirmed that transfection was efficient (Figure S6C–C′), whereas RT-PCR on genomic DNA showed that HBE deletion frequency was close to 90% 4 d after transfection (Figure S6D). We found that 6 d after transfection the expression of NodalΔHBE-YFP was maintained at a level similar to that of the undeleted allele (Figure 5D–E″, Figure S6E). This result indicates that HBE is not required for the expression of Nodal in EpiSCs, which is thus driven by another Nodal enhancer, presumably ASE.
To investigate the dynamics of the transition from an HBE-driven Nodal expression to an ASE-driven one, we induced ESCs carrying either the conditional HBE allele NodalcondHBE-YFP or the HBE-deleted allele NodalΔHBE-YFP to differentiate into EpiSCs. RT-PCR analysis showed, as expected, that the expression of NodalΔHBE-YFP was much lower than that of NodalcondHBE-YFP at the beginning (Figure 6A). Surprisingly, it did not recover, even after 10 d of differentiation. Comparison with the expression of the undeleted allele showed an average difference of about 80%, and immunofluorescence detected the YFP in just a few cells (Figure 6B–D″). Like in ESC colonies, these rare YFP-positive cells had lower or no Oct4 expression (Figure 6D″). Together with the earlier finding that HBE is not required for NodalΔHBE-YFP expression in EpiSCs, this indicates that prior to its deletion in EpiSCs, HBE contributed to a modification of the locus critical for the activation of ASE, which allowed NodalΔHBE-YFP to be expressed in EpiSCs. These results demonstrate that during the conversion of ESCs into EpiSCs, HBE is initially required to promote the activation of ASE.
As ASE is dependent on Activin/Nodal signaling and as Nodal in NodalΔHBE-YFP cells is still produced by the WT allele, we hypothesised that HBE is required to potentiate the activation of ASE at the chromatin level. We used ChIP to track changes in the distribution of the mutually exclusive H3K27me3 and H3K27ac histone marks at different positions in the locus. This analysis revealed that after HBE deletion, a region 5′ to the ASE sees a 2.5-fold increase of the repressive H3K27me3 mark and a 2-fold decrease of the active H3K27ac mark. These modifications are specific to the recombinant allele. No changes were detected at the 3′ end of the autoregulatory enhancer. No changes either were detected immediately upstream and downstream of the deleted HBE (Figure 6E–G). This result demonstrates that HBE controls the chromatin status of a region adjacent to ASE and therefore suggests that it is via the recruitement of chromatin modifiers that HBE exerts an influence over ASE activation.
To investigate whether HBE is necessary for the expression of Nodal in vivo as it is in vitro, chimeric embryos were generated. NodalcondHBE-YFP and NodalΔHBE-YFP cells were first stably transfected with mCherry so that they could be traced in chimeric embryos. Small groups of these cells were then aggregated with E2.5 morulae, and the resulting blastocysts were either cultured in vitro until the equivalent of stage E4.5 or reimplanted into pseudopregnant mice and allowed to develop in utero until the equivalent of stage E6.5. Chimerism was very high as judged by the number of mCherry-positive cells in the epiblast of the aggregation chimeras. Embryos generated from NodalcondHBE-YFP cells expressed YFP in the epiblast (n = 34/48 of stage E4.5 and 7/7 of stage E6.5 embryos analyzed; Figure 7A and C), and this expression was consistent with the expected expression profile for Nodal, notably showing a restriction to the proximal posterior epiblast at E6.5. In contrast, embryos generated from NodalΔHBE-YFP cells did not express the fluorescent marker or expressed it at very low levels in just a few cells (n = 44/45 of stage E4.5 and 7/7 of stage E6.5 embryos analyzed; Figure 7B and D), indicating that HBE is required for the activation of Nodal transcription in epiblast cells in vivo, as in vitro differentiation experiments suggested.
Genome-wide ChIP studies have shown that in ESCs, pluripotency factors co-occupy the genome at specific multitranscription factor-binding loci (MTL) through which they control the pluripotent state of the cells [22],[24]–[26],[44]. These studies led to the view that the core transcription factors of the pluripotency gene regulatory network (GRN), Oct4, Nanog, and Sox2, form an interconnected autoregulatory loop that positively regulates their own promoters, activate the expression of genes necessary to maintain the pluripotent state, and contribute to the repression of genes promoting differentiation [45]–[47]. We identified HBE as an MTL at the Nodal locus. Our results confirm that this region is a target of the molecular machinery of pluripotency and of the Activin/Nodal signaling pathway, as ChIP studies predicted [22],[24]–[26],[42].
We found that HBE has enhancer activity in ESCs, as was the case for all Oct4/Sox2/Nanog MTLs tested so far [22],[47]. HBE is in fact the only Nodal enhancer active in ESCs. Moreover, it is activated early on during mouse embryonic development. Transgenic embryos expressing YFP under the control of HBE up-regulate the fluorescent marker in the ICM of the E3.5 blastocyst. Its expression is then restricted to the embryonic epiblast and is maintained in its embryonic and extra-embryonic derivatives until organogenesis starts at E8.5, at which point Oct4 expression and pluripotency are lost [48].
We showed that the enhancer activity of HBE is dependent on Oct4 and Klf family members. In fact Oct4 is the master pluripotency factor most critical to this activity. This is consistent with studies suggesting that unlike other master pluripotency factors, Oct4 is a strong transcriptional activator [49]. It appears to function as a pioneer factor at enhancers, opening up the chromatin and allowing other factors, such as pSmad3, to access their binding sites [42]. The main Oct4 binding site in HBE is the only one of all the putative pluripotency factors binding sites we identified that is extensively conserved among placental mammals, suggesting that HBE evolved around this particular sequence.
We also found that the enhancer activity of HBE is dependent on Activin/Nodal signaling and we showed previously that Activin/Nodal signaling is activated in Nodal−/− blastocysts [6]. In other animal models, there is consistent evidence of another TGFβ family member acting upstream of early Nodal expression [50]–[54]. Gdf1 and Gdf3, two possible TGFβ-related candidates in the mouse, appear however unable to activate the Smad2/3 pathway at physiological concentrations [55]–[57]. This was confirmed when we showed that Gdf3 cannot replace Nodal in vivo [6]. Better candidate ligands for the early activation of the Smad2/3 pathway and of HBE are thus Activins, which are present in the ICM as well as in the oviduct and uterine epithelia prior to implantation [58]. Because Nodal was also found to be expressed in the endometrium of E3.5 pregnant females, one cannot discount the possibility that Nodal of maternal origin might be involved in the induction of Nodal expression in the embryo [59].
The finding that the onset of Nodal expression is dependent on the pluripotency GRN coincides with a growing realization that in the context of the embryo so-called pluripotency factors are in fact actively engaged in promoting development. Nanog, described as the guardian of pluripotency in ESCs [37], is required in epiblast precursors to promote, by a non-cell-autonomous mechanism, the differentiation of adjacent PrE precursors [60]. It has also been shown recently that Oct4 promotes PrE development through both cell-autonomous and non-cell-autonomous mechanisms, and more generally favors embryo development via its control of multiple metabolic pathways [61]. Recent work indicates that Activin/Nodal signaling may first be required in the PrE around E4.0 to specify a subset of Lefty1-expressing PrE cells, the descendants of which will later give rise to the distal visceral endoderm (DVE), a group of cells playing a critical role during the establishment of AP polarity [2],[7]. It is therefore possible that the HBE-dependent expression of Nodal in the blastocyst contributes to this initial regionalization of the PrE. During the transition from pre-implantation to postimplantation epiblast, Nodal undergoes a regulatory shift, from an HBE-driven phase to an ASE-driven one, which correlates with an increase in its expression levels and an up-regulation of differentiation promoting downstream targets, also seen in EpiSCs [6],[13],[14]. In ESCs, most genes involved in lineage specification are in a poised state that is transcriptionaly silent but ready to be activated by developmental signals. This state is defined by the presence of both active and repressive histone marks on the promoters of these genes. Repressive marks are introduced by chromatin modifiers locally recruited by Oct4, Sox2, and Nanog [47]. Smad2/3 complexes, activated by the Activin/Nodal pathway, can remove these repressive marks and induce the expression of downstream targets such as Gsc and Mixl1. Yet although Nodal is expressed in ESCs, Gsc and Mixl1 remain poised in these cells. This can be partly explained by the relatively low level of Nodal expression in ESCs and by the co-expression of genes known to restrain its signaling activity, such as Smad7, Lefty1, and Lefty2. These data suggest that in the blastocyst components of the Activin/Nodal signaling pathway are tightly regulated to ensure proper embryonic and extra-embryonic development. Initially, activation of Nodal by HBE produces low levels of the signal that specify certain extra-embryonic precusors, possibly of the DVE, while minimizing the exposure and the response of nascent epiblast to prevent its premature differentiation. During subsequent stages of development the autoregulatory ASE takes over. This shift from an HBE-driven phase to an ASE-driven one results in an amplification of the Nodal signal, which triggers the differentiation of the epiblast.
We found that HBE is required in differentiating ESCs for the activation of ASE. When HBE is deleted in EpiSCs, ASE, the predominant Nodal enhancer in this cell type, is active. However, if HBE deletion occurs in ESCs, before their differentiation into EpiSCs, ASE does not drive expression of the gene. Our results suggest that once bound to HBE, master pluripotency factors induce local modifications of the chromatin that in turn affect the ability of the ASE to interact with the adjacent promoter, and thus Nodal expression levels. Changes in the combination of HBE-bound factors, such as those taking place during epiblast maturation or ESC to EpiSC transition, could modify the effect HBE has on ASE.
Although Nodal is expressed in ESCs, the autoregulatory enhancer ASE is not active in these cells. One hypothesis is that Nanog acts at the Nodal locus to prevent ASE activation. We found previously that the expression of the ASE-YFP reporter transgene is only detected in epiblast cells with low or no Nanog [6]. This is consistent with the results of luciferase assays in ESCs and EpiSCs that correlate a higher level of ASE transcriptional activity with a lower level of Nanog. Nanog depletion in ESCs results in an increase in Nodal expression [19], yet we found that Nanog depletion, or the elimination of Nanog binding sites, had no effect on the transcriptional activity of HBE. Because Nanog binds only HBE at the Nodal locus in ESCs, it must act from this position to prevent ASE activation. This would keep Nodal expression, and thus Activin/Nodal signaling, low as long as Nanog is present. Its down-regulation during the conversion of ESCs into EpiSCs signaling by unlocking ASE would then allow an increase in Activin/Nodal.
The dependency of ASE activity on HBE may also involve Oct4, but in a role opposite to that proposed for Nanog. HBE-bound Oct4 could promote ASE activation. The mechanism described for the activation of poised genes by companion Trim33-Smad2/3 and Smad4-Smad2/3 complexes [62] suggests a similar scenario for the HBE-dependent activation of ASE. The Oct4-Smad3 complex bound on HBE could initiate chromatin modifications that would then allow the interaction of ASE with the adjacent promoter, leading to the transcriptional activation of Nodal by the autoregulatory element and the amplification of the Nodal signal. The results obtained in aggregation chimeras suggest that ASE may not be the only Nodal enhancer whose activation is controlled by HBE. The lack of expression of the NodalΔHBE-YFP allele in proximal and posterior epiblast cells at E6.5, where Nodal expression was shown to be independent of ASE, but where transgenic PEE reporters were found to be expressed [6],[20],[63], do suggest a similar influence on PEE.
The implication of Oct4 in such an unlocking mechanism would be consistent with recent studies showing that the capacity of ESCs to differentiate is critically dependent on the level of Oct4 not being too low [64],. Such a mechanism may concern the regulation of differentiation-promoting genes other than Nodal. Further studies will be necessary to test these hypotheses and get a better understanding of how HBE-bound factors contribute to the regulation of Nodal expression.
To conclude, our results complete the picture on the regulation of Nodal at early stages. They show that HBE has a dual role, acting both as an enhancer and as a modulator of the activity of other regulatory elements. Our analysis of its regulation and mode of action furthers our understanding of the distinct roles assumed by master pluripotency factors and of the complex fashion in which the molecular machinery of pluripotency controls gene expression (Figure 8). It is likely that similar mechanisms are involved in the regulation of genes other than Nodal. Our results are consistent with the notion that the need to control Activin/Nodal signaling is one of the leading influences on the evolution of the pluripotency GRN.
Experiments were performed in accordance with French Agricultural Ministry and European guidelines for the care and use of laboratory animals. The project has been reviewed and approved by the Animal Experimentation Ethical Committee Buffon (CEEA-40). It is recorded under the following reference: CEB-35-2012.
Potential binding sites at endogenous and mutated sequences were scored statistically using the Binding Factor (BiFa) tool [6]. Weight matrices from the TRANSFAC database v2009.4 [66] were used. The alignment of the main Oct4 binding site was retrieved from the Ensembl database release 73. It belongs to «36 eutherian mammals EPO LOW COVERAGE» (positions 61,416,797 to 61,416,833 on mouse chromosome 10). The alignment was visualized using Jalview 2.8 [67] using data for a subset of available species.
See Materials and Methods S1 for detailed CCE, ZHbTc4, and RCNbH mouse ES cell culture conditions. Inhibition of Oct4 expression in ZHbTc4 cells was induced with 0.1 mg/ml Doxycyclin (Sigma), whereas Nanog knock-down in RCNbH cells was induced with 1 µM 4-Hydroxy-Tamoxifen (Sigma). We transiently transfected 200,000 ES cells with 1 µg of any of Firefly Luciferase constructs and 0.05 µg of the pCAG-Renilla Luciferase construct (in 50 µl DMEM) and 2 µl Lipofectamine 2000 (Invitrogen—in 50 µl DMEM) according to the manufacturer's instructions and harvested them 24 h after transfection.
ES cells were grown as previously described [13]. EpiSC-like colonies start to appear at passage 3 (day 6), and colonies were passaged by mechanical dissociation after 30 s treatment with accutase at room temperature. Colonies were passaged every 2 d and diluted 3 to 4 times.
Site-directed mutagenesis of HBE was performed by two rounds of PCR amplification. First, complementary primers containing the point mutations as well as primers complementary to the 5′ or the 3′ ends of the sequence were used to amplify the two parts of HBE that contain the mutated sequence at one end. Then, the two parts were used as the template for the amplification of the whole sequence, using the end primers alone. Multiple point mutations were introduced sequentially. See Materials and Methods S1 for primer sequences.
The luciferase activities of the cell lysates were measured by means of the Dual-Luciferase Reporter Assay System (Promega) in a Berthold Centro LB 960 device. The activity of the firefly luciferase was measured for 60 s, whereas the activity of the Renilla luciferase was measured for 0.5 s. Finally, the normalised values for HBE and HBE23 were arbitrarily set to 10. Activities are reported as mean standard errors of a minimum of three independent experiments.
Total RNA was prepared using NucleoSpin RNA Kit (MN) followed by DNaseI (Roche) treatment. First-strand cDNA was synthesised using Vilo reverse transcriptase (Invitrogen). Real-time PCR was performed using FastStart SYBR Green Master (Roche). Gene expression was determined relative to Gapdh using standard curve calibration. All quantitative PCR reactions were performed in LightCycler 480 (Roche). See Materials and Methods S1 for primer sequences.
A DNA construct expressing Venus-YFP fused to 3 NLS was linearised, gel-purified, and resuspended in Tris 10 mM, EDTA 0.25 mM, pH 7.5. Transgenic founders were obtained after microinjection of the DNA into (C57BL/6 × CBA) F2 fertilized eggs (1 or 2 ng/ml in injection buffer). Heterozygous embryos carrying the HBE-Venus transgene were generated by mating homozygous transgenic males with WT Sw females. The genotyping was done as described for the ASE-YFP transgene [6].
Mice mating and embryo collection were as described [6]. Eight-cell stage uncompacted Swiss × Swiss mouse embryos were collected in M2 (Sigma), shelled in Tyrode's solution (Sigma), and electroporated in a flat electrode chamber with a 1 mm gap between the electrodes (BTX Inc., San Diego, CA) in 1× HBS DNA solution containing 0.25 µg/µl of the mCherry expressing control plasmid and 1 µg/µl of the Venus expressing experimental plasmid. Two sets of four pulses of 1 ms each at 25 V were delivered, with 100 ms intervals between the pulses and a 1 min interval between the two sets of inverted polarity. The embryos were then cultured in G2 (Vitrolife) at 37°C and 5% CO2 for 30 h.
Eight-cell stage uncompacted transgenic ASE-YFP embryos were transferred to an eight-well Netwell plate (Costar) with 400 µl of G2v5PLUS (Vitrolife). They were cultured for 48 h at 37 °C/5% CO2 in the presence of 20, 40, or 50 µM SB-431542 (Sigma) in DMSO, to test for dose toxicity and effectiveness. Control embryos were cultured in the presence of the same amount of DMSO. We found as previously that treatment with 40 µM SB-431542 was required to significantly decrease the activity of the ASE-YFP transgene [6]. This dose was not toxic for cultured embryos and was thus chosen to perform similar inhibition experiments on eight-cell stage uncompacted transgenic HBE-YFP embryos.
Cells on coverslips were fixed in 4% paraformaldehyde, permeabilized in PBS/0.3% Triton blocked with 10% FBS in PBS, and incubated with the primary and secondary antibodies (diluted in blocking solution). Nuclei were marked with DAPI- D9564 (Sigma) and cortical actin was marked with 0.5 µg/ml Alexa 647-conjugated Phalloidin (both Molecular probes) and the coverslips mounted on slides with Mowiol 4–88 (Sigma). Immunofluorescence on embryo were done as described [6]. See Materials and Methods S1 for antibody combinations.
ISH was performed as described previously [6].
16×106 CK35 ES cells were transfected with 20 µg of linearised homologous recombination construct containing 12 Kb of the Nodal locus with Venus-YFP fused to three NLS and a PEST sequence replacing the first exon of the gene, two loxP sequences flanking the HBE, a Neo cassette flanked by two FRTs, and a dtA cassette. Transfection was performed by electroporation in two batches of 0.5 ml each in an 0.4 mm gap Biorad cuvette using the Biorad GenePulser and its Capacitance Extender at 200 V and 950 µF capacitance. Selection was performed with 0.2 mg/ml G418. Recombinant clones were further tested by PCR and Southern hybridization.
ChIP experiments were performed as described [68]. All ChIPs were done in triplicate and analyzed by duplicate qPCRs. Real-time PCR was performed on Roche Lightcycler using Roche SYBR Green mix (Roche, Switzerland). Five genomic regions were chosen on the Nodal locus as shown on Figure 6F. The occupancy of these regions was quantified by quantitative PCR analysis of the ratio of the ChIP signal versus the input signal. The following antibodies were used: anti-acetyl K27-Histone H3 (abcam, ab4729) and anti-trimethyl K2-Histone H3 (Millipore, 07-449), and for mock ChIP, anti-GFP (lifetechnologies, A11122). See Materials and Methods S1 for primer sequences.
NodalcondHBE-YFP and NodalΔHBE-YFP ES cells were labelled with nuclear mCherry by transfection with a plasmid expressing mCherry under the control of the strong promoter CAG and the neomycin resistance gene. mCherry-positive cells were selected with 0.2 mg/ml G418. Eight-cell stage Swiss × Swiss mouse embryos were collected in M2 (Sigma), shelled in Tyrode's solution (Sigma), and co-cultured in G2 (Vitrolife) at 37°C and 5% CO2 with groups of 10–15 of mCherry labelled, NodalcondHBE-YFP, or NodalΔHBE-YFP ES cells. Aggregated chimeras were cultured in G2 for 60–72 h until they reached the equivalent of stage E4.5 or transferred 36 h later into the uterus (up to 10 blastocysts) of E2.5 pseudopregnant mice, where they developped until they reached the equivalent of stage E6.5.
Acquisitions of fixed embryos were performed at Imago Seine Core Facility using confocal microscopes (Zeiss LSM 710 and 780). See supplementary experimental procedures for details (Materials and Methods S1). The total number of cells and/or of labeled cells was obtained by counting cell nuclei manually. All images shown in the article are one 5 µm confocal section.
|
10.1371/journal.pgen.1003453 | Signatures of Diversifying Selection in European Pig Breeds | Following domestication, livestock breeds have experienced intense selection pressures for the development of desirable traits. This has resulted in a large diversity of breeds that display variation in many phenotypic traits, such as coat colour, muscle composition, early maturity, growth rate, body size, reproduction, and behaviour. To better understand the relationship between genomic composition and phenotypic diversity arising from breed development, the genomes of 13 traditional and commercial European pig breeds were scanned for signatures of diversifying selection using the Porcine60K SNP chip, applying a between-population (differentiation) approach. Signatures of diversifying selection between breeds were found in genomic regions associated with traits related to breed standard criteria, such as coat colour and ear morphology. Amino acid differences in the EDNRB gene appear to be associated with one of these signatures, and variation in the KITLG gene may be associated with another. Other selection signals were found in genomic regions including QTLs and genes associated with production traits such as reproduction, growth, and fat deposition. Some selection signatures were associated with regions showing evidence of introgression from Asian breeds. When the European breeds were compared with wild boar, genomic regions with high levels of differentiation harboured genes related to bone formation, growth, and fat deposition.
| The domestic pig, an important source of protein worldwide, was domesticated from the ancestral wild boar in multiple locations throughout the world. In Europe, local types were developed following domestication, but phenotypically distinct breeds only arose in the eighteenth century with the advent of systematic breeding. Recently developed molecular tools for pigs (as well as other livestock species) now allow a genetic characterisation of breed histories, including identification of regions of the genome that have been under selection in the establishment of breeds. We have applied these tools to identify genomic regions associated with breed development in a set of commercial and traditional pig breeds. We found strong evidence of genetic differentiation between breeds near genes associated with traits that are used to define breed standards, such as ear morphology and coat colour, as well as in regions of the genome that are associated with pork production traits. It is well documented that crosses with Asian pigs have been used to modify European breeds. We have found evidence of genetic influence from Asian pigs in European breeds, again in regions of the genome associated with breed standard characteristics, including ear shape and coat colour, as well as production traits.
| The domestic pig is an important livestock species and an important protein source worldwide. The pig originated from the wild boar, Sus scrofa, by multiple independent domestications, mainly in Asia Minor, Europe and East Asia [1], [2]. Domestication and subsequent selective pressures altered the behaviour and phenotypic characteristics of these animals [3]. Local pig types were developed in Europe and Asia after domestication, but the development of phenotypically distinct breeds chiefly occurred with the commencement of organised breeding in the 18th century [4].
Strict organised breeding was adopted to improve and develop livestock breeds and Britain in particular was a main centre of the early improvement of pig breeds [5], [6], as a reaction to increasing demand for meat in the wake of the industrial revolution. From the 18th century pig breeds were selectively bred for specific production traits such as early maturation, rapid growth and increased prolificacy. In addition, the coat colour phenotype (which includes both skin and hair pigmentation) was another morphological trait often used during the selective breeding process. Substantial morphological changes occurred in breeds over a short period of time, resulting in the development of numerous distinct pig breed phenotypes in Britain. Charles Darwin commented on the rapid morphological changes in pig breeds at that time: “Chiefly, in consequence of so much crossing, some well-known breeds have undergone rapid changes; thus, according to Nathusius […] the Berkshire breed of 1780 is quite different from that of 1810; and, since this latter period, at least two distinct forms have been borne the same name.” [4]. Although breeds tended to be formed by complex crossing with numerous other breeds, including a number from Asia, to introduce desirable traits [4]–[6], after improvement the breeds were kept distinct, resulting in highly specialised phenotypically distinct and genetically differentiated pig breeds [7]. From the 20th century, with the recognition of the benefits of genetic improvement and changing consumer preferences, certain pig breeds experienced further strong selection for lean meat content, muscularity and enhanced reproduction [5], [6].
To better understand the genetic basis for phenotypic variation in the pig, studies have focused on important traits relevant to the breed development process with the aim of identifying, characterising and mapping candidate genes, and subsequently identifying the underlying causal mutations and allelic differences between breeds [8], [9]. Studies mapping quantitative trait loci (QTL) have particularly focused on muscle growth. Fine mapping of one of these regions (SSC2) identified a causal mutation in the IGF2 gene, where a single nucleotide change is associated with high muscle content in some commercial pig populations [10]. The level of fat on the carcass is also a production trait of economic impact and QTL studies have mapped loci associated with fat deposition to various chromosomes, in particular SSC4 and SSC7 [11], [12]. Reproductive traits have received attention in pigs with several genes investigated in relation to litter size and the number of teats (ESR, PTHLH and PTHR1) [13]. Coat colour is considerably varied amongst breeds within domesticated animal species and investigations into the genetics of pigmentation have identified numerous loci influencing these traits [8], [9]. Variation at two genes, KIT and MC1R, is associated with a variety of pig breed colour types including red, black and white colouring and belted and spotted phenotypes [14]–[16].
With growing genomic resources, selection mapping approaches are increasingly being implemented to identify genetic variants that underlie the phenotypic diversity in domesticated animals. These approaches involve scanning the genome for levels of population differentiation and diversity [17]. Genome-wide scans for signatures of diversifying selection in livestock species have detected signals revealing candidate genes related to morphological variation such as body size, skeletal formation, cranial structure and coat patterns, and production traits such as muscle conformation and milk yield [18]–[25].
To further explore the genetic variation underlying the phenotypic diversity of pig breeds, a genome-wide scan of a diverse set of commercial and traditional British/European pig breeds was performed to identify genomic regions showing signatures of between-breed (diversifying) selection using levels of breed genetic differentiation (FST). Based on these results, sequence data from three candidate regions was analysed to investigate potential causative variants.
A genome-wide scan for signatures of selection in 13 European pig breeds (Table 1) was carried out by estimating Wright's FST, a measure of population genetic differentiation, at each genetic marker. After adopting a sliding window approach, candidate regions that may have experienced diversifying selection were identified by taking the 99th percentile of the empirical distribution of FST–windows (Figure S1). A total of 491 FST–windows per breed were deemed as outlier regions and as many were adjacent SNPs that clustered together, a total of 446 genomic regions displayed strong breed differentiation.
The genome-wide scan revealed five genomic regions of extremely high levels of differentiation that overlapped in five or more breeds; all of these regions contain biologically interesting candidate genes (Table 2). One such region was observed in eight breeds on SSC5 (32.32–34.06 Mb). In all but two of the breeds, the peak FST–window (∼32.6–32.8 Mb) overlapped with the genes WIF1 (32.66–32.72 Mb) and LEMD3 (32.77–32.89 Mb). This region is orthologous to a region in dogs associated with ear morphology [19], [24]. Another region was detected in five breeds on SSC7 (54.00–57.00 Mb), where at the 97.5th percentile a further four breeds also exhibited a signal. On SSC8, a region of high differentiation spanning 71.84–75 Mb was observed in nine breeds. More striking was the extended region of differentiation on SSC8 spanning 40–75 Mb observed in most breeds, with numerous overlapping and non-overlapping peaks of FST across a large genomic region on that chromosome (Figure S1), although fewer than five breeds overlapped directly in their peak FST–windows, except in the narrow interval mentioned above. Duroc was the only breed that did not show high levels of differentiation in this region, or even on that chromosome, at either the 99th or 97.5th percentile. Outlier regions were also found on SSC15 (139.60–142.10 Mb), observed in six breeds, and on SSC16 (18.72–20.63 Mb), observed in five breeds.
Most extreme genomic regions were observed in fewer breeds (1–4) (Figure S1) and we highlight examples of those found in the within-breed 99.9th percentile that overlapped QTLs and contained biologically interesting genes (Table S1). The Duroc breed exhibited several signatures of diversifying selection on two chromosomes. On SSC14 a highly differentiated region (123.08–123.41 Mb) overlapped with QTLs for fatty acid composition in Duroc [26], [27] and includes a gene involved in fatty acid biosynthesis, ELOVL3 (123.08–123.083 Mb) [28]. On SSC15 a highly differentiated genomic region (85.73–86.62 Mb) contained the MYO3B (Class III myosin B) gene (85.63–85.93 Mb), which directly overlapped the peak FST-window (85.83 Mb). An extended differentiated genomic region was observed in the Landrace breed on SSC13, with the highest FST–window occurring at 73.06 Mb, close to the GHRL gene (73.47–73.48 Mb). In addition, QTLs related to various reproductive traits in pigs have been mapped to SSC13 [29] and overlap with the extended differentiated genomic region.
Large, breed-specific signatures of diversifying selection were not limited to the commercial breeds, but also were observed in the traditional breeds (Table S1). Gloucestershire Old Spots displayed a signal of diversifying selection on SSC11, close to EDNRB (54.69–54.72 Mb), a gene implicated in coat colour pattern in mammals [30]. Near the peak FST–window (55.20 Mb) many SNPs in this region were fixed in this breed whereas alleles were segregating in all other pig breeds (Figure 1). A weaker signal in the region of this gene (seen in the 99th but not 99.9th percentiles) appeared in Mangalica and British Saddleback breeds (Figure S1). Another breed-specific signature of selection was observed on SSC5 at a different coat colour locus in the Berkshire. KITLG (KIT ligand, 98.74–98.78 Mb) was just upstream from a 99.9th percentile FST–window (98.84 Mb) on SSC5 and KITLG fell within the 99th percentile differentiation region. Many SNPs in the region of this gene were almost fixed for the same allele in Berkshire and the Asian breed, Meishan, whilst alleles were segregating in the other European pig breeds (Figure 1).
Levels of genetic differentiation were examined between the European pig breeds and wild boar (Table 1). None of the SNPs were found to be fixed for alternative alleles in the pig breeds and wild boar. The genome-wide distribution of FST for domestic pig breeds compared with wild boar is shown in Figure 3A. FST–windows falling into the 99th percentile were viewed as candidates of signatures of selection (Table S11) and contained some biologically interesting genes, as described below.
A genomic region on SSC1 showed high levels of differentiation (1.07–3.19 Mb, Table S11), homologous to a region of the canine genome associated with brachycephaly (broad and short skull shape) in dog breeds [31], [32]. This region contains, amongst seventeen characterised and uncharacterised genes, THBS2 (1.59–1.62 Mb) and SMOC2 (2.23–2.24 Mb), which were suggested as candidates for brachycephaly in the above-mentioned papers (Figure 3B). Pairwise FST–SNPs between wild boar and each breed in this region (48 SNPs) revealed maximum breed average FST values for Tamworth (0.42), Welsh (0.43) and Landrace (0.45), none of which have extremely brachycephalic skulls. A highly differentiated genomic region was also observed on SSC7 (31.30–38.89 Mb, Table S11). This region is close to the pig major histocompatibility complex: class I (∼24–26 Mb), class II (∼29 Mb) and class III (∼27 Mb). Within the differentiated region there are several genes of biological interest, including PPARD (36.14–36.22 Mb) (Figure 3C). Pairwise FST–SNPs (207) between wild boar and each breed in this region revealed highest breed average FST–SNPs in two commercial breeds, Duroc (0.50) and Landrace (0.37), and one traditional breed, Large Black (0.38); the minimum value of breed average FST was in Tamworth (0.09). Another interesting differentiation region observed between the domestic pigs and wild boar was on SSCX (Table S11). Amongst other genes, this region contained AR (60.31–60.50 Mb), the androgen receptor, previously suggested as a candidate gene for backfat thickness in pigs due to its proximity to mapped QTLs [33]. Other regions showing substantial differentiation between wild boar and pig breeds were found on SSC12, SSC13 and SSC14 but no clear candidate genes could be identified.
Consistent with previous studies [34], [35], genome-wide clustering results indicated substantial Asian ancestry for the European breeds. The clustering results indicated that the inferred ancestry of all Meishan individuals (a breed of Chinese origin, Table 1) to the first (“Asian”) cluster was high (92.3–93.9%). In contrast, the inferred ancestry of the European individuals to the second (“European”) cluster was lower (breed averages ranged from 69.6% for Large White up to 87.3% for Mangalica). With levels of ancestry varying across the genome, regions with particularly strong signals of Asian introgression into European breeds were identified according to two criteria: (1) high introgression probabilities (99th percentile) calculated by STRUCTURE software and (2) low differentiation based on FST (below the 1st percentile of individual European breeds versus Meishan) (Table S12). Two candidates of introgression overlapped with signals of selection associated with ear morphology. A genomic region on SSC5 (32–35 Mb), overlapping the region of differentiation detected when prick-eared breeds were contrasted with flat-eared breeds, was found in Gloucestershire Old Spots, Large Black and Mangalica (a signal of introgression in British Saddleback, the other flat-eared breed, was observed in this region only in the FST analysis). A genomic region on SSC7 (33–38 Mb), overlapping with one of the regions of differentiation detected when prick-eared breeds were contrasted with intermediate-eared breeds, was found in British Saddleback, Duroc, Landrace and Welsh. Another signal of introgression was detected on SSC11 (54–55 Mb) in Gloucestershire Old Spots, which overlapped with the differentiated region found in this breed and may be associated with coat pattern. The chromosome with the greatest number of regions showing evidence of Asian introgression was SSC14, where several regions overlapped across multiple breeds (81–85 Mb, eight breeds; 93–94 Mb, four breeds; 96–98 Mb, three breeds; 103–107 Mb, three breeds).
Based on the differentiation results, three genomic regions were further investigated using genome sequence data for 76 individuals from European and Asian breeds (Table S13).
Over the past 300 years, intense artificial selection in European pig breeds for production traits has led to the development of a number of pig breeds with well-defined, specialised phenotypic traits. In this study a number of regions showing between-breed signatures of selection have been identified. Various genes found within these regions can be considered as candidates under selection based on function or previous association with traits that are known to be favoured in pig breeds.
Signatures of diversifying selection were found for traits related to morphological variation described by breeding criteria. Ear morphology is one trait that plays a major role in pig breed standards with strict conditions over ear form. By grouping breeds based on this phenotypic trait, the genome-wide scan suggested that the genetic basis of ear variation in pigs involves at least three genomic regions, located on SSC5 and SSC7. The region on SSC5 was associated with the difference between prick or intermediate ears and large, flat ears and the signals on SSC7 were associated primarily with the differences between prick- and intermediate-eared breeds. Our results from an introgression analysis also suggest that the SSC5 region of flat-eared breeds derives from Asian pigs.
The signatures of selection associated with ear morphology concurred with an earlier QTL study of the trait in pigs [38]. The SSC7 QTL of Wei et al [38] overlaps directly with the first differentiated region (31.82–34.19 Mb) on that chromosome. The suggestion that PPARD located on SSC7 plays a role in ear variation in pig breeds could not be supported as it was not positioned near either of the two signals of selection identified on this chromosome. However, as PPARD is involved in many biological processes and is located next to major QTLs for fat deposition and growth, its role in ear morphology warrants further investigation [39]. The QTL peak on SSC5 reported by Wei et al [38] is located approximately 10 Mb upstream of the peak FST signal but the confidence interval for the QTL location could overlap this position. Genome-wide association studies (GWAS) on ear morphology in dog breeds identified a region underlying this trait that was syntenic to the region on SSC5 in this study [19], [24]. Both these studies suggest MSRB3 and HGMA2 as candidate genes due to the proximity of the associated SNP. However, in the pig breeds the peak signal was located closer to LEMD3, which is involved in bone morphogenetic protein (BMP) signalling. Recently, a fine mapping study in pigs has suggested HMGA2 as a candidate locus for this QTL [40]. Mutations in the human version of this gene are associated with disorders involving increased bone density, suggesting a possible role in bone development [41]. However, analysis of coding sequences of these genes in this region of SSC5 for prick- and flat-eared pig breeds did not reveal any shared non-synonymous differences between the two groups, suggesting that changes in regulatory elements or miRNA genes may be responsible. Expression studies are required to test this hypothesis.
Like ear morphology, variation in coat colour patterns occurred post-domestication and signals of selection related to the traits indicate strong historic selection for the different phenotypes. Molecular studies have already identified the major coat colour loci in pigs, KIT and MC1R, for which allelic variation is associated with many of the coat colour variants (see references in [9], [17]). However, in this study signals of selection were not observed at or near MC1R (SSC6) for individual breeds that have an allele associated with a particular coat colour or when breeds were grouped by coat colour traits. The other locus, KIT (SSC8), is found ∼1 Mb downstream from a differentiated region shared by three breeds (British Saddleback, Hampshire, Pietrain). Several possible explanations could account for weak and absent signals of diversifying selection at KIT and MC1R, respectively. The differentiated region on SSC8 was quite extensive in genomic size and KIT may have been one of several targets of selection in that region, thus dampening any KIT-specific signals. Furthermore, allelic variation at both KIT and MC1R is associated with a large variety of coat colours and patterns for many breeds. With the breed set analysed in this study, there is no simple dichotomous division of the breeds based on coat type for these two genes, which could have weakened the power of this approach. Lastly, the inter-SNP distances in the MC1R region of SSC6 were particularly high (the distance between the flanking markers was in the 99th percentile of the genome-wide distribution of inter-SNP distances). Thus it appears that the MC1R region was not well covered by the PorcineSNP60 chip, which may explain why no signals of diversifying selection were detected there.
In contrast to the weak or absent signals of selection at the two major coat colour loci, KIT and MC1R, strong breed-specific signals of diversifying selection were observed near other coat colour loci. Two non-synonymous mutations were found in the endothelin receptor B (EDNRB) gene, in a region exhibiting substantial differentiation unique to Gloucestershire Old Spots. EDNRB encodes a G protein-coupled receptor that binds to the different isoforms of endothelins. The EDNRB-endothelin interaction plays a role in a range of critical physiological processes including the formation of enteric nerves and melanocytes (pigment-producing cells), both of which are neural crest derivatives [42], [43].
Mutations in EDNRB, leading to a reduced expression of the gene and partial or complete loss-of-function, have been shown to be associated with changes in pigmentation due to its role in melanocyte development [43], [44]. The piebald phenotype in mouse, characterised by white coat spotting [43], results from the insertion of a large retrotransposon in the first intron of EDNRB [45]. Several different mutations in humans are associated with a loss of pigmentation in the hair, skin and iris (Hirschsprung's disease/Waardenburg syndrome) [43] while a missense mutation gives rise to the Lethal White Foal Syndrome [46], where homozygous foals are completely white (and die early due to intestinal blockage) while heterozygous animals usually have distinctive white patches.
The mechanism(s) by which point mutations in EDNRB could be associated with (partial) loss of function is not yet known. The amino acid changes at residues 64 (Jinhua) and 68 (Gloucestershire Old Spots and Xiang) are both located in the N-terminal extracellular domain of the protein. One of the non-synonymous EDNRB mutations associated with Hirschsprung's disease is located in the same domain, at residue 57. This domain has been suggested to be important for stable ligand binding [47]–[49]. Furthermore, human EDNRB is believed to be cleaved by a metalloprotease at R64|S65 (R65|S66 in pig) and a truncated EDNRB (missing the first 64 residues) was found to be functional but had significantly reduced cell surface expression [50]. Using a program that predicts cleavage sites by membrane-type metalloproteases (SitePrediction, [51]), the reference pig EDNRB with S68 was, like its human homologue, found more likely to be cleaved at the R65|S66 site than the Gloucestershire Old Spots protein with F68 (unpublished results). The SNPs that alter residues 64 and/or 68 may result in an incomplete or uncleaved EDNRB and hence altered expression on the cell surface.
Black spotting in the Gloucestershire Old Spots has been previously associated with the EP allele at the melanocortin receptor 1 (MC1R locus): a 2-bp insertion in MC1R causes a frameshift mutation which results in a premature stop codon further downstream [15]. That study also demonstrated irregular somatic reversion to the black form of MC1R in two spotted breeds, Pietrain and Linderod, such that some regions of the body (black spots) expressed the form of the protein that enables black pigment production, whereas other (white) regions mainly expressed the mutated (non-functional) form of the protein. However, as breeds with various spotted and non-spotted patterns carry the 2-bp insertion, it is likely that additional loci also influence coat pattern and colour in these breeds. A recent paper demonstrated the complex interactions between melanocortin and endothelin signalling in determining coat patterns in cats [52] and similar interactions may also influence coat pattern diversity in pigs. We propose that the variant MC1R, resulting from the 2-bp insertion (and somatic reversion), may interact with partial loss of function in EDNRB such that only part of the body is populated by melanocytes which have the potential to revert and become pigmented. This in turn could give the Gloucestershire Old Spots its characteristic spotting pattern of relatively few and small spots compared to those observed, for example, in Pietrain. Functional analyses are required to characterize the effects of the Gloucestershire Old Spots variants on EDNRB function and on pigmentation patterns.
Although the variants at EDNRB were unique to the Gloucestershire Old Spots in the analysis of European breeds, they were shared by the Asian breed Xiang. We do not have phenotypic information for the Xiang individual who shares the Gloucestershire Old Spots variants but one of the most common Xiang subtypes is two-end black with a white middle body, akin to the familiar piebald mouse (http://www.viarural.com.pe/ganaderia/a-porcinos/exteriorcerdos/paises/china.htm). The Jinhua breed, which carries a proline to serine change at nearby residue 64 (Figure 4; [53]), has a similar phenotype. The difference in the phenotypes between the Asian breeds and Gloucestershire Old Spots is likely to be related to their different MC1R genotypes. The Asian breeds with EDNRB mutations do not carry the MC1R insertion (unpublished results), consistent with previous studies that show a low frequency or absence of this allele in Asian pigs [54], [55]. The two Gloucestershire Old Spots individuals are substantially more similar to the Asian breeds than the European ones in the EDNRB region. This finding, the shared EDNRB genotypes of Gloucestershire Old Spots and Xiang, and the introgression results described above together suggest an Asian origin for the Gloucestershire Old Spots mutations.
A putatively selected region identified in the Berkshire breed includes the KITLG locus and further sequence analysis revealed several non-synonymous variants in this breed. KITLG binds to the KIT receptor and plays a role in the melanocyte production pathway. Variation at the locus has been implicated in different skin pigmentation phenotypes in mice (i.e. steel mutant) [44], [56] and humans [57], [58], including hypo- and hyper-pigmentation, and has been investigated previously for its role in pig colouration [59]. The breed standard for Berkshire is a black animal with six white points (on the snout, tip of the tail and tips of each of the legs). The Berkshire was allegedly highly variable in coat colour until introgression of Asian genetic material and selection for breed homogeneity led to its contemporary coat pattern [5], [6]. Our tests using PorcineSNP60 data did not detect evidence of Asian introgression for Berkshire in the KITLG region (as assessed using comparisons with Meishan), although Berkshire shared the C1089T variant with Jiangquhai, another Asian breed, but not with any other European or Asian individuals. Furthermore, the two non-synonymous variants found in Berkshire were more common in the Asian than the European breeds. Similarly, Okumura and colleagues [37], [60] found evidence for an Asian origin of KITLG in Berkshire, as the breed shared haplotypes similar to Asian breeds at the locus whilst differing from other European breeds.
We identified the same two non-synonymous variants (A919G, G458A) in Berkshire and several Asian breeds as Okumura and colleagues [37], [60]. However, these variants cannot on their own explain the Berkshire phenotype because they were also found in three European individuals, including a Pietrain and a Tamworth (both homozygous), the latter breed which is red. Alternatively, the Berkshire phenotype might be attributed to differential regulation of KITLG, in conjunction with variation at other pigmentation genes (e.g. MC1R—Berkshire also carries the black spotting allele discussed above—and KIT). This could be related to the C1089T 3′-UTR variant that was only seen in Berkshire and Jiangquhai (also a black breed) or another regulatory element. Cis-regulatory differences in KITLG expression have been associated with pigmentation differences in stickleback fish [61] and a SNP located 350 Kb upstream of the KITLG gene was found to be associated with human hair colour, suggesting a possible regulatory role [62]. However, we were unable to search for variants in either proximal or distant enhancer/repressor elements due to errors in this region of the current pig genome assembly.
Signatures of diversifying selection were found that may be associated with important pig production traits. Teat number is an important reproductive trait because with increased litter size, which is often selected for in pig breeds, a sufficient number of teats are required to support the litter [13]. Although the FST teat-trait analysis results had some ambiguity, the signal on SSC12 seen in the 14 vs 12 teats comparison but not the ‘control’ comparison (breeds with 14 teats compared with one another and breeds with 12 teats compared with one another) overlapped with documented QTL. Both Hirooka et al [63] and Rodriguez et al [64] reported a significant QTL for teat number on this chromosome, with the latter study suggesting that the most likely position of the QTL was between markers SW874 (23.67 Mb) and SW1956 (40.77 Mb), which overlapped with the region of high differentiation observed in the current study. The NME1 gene, which is found in this region (27.46–27.50 Mb), plays a role in mammary gland development. NME1-deficient mice, although they reproduce normally, have delayed mammary gland development [65] and incomplete maturation of the lactiferous duct in the nipple [66].
Amongst the production characteristics that commercial pig breeds share, they also possess breed-specific characteristics. Duroc pigs are known for their high intramuscular fat content (IMF) in comparison to other commercial pig breeds [67] and for their higher concentrations of saturated and mono-unsaturated fatty acids (and lower concentrations of poly-unsaturated fatty acids) [68], characteristics that play key roles in meat quality. Uemoto et al [27] found a significant QTL for fatty acid composition in Duroc on SSC14 that has not been reported for other breeds. This QTL region overlaps with an extreme differentiation region observed only in the Duroc breed and contains ELOVL3, a gene involved in the synthesis of fatty acids; in mice a lack of ELOVL3 resulted in decreased levels of certain fatty acids due to an inability to convert saturated fatty acyl-CoAs into very long chain fatty acids [28]. In addition, SCD (stearoyl-CoA desaturase), a gene located close to the peak differentiation region, encodes a key enzyme in the synthesis of fatty acids and has thus been proposed as a candidate gene for the fatty acid composition QTL [27].
Landrace also exhibited high levels of differentiation, in this case in an extended region of SSC13. The peak differentiation values were found close to the grehlin (GHRL) gene, which is a candidate for associations with appetite and feeding behaviour. The regulation of voluntary food intake is controlled by a biological cascade of chemical signals that controls appetite and satiation, where various hormones are involved in the starting and/or termination of an eating episode. Grehlin has been specifically proposed in prompting hunger feelings and therefore initiating eating [69]. Its involvement in regulating feeding behaviour in pigs has only recently been considered [70].
By comparing pig breeds with their ancestral species, the wild boar, we sought to identify genomic regions and genes that could be involved in the domestication process. The largest differentiated genomic region between the domestic pig breeds and wild boar was observed on SSC7. Numerous QTLs have previously been mapped to this chromosome for traits such as growth, carcass length, skeletal morphology and backfat depth using several types of crosses [11], [12]. Several genes located in the region of differentiation have been investigated for possible physiological roles: PPARD and CDKN1A have been considered candidates for fat deposition [71] and, as mentioned above, PPARD has also been considered a candidate gene for ear structure variation [39]. In addition, the genomic signal of selection is close to the MHC region, a complex that is crucial in vertebrate immunity, making it a potential source of evolutionary change on the chromosome. The large differentiated region on SSC7 may reflect strong diversifying selection in domestic pig breeds as this chromosome appears to influence many pig production traits.
Domestic pig breeds are also different from wild boar in skeletal morphology. Substantial changes have occurred in the body and cranial dimensions following domestication [72]. In the comparison of pig breeds with wild boar, a region of genetic differentiation identified on SSC1 is syntenic to a region associated with cranial dimensions in dog breeds [32]. The cranial trait under investigation in the dog studies, brachycephaly, is characterised by a strong alteration of the facial bone structure through shortening of the muzzle and shortening and widening of the skull [31]. Pig breeds possess variable skull morphology ranging from a long snout (Tamworth) to shorter wider faces (Berkshire, Gloucestershire Old Spots, Large Black) to very short faces with upturned snouts, similar to brachycephaly in dogs (Middle White) (see Figure S1). However, Middle White, the most brachycephalic-like breed, did not show significant differentiation from wild boar in the SSC1 region. Incidentally, it has been suggested that Middle White acquired its ‘dished’ face from Asian pigs [6]. However, there was no evidence of Asian introgression into the Middle White in the regions orthologous to the dog brachycephaly regions, suggesting that if it did indeed acquire its squashed face from Asian pigs, there has been independent evolution for this trait in dogs and pigs. As various skeletal and cranial changes occurred after domestication of the wild boar [72], the region of high differentiation overlapping the brachycephaly region in dogs could be associated with other bone alterations.
The putative genomic signatures of selection for breed-defining phenotypic traits and levels of breed genetic differentiation reflect the historical development of the pig breeds. The Duroc had the strongest signals of diversifying selection, evidenced by the levels of genomic differentiation, which were observed to be unique to this breed and unlike the other breeds, no signals of diversifying selection were observed on SSC8 for the Duroc, indicating that this breed may have a distinct genetic origin, as previously noted from microsatellite and sequence data [35], [73]. Some of the clearest signals of both diversifying selection and introgression from Asian pigs were associated with highly visible phenotypes such as coat pattern and ear morphology, suggesting that these traits have been under particularly strong selection during the development of European pig breeds. In particular, selection associated with flat ears was detected in breeds that do not appear to share recent ancestry [7], [73], which may reflect convergent evolution through independent selection for that trait. In contrast, although microsatellite markers indicate a common ancestry for Berkshire and Gloucestershire Old Spots [7], [73], shared differentiation signals were not seen, illustrating differing breed development trajectories. Signatures of selection were also observed in regions associated with certain quantitative traits in pig production, but there was a paucity of signals at loci associated with those related to reproduction. The lack of differentiation signals associated with such traits may reflect their control by many genes of small effect, as suggested by Boyko and colleagues [19].
The genomic regions identified in this study using the genetic differentiation approach generally did not overlap with those identified in a scan for extreme homozygosity in European pig breeds: none of the regions identified in five or more breeds overlapped with the regions reported by Rubin and colleagues [25] and only two out of 109 regions identified in individual breeds overlapped (SSC1:172.13 Mb and SSC15:115.17–115.77 Mb). The Rubin study used more dense genomic data so it is possible that the Porcine SNP60 chip did not contain variants close to the regions they identified. However, in our study we have detected what appear to be genuine signals of selection in pig breed development. Another explanation for the lack of overlap between the studies is that, by pooling genomic data across several breeds, Rubin and colleagues [25] identified regions of homozygosity that were shared amongst the breeds, arguably picking out candidates more likely to be involved in the domestication process and early, post-domestication pig development. In contrast, our methodological approach searched for between-breed differences, thus revealing candidates arising from diversifying selection that occurred during breed development.
DNA samples were obtained from blood samples collected by veterinarians according to national legislation, from tissue samples from animals obtained from the slaughterhouse or, in the case of wild boar, from animals culled within wildlife management programs.
DNA samples were obtained from blood samples collected by veterinarians according to national legislation, from tissue samples from animals obtained from the slaughterhouse or in the case of wild boar, from animals culled within wildlife management programs. Samples for SNP genotyping were obtained from between 24 and 34 individuals for 14 pig breeds, described in Table 1, and were genotyped using the PorcineSNP60 chip assay [74]. Most breed samples (including the Asian breed, Meishan) were from the PigBioDiv study whereby a maximum of two individuals were sampled from a litter from as many herds as possible, so as to have as few related individuals as possible in the sample set [75]. For the four commercial breeds (Duroc, Landrace, Large White and Pietrain), the data was from individual commercial populations, which were found to be good representatives of the breeds based on clustering analysis of multiple populations (unpublished results). Welsh samples were provided by the Pedigree Welsh Pig Society. Wild boar samples were those used in the original SNP discovery procedure [74]. Genotype data are deposited in the Dryad repository (http://dx.doi.org/10.5061/dryad.c2124).
All analyses were carried out in R ([76], http://www.r-project.org/). A series of quality control measures were applied to the dataset to filter out any possible genotyping anomalies. First, SNP markers that had greater than 10% missing genotypes were discarded. Second, markers that were monomorphic across all the breeds (i.e. MAF<0.01) were also discarded from further analysis. Third, SNP markers were tested for deviations from Hardy-Weinberg equilibrium within each breed using an exact test [77]. At a critical rejection region of 8.33×10−7 (0.05/60,000) a total of 66 SNPs did not conform to HWE expectations in one or more breeds and were removed from the analysis. Of these, 46 deviated from HWE due to excess of heterozygote genotypes in one or more breeds. The other 20 SNPs deviated from HWE due to heterozygote deficit in one or more breeds. Fourth, markers that were not mapped to the porcine genome were removed, based on the current pig genome assembly, Sus scrofa (SSC) Build 10.2. For the remaining markers, SNPs that were not yet mapped to a specific location on a specific chromosome of the pig genome were also filtered out. Following quality control, 49 260 markers were considered for the majority of analyses (see below for one exception). After QC, average individual genotype coverage was 99.20% across all breeds and average individual genotype coverage in individual breeds ranged from 96.09% in the Mangalica breed to 99.96% in the Hampshire breed.
Pairwise Wright's FST [78], the classical measure of population genetic differentiation, was used to detect signatures of diversifying selection. We previously showed [79] that pairwise measures of differentiation were better at identifying markers that distinguished breeds than global measures and that Wright's estimate of FST was highly correlated to that of Weir & Cockerham's [80]. The use of population (breed) differentiation to identify candidate selected regions, as implemented in the current study, was originally suggested by Akey and colleagues [81]. This approach was justified by use of simulations in a follow-up study on dogs [18] and has subsequently been implemented in various empirical studies [22], [24], [82].
The PorcineSNP60 chip assay was designed to include SNPs evenly distributed across the genome, with per-chromosome average inter-SNP distances ranging from 30 to 40 kb (except for SSCX) (based on builds 7 and 8) [74], with a median of 30 kb for the genome-wide distribution. Across the genome, the majority (80%) of inter-SNP distances were less than 70 kb in this study. Recent studies (e.g. Ref. [83]) show high linkage disequilibrium across commercial pig genomes (r2∼0.4 between adjacent SNPs on the PorcineSNP60 chip), suggesting that our study is likely to detect most signals. To account for stochasticity in locus-by-locus variation, for all of the FST analyses a 13-SNP sliding window was implemented on the estimated values, with the mid-SNP determining the genomic location of the window (hereafter designated as FST-window). To allow 13-SNP sliding windows across a whole chromosome, the first window on a chromosome was centred at the 7th SNP position and the last window on a chromosome was centred at the 7th from last SNP position. Candidate selected regions were defined as the 99th percentile of the empirical distributions of FST-windows, except where indicated otherwise.
A breed average FST was first calculated. FST was estimated between pairs of European breeds at each SNP marker using the breed allele frequencies. For each breed this produced 12 breed-pairwise FST comparisons at each SNP marker. The FST at each SNP marker for all of these pairwise comparisons were averaged to produce an overall FST for each SNP marker in each breed (here after designated as FST-SNP).
The FST analysis was extended to compare groups with different phenotypic traits. For each trait classes were formed, based on the observed phenotypic variation between breeds (see below), and breeds were placed into one of the classes. For each trait, FST was estimated between each breed in one class compared against each breed in the next class and averaged across the pairwise comparisons to obtain a FST-SNP estimate. A summary table of the different traits, the phenotypic classes and the class designation of each breed is shown in Table S2.
Ear morphology in European pigs is variable, ranging from upright or prick ears that may be slightly inclined forwards (the ancestral state as seen in wild boar), to a medium sized ear that points forwards and downwards but is not too heavy, to a completely dropped ear that is long, thin and lies relatively flat against the face slightly curbing vision of the animal (see Figure S2). Ear morphology was grouped into the following classes: prick-eared breeds, intermediate-eared breeds and flat-eared breeds.
Coat colour in European pigs is a highly variable phenotypic trait including from black, red, brown and white, with and without spots and belts. The coat colour was grouped into the following classes: red coat breeds compared with non-red coat breeds; saddleback breeds compared with non-saddleback breeds; white coat breeds compared with non-white coat breeds; red coat breeds compared with black coat breeds.
Amongst the breed standard requirements set by the British Pig Association (BPA), the number of teats is one listed criterion. Using this trait, breeds were grouped in the following classes: breeds where the BPA standards required a minimum of 14 displayed teats compared with breeds where the BPA standards required a minimum of 12 displayed teats, Berkshire and Middle White were removed from this trait comparison because there was not a definitive breed standard requirement (breed standards suggested a “minimum of 12 but preferably 14 teats”) and Mangalica was also removed because the breed standard number of teats was unknown.
Levels of genetic differentiation between the domestic pig breeds and wild boar were estimated. The SNPs that were monomorphic in the pig breeds were compared with wild boar genotypes to determine if some were segregating in the wild boar. The (mapped) breed-monomorphic SNPs that were segregating in the wild boar were added to the set of polymorphic SNPs described above, giving a total of 49 556 markers. FST was estimated between wild boar and each pig breed, which produced 13 pairwise comparisons at each SNP marker. The FST at each SNP marker for each of these pairwise comparisons were averaged to produce an overall FST for each SNP marker (here after designated as FST-SNP).
Two methods were employed to infer signals of Asian introgression in European breeds. First, an FST analysis, as described above, was used to quantify differentiation between the Asian Meishan breed and each of the 13 European breeds. Regions of particularly low differentiation (below 1st percentile) were interpreted as showing evidence of Asian introgression.
Second, a Bayesian analysis was performed using the site-by-site linkage model in STRUCTURE software [84]. This model was designed to infer the ‘population-of-origin’ assignment of genomic regions and has been used to determine levels of introgression between populations (e.g. Ref. [85]). Each of the 13 European breeds was compared with the Meishan breed, using no a priori population information: at a pre-defined number of clusters, K = 2, the linkage model was run five times for 20,000 iterations after a burn-in of 40,000 iterations (which included 20,000 iterations with the admixture model). Due to computer memory limitations, for the analysis 15 individuals per breed (approximately half of the total dataset) were chosen at random and every second marker across each chromosome was removed from the input data set leaving a total of 24 630 markers.
Ancestry proportions across the two clusters (“Asian” and “European”) were estimated for each of the European individuals. Estimates of Asian ancestry for each European animal for each SNP were obtained from the probability of assignment to the Asian cluster and then averaged across the individuals within each breed. As described above, a sliding window average of Asian ancestry values across each chromosome was calculated, with windows composed of 7 SNPs (half the number used for the analyses of the full set of SNPs). The average value for the window was assigned to the position of the central SNP. These values were interpreted as probabilities of introgression from Asian to European breeds.
In order to identify genomic regions with clear signals of Asian introgression, we identified SNP positions (to the closest Mb) that met two criteria: (1) values below the 1st percentile of the Meishan-European breed FST-windows distribution and (2) found in the 99th percentile of STRUCTURE-calculated introgression probabilities for that breed.
DNA samples for sequencing were obtained as described above for SNP genotyping. Individual samples (52) from 12 of the European breeds analysed above (no Welsh pigs were included) as well as 24 samples from eight Asian breeds (Table S13) were sequenced using the Illumina HiSeq2000 platform, with library preparation and sequence generation per manufacturers protocols. Sequence mapping and variant calling were carried out as described previously [25], [34]. Briefly, Illumina (v. 1.3–1.8) formatted fastq files, with sequence reads of 100 bp were subject to quality trimming prior to sequence alignment. The trimming strategy involved a 3 bp sliding window, running from 5′ to 3′, with sequence data upstream being discarded if the 3 bp window average quality dropped below 13 (i.e. average error probability equal to 0.05). Only sequences of 45 bp or more in length were retained. In addition, sequences with mates <45 bp after trimming were also discarded. During trimming, quality scores were re-coded to follow the Sanger fastq format to standardize downstream processing.
Sequences were aligned against the Sscrofa10.2 reference genome using Mosaik 1.1.0017. Alignment was performed using a hash-size of 15, with a maximum of 10 matches retained, and 7% maximum mismatch score, for all pig populations and outgroup species. Alignment files were then sorted using the MosaikSort function, which entails removing ambiguously mapped reads that are either orphaned or fall outside a computed insert-size distribution. Alignment archives were converted to BAM format using the Mosaiktext function. Manipulations of BAM files, such as merging of alignments archives pertaining the same individual, were conducted using SAMtools v. 1.12a [86].
Variant allele calling was performed per individual using the pileup function in SAMtools, and variations were initially filtered to have minimum quality of 50 for indels, and 20 for SNPs. In addition, all variants showing higher than 3x the average read density, estimated from the number of raw sequence reads, were also discarded to remove false positive variant calling originating from off-site mapping as much as possible. Heterozygous variants and those with minimal SNP/indel qualities were further inspected manually to ensure that they were true variants.
We examined the sequence variation in three genomic regions that showed extreme differentiation in one or more breeds (Table S1) for the individuals from the 12 European breeds: (1) SSC5:31.0–34.0, (2) SSC5:98.0–99.0 and (3) SSC11:53.5–55.5 Mb. Information for the relevant regions was excised from the BAM files using SAMtools v. 1.12a [86]. Alignment files and variants called in these regions for all animals considered in this manuscript are deposited in the Dryad repository (http://dx.doi.org/10.5061/dryad.c2124). For the first region, we identified all variants that were shared by the individuals from flat-eared breeds but differed from all individuals from the prick-eared breeds (Table S2); for the second region, we identified all variants that were shared by the two Berkshire individuals but differed from the other individuals; and for the third region, we identified all variants that were shared by the two Gloucestershire Old Spots individuals but differed from the other individuals. Data for the individuals from Asian breeds was then used to examine specific sequence variants, as described in the Results.
|
10.1371/journal.pntd.0000407 | Phenotypic and Functional Characterization of Human Memory T Cell Responses to Burkholderia pseudomallei | Infection with the Gram-negative bacterium Burkholderia pseudomallei is an important cause of community-acquired lethal sepsis in endemic regions in southeast Asia and northern Australia and is increasingly reported in other tropical areas. In animal models, production of interferon-gamma (IFN-γ) is critical for resistance, but in humans the characteristics of IFN-γ production and the bacterial antigens that are recognized by the cell-mediated immune response have not been defined.
Peripheral blood from 133 healthy individuals who lived in the endemic area and had no history of melioidosis, 60 patients who had recovered from melioidosis, and 31 other patient control subjects were stimulated by whole bacteria or purified bacterial proteins in vitro, and IFN-γ responses were analyzed by ELISPOT and flow cytometry.
B. pseudomallei was a potent activator of human peripheral blood NK cells for innate production of IFN-γ. In addition, healthy individuals with serological evidence of exposure to B. pseudomallei and patients recovered from active melioidosis developed CD4+ (and CD8+) T cells that recognized whole bacteria and purified proteins LolC, OppA, and PotF, members of the B. pseudomallei ABC transporter family. This response was primarily mediated by terminally differentiated T cells of the effector–memory (TEMRA) phenotype and correlated with the titer of anti-B. pseudomallei antibodies in the serum.
Individuals living in a melioidosis-endemic region show clear evidence of T cell priming for the ability to make IFN-γ that correlates with their serological status. The ability to detect T cell responses to defined B. pseudomallei proteins in large numbers of individuals now provides the opportunity to screen candidate antigens for inclusion in protein or polysaccharide–conjugate subunit vaccines against this important but neglected disease.
| The Gram-negative bacterium, Burkholderia pseudomallei, is a public health problem in southeast Asia and northern Australia and a Centers for Disease Control and Prevention listed Category B potential bioterrorism agent. It is the causative agent of melioidosis, and clinical manifestations vary from acute sepsis to chronic localized and latent infection, which can reactivate decades later. B. pseudomallei is the major cause of community-acquired pneumonia and septicemia in northeast Thailand. In spite of the medical importance of B. pseudomallei, little is known about the mechanisms of pathogenicity and the immunological pathways of host defense. There is no available vaccine, and the mortality rate in acute cases can exceed 40% with 10–15% of survivors relapsing or being reinfected despite prolonged and complete treatments. In this article, we describe cell-mediated immune responses to B. pseudomallei in humans living in northeast Thailand and demonstrate clear evidence of T cell priming in healthy seropositive individuals and patients who recovered from melioidosis. This is the most detailed study yet performed on the cell types that produce interferon-gamma to B. pseudomallei in humans and the antigens that they recognize and the first to study large sample numbers in the primary endemic focus of melioidosis in the world.
| Melioidosis is a serious infectious disease in Southeast Asia and Northern Australia caused by the soil-dwelling Gram-negative bacterium, Burkholderia pseudomallei [1]. In Northeast Thailand, the mortality rate for acute melioidosis remains high, approximately 50%, despite recent advances in antibiotic treatments. Serological evidence, based on the indirect hemagglutination assay (IHA), suggests that 80% of people living in endemic areas have been exposed to B. pseudomallei, without showing clinical symptoms [1]–[3]. Recurrent melioidosis can also occur either as relapse after antibiotic treatment or re-infection [3],[4]. B. pseudomallei is classified as a NIAID category B potential agent for biological terrorism [5]. The mechanism that enables the organism to avoid the bactericidal effects of the host immune response has never been fully understood, and there are no licensed vaccines.
B. pseudomallei is able to disseminate throughout the body, invades non-phagocytic cells and replicates in phagocytes [6],[7]. In mice, B. pseudomallei is a potent inducer of IFN-γ and IFN-γ inducing cytokines such as IL-12, IL-18 and TNF in vitro and IFN-γ is essential for resistance in vivo via the activation of macrophages for both oxygen dependent and independent killing mechanisms [8]. In mice, NK cells and bystander CD8+ T cells provide innate production of IFN-γ [9], while IFN-γ secreting, antigen-specific CD4+ T cells contribute to protection against primary infection with B. pseudomallei and following immunization with experimental vaccines in vivo [10],[11]. In addition, murine models of vaccination with dendritic cells pulsed with heat killed B. pseudomallei in the presence of CpG oligodeoxynucleotides showed significant levels of protection [12] suggesting the role of specific T cells in host protection.
In contrast, the mechanisms of cell-mediated immunity to B. pseudomallei in humans are poorly understood. IFN-γ, IL-12, IL-18 and TNF are found in plasma samples from acute, septic melioidosis but the IFN-γ producing cells have not been well characterized [13]. Previous studies in small numbers of patients in northern Australia and Papua New Guinea who recovered from melioidosis have demonstrated evidence of T cell priming to B. pseudomallei, but the characteristics of the responding cell populations and the antigens recognized have not been defined [14],[15].
Here, we analyzed a large cohort of individuals from the melioidosis endemic region of Thailand to identify the cellular sources of IFN-γ in response to whole B. pseudomallei and the bacterial ABC transporter proteins LolC, OppA and PotF which are T cell immunogens in mice and candidate vaccine antigens [16],[17]. Peripheral blood cells from healthy individuals with serological evidence of exposure to B. pseudomallei and recovered melioidosis patients (but not seronegative control subjects) showed evidence of CD4 and CD8 T cell priming to both whole bacteria and purified B. pseudomallei antigens. Together with a prominent IFN-γ response from NK cells, these sources of IFN-γ may contribute to host resistance against melioidosis in the endemic setting.
The study and the consent forms were approved by the Khon Kaen University Ethics Committee for Human Research (Project number HE470506). Informed consent was obtained from all the subjects recruited into the study.
Blood samples from 133 healthy donors who had no clinical history of melioidosis were collected at the Blood Bank, Khon Kaen University, Thailand. Another set of blood samples was obtained from patients and control subjects at Sappasithiprasong Hospital, Thailand for cellular studies by ELISPOT assay. Patients were defined as those who had recovered from melioidosis (previously diagnosed by isolation of B. pseudomallei from blood or tissues) and completed antibiotic treatment (n = 36). Non infected control subjects (n = 21) were those who attended the hospital for non infectious reasons at the diabetic clinic and had no history of clinical melioidosis, and were matched for age, sex, occupation, the presence of diabetes as an underlying condition and lived in the same endemic area. In addition, 24 recovered melioidosis patients and 10 healthy control subjects were enrolled, using the same criteria, at Srinagarind Hospital, Thailand for cellular sources of IFN-γ, kinetics and memory cells assayed by flow cytometry. The subjects who had antibodies to B. pseudomallei at a titer of 1∶40 or greater by IHA were considered seropositive [15],[18]. None of the subjects had any clinical sign or symptoms of any infection including HIV/AIDS at the time of blood collection.
B. pseudomallei strain K96243 is a clinical isolate from Thailand and is the prototype genome sequence strain [19]. Whole heat-killed B. pseudomallei (hkBp) was prepared by heating the bacteria at 100°C for 20 minutes, washed twice with PBS pH 7.4, aliquoted and stored at −80°C. The number of viable bacteria was determined by colony-forming counts and defined as colony-forming units (CFU) prior to heating. Recombinant B. pseudomallei ABC transporter proteins (LolC, OppA and PotF) were prepared as previously described [16],[17] and used as test stimulators in this study. Phytohemagglutinin (PHA) (Biochrom AG, Germany), human recombinant IL-12, and IL-15 (BD Biosciences, USA) and a MHC class I-restricted T cell epitope control of pooled peptides of cytomegalovirus, Epstein Barr virus and influenza virus (CEF) were used as positive controls (Mabtech, AB, Sweden). Recombinant protein from Francisella tularensis, FT1823 [20] was included as a non related protein/negative control.
Peripheral blood mononuclear cells (PBMCs) from each subject were isolated from heparinized blood samples by density centrifugation on Ficoll-Hypaque and adjusted the number of cells as required prior to stimulation. In brief, 96-well PVDF-plates (MSIP, Millipore) were previously coated overnight with 15 µg/ml 1D1K anti-human IFN-γ at 4°C. Fresh PBMCs were added in duplicate wells at 5×105 PBMCs/well and each stimulator was added at the optimal concentration. After 42 hours, secreted IFN-γ was detected by adding 1 µg/ml biotinylated mAb 7-B6-1-biotin for IFN-γ for 3 hours and followed by 1 µg/ml streptavidin-alkaline phosphatase (Mabtech, AB, Sweden) prior to enumeration under a stereomicroscope. The responses were compared in the absence or presence of 0.3 µg/ml cyclosporin A (CsA, Sigma, USA).
Whole blood samples were firstly analyzed for complete blood count using an automatic machine (Sysmex, Germany). The number of absolute lymphocytes was then adjusted to 9×105 lymphocytes/ml by diluting with completed RPMI medium (10% FBS supplement). The adjusted cells in 100 µl were added into 96 well culture plates and added up by another 100 µl of 2× concentration of stimulators and incubated at 37°C with 5% CO2. Cultured cells were blocked with 10 µg/ml brefeldin A (Sigma, USA) for 3 hours prior to the end of the incubation time. Then washed and blocked with anti-CD16 (BD Biosciences). The following antibodies were used for immune cell surface staining: FITC anti-CD4, PE anti-CD8 or PE anti-CD56 (BD Biosciences) and Tricolor anti-CD3 (Invitrogen, USA). In addition, cell surface markers for memory T cell phenotypes were included: FITC anti-CCR7 (R&D systems, USA), PE anti-CD45RA (Invitrogen) and Tricolor anti-CD4 or CD8 (Invitrogen). Isotype-matched control antibodies were used in each analysis. After 30 min of staining, followed by fixation with 10% paraformaldehyde-PBS overnight at room temperature, cells were then permeabilized by 0.12% saponin (Sigma, USA) for 15 min followed by APC anti-IFN-γ (Invitrogen, USA) for 30 min prior to analysis by FACScalibur with CELLQuest software (BD Biosciences, USA).
Statistical analysis (one way-ANOVA, unpaired and paired t-test) was performed using Graphpad Prism version 5 software (GraphPad, San Diego, CA, USA). A P-value<0.05 was considered statistically significant.
To examine the cellular immune response to B. pseudomallei of healthy individuals living in Northeast Thailand, PBMCs of 133 donors from the Blood Bank at Khon Kaen University, Thailand were cultured with whole bacteria, recombinant B. pseudomallei ABC transporter proteins (LolC, OppA and PotF) or control stimulators and 42 hours later assayed for IFN-γ production by ELISPOT. We have previously shown in mice that several different cell types contribute to IFN-γ responses to B. pseudomallei in vitro; NK cells and bystander T cells produce IFN-γ indirectly via a cytokine mediated pathway which is not blocked by cyclosporin A (CsA), whereas specific B. pseudomallei primed T cells respond via a CsA sensitive T cell receptor (TCR) dependent process [9],[11]. To validate this approach in human peripheral blood, we initially compared the CsA sensitivity of cytokine (IL-12+IL-15), mitogen (PHA) or antigen specific IFN-γ responses in vitro in the presence or absence of CsA. Compared to medium alone controls, cells incubated with PHA or a pooled cocktail of established T cell reactive peptides from pathogens known to be present in the Thai population (CMV, EBV and influenza) showed strong IFN-γ responses which were inhibited in the presence of CsA (Figure 1A; P<0.0001, paired t-test). In contrast, the IFN-γ response to IL-12/IL-15 or the low but detectable background response observed in cells incubated with an irrelevant Francisella tularensis control protein were not CsA susceptible. Moreover, the results revealed that whole B. pseudomallei (hkBp) and three Bp-derived ABC transporter proteins (LolC, OppA and PotF) could induce IFN-γ responses via TCR independent (innate) and dependent (specific) pathways in healthy individuals in vitro (Figure 1B).These IFN-γ responses were in a dose dependent manner ranging between 1×104–1×107 CFU/ml hkBp and 0.1–3.0 µg/ml of the 3 proteins (data not shown).
IHA has been widely used as routine serologic test for melioidosis with the threshold titer of 1∶40 in the endemic area indicating previous exposure to B. pseudomallei [15],[18]. To investigate whether the magnitude of the cellular immune response correlated with evidence of exposure to B. pseudomallei by serology, 133 healthy donors were classified into five groups based on their B. pseudomallei antibody IHA titers (as 1∶20, 1∶40, 1∶180, 1∶160 and 1∶320 (n = 6, 19, 60, 45 and 3, respectively; Figure 2). The results revealed that the continual increase of the average values of specific (CsA sensitive) IFN-γ spots in response to B. pseudomallei and its proteins was significantly correlated with increasing antibody titers (P<0.0001, one way ANOVA). No such correlation was observed in the response to CEF vs. medium controls or for the innate (CsA resistant) IFN-γ spots to B. pseudomallei (P>0.05, one way ANOVA). Thus environmental exposure to B. pseudomallei in the endemic region of NE Thailand generates both T cell and B cell responses to B. pseudomallei in healthy individuals even in the absence of disease.
To assess the extent of T cell priming in patients who had survived active infection, specific T cell responses to whole bacteria and recombinant proteins of B. pseudomallei were studied in 36 recovered melioidosis cases and 21 other patient control subjects from the same endemic region chosen on the basis of same age, sex and occupation with no history of clinical melioidosis but who were seropositive for B. pseudomallei exposure. The frequency of IFN-γ producing cells was significantly increased in recovered patients compared to seronegative control subjects (data not shown) but was similar to that observed in seropositive healthy individuals (Figure 3A) (P>0.05, unpaired t-test). However, there were some individuals who had no specific IFN-γ producing cells to B. pseudomallei above the background of medium control in both groups. Of note, IFN-γ levels as quantified by ELISA were significantly higher in recovered melioidosis patients than seropositive individuals (Figure S1). According to Figure 3B, the comparison of IFN-γ responses between patients who recovered from melioidosis with a history of localized infection (n = 13) and severe sepsis (n = 11) did not show any significant difference. These specific T cell responses declined over the time but remained detectable after 80 weeks (Figure 3C). These results indicated that either whole B. pseudomallei or its proteins could trigger the cellular immune response following re-exposure to the microorganism in vitro up to 80 weeks post admission.
Diabetes mellitus (DM) is a major risk factor for human melioidosis [21], and only 4 cases of recovered melioidosis without DM were found in this study. Even though we observed no difference between IFN-γ responses of these groups, it remains inconclusive for the effect of diabetic condition on host T cell responses (data not shown). In addition, recovered melioidosis patients with a history of recurrent infection (n = 6) compared to those with a single disease episode (n = 30) also showed no statistically significant difference (data not shown); suggesting that under these conditions IFN-γ responses do not differentiate between primary and recurrent melioidosis.
To identify the cellular sources of IFN-γ responses to B. pseudomallei, whole blood samples from six seropositive healthy individuals and ten recovered melioidosis cases (all with IHA antibodies 1∶40 or greater) were restimulated with B. pseudomallei in the absence of CsA and analyzed by four-color flow cytometry. As shown from one representative of seropositive group, the small lymphocyte area was gated (Figure 4A) and analysis of IFN-γ+ cells showed that NK cells (CD3−CD56+), CD4+ T (CD3+CD4+) and CD8+ T (CD3+CD8+) cells all contributed to IFN-γ production to hkBp (Figure 4B). A dominant contribution of CD4+ and CD8+ T cells on IFN-γ production was confirmed by significant reduction of specific IFN-γ (CsA sensitive) spots following depletion of CD3, CD4 and/or CD8 cells (Figure S2). In addition, the mean fluorescent intensities (MFI) of intracellular IFN-gamma staining of CD3+ and CD3− (NK) cells of 4 recovered melioidosis cases were analyzed and revealed that the average MFI of IFN-gamma gated on CD3+ cells was significantly higher than CD3− (NK) cells (P<0.05, paired t-test) (Figure S3). Together with our finding that the ELISPOT size of remaining IFN-γ+ cells after T cell depletion was very small suggests that innate (CsA resistant) cells produce less of this cytokine compared to specific T cells.
The analysis of blood samples from 6 seropositive individuals and 10 recovered melioidosis cases showed background staining of total IFN-γ producing cells in medium alone at 0.03±0.01 and 0.3±0.09% (mean±SE), respectively. The relative contribution of each cell type to the IFN-γ response to B. pseudomallei also varied according to the time point examined in culture after addition of the bacteria. NK cells appeared to respond more rapidly than T cells and significantly contributed to the production of rapid IFN-γ at 4 hours and decreased over time in both groups (Figure 5). On the one hand, the frequency of IFN-γ producing NK cells were significantly higher in seropositive healthy individuals than recovered melioidosis at 4 and 12 hours (P<0.05, unpaired t-test). On the other hand, there was a statistically significant difference of IFN-γ producing CD4+ T cells being higher in recovered melioidosis than the seropositive individuals at both time points and all three time points for IFN-γ producing CD8+ T cells. These results demonstrated the increasing contribution to IFN-γ production over the time from T cell subsets, particularly in the recovered melioidosis group.
To investigate whether the rapid IFN-γ producing T cells in response to B. pseudomallei were memory T-cell phenotypes, immune subsets of human memory T cells were identified based on the cell surface expression of CD45RA and CCR7 [22] in 16 recovered melioidosis cases and 7 seropositive control subjects. The results demonstrated the percentages of IFN-γ producing memory T cells and the majority of memory CD4+ and CD8+ T cells of both groups were revealed as terminally differentiated T effector memory (TEMRA) cells which was significantly higher than the other memory phenotypes of effector memory (TEM) and central memory (TCM) cells of CD4+ and CD8+ T cells (P<0.0001, unpaired t-test) (Figure 6A).
When the clinical histories of these 16 recovered melioidosis subjects were analyzed, distinctive patterns of memory T cell phenotypes were revealed. The memory T cells of the septicemic melioidosis group (n = 12) were TEMRA significantly greater than TEM and TCM of both CD4+ and CD8+ subsets (P<0.0001, unpaired t-test). Interestingly, there was a trend of localized melioidosis group (n = 4) showed stronger responses of TEMRA with small contribution of TCM cells and TEM cells, but it was not statistically significant (Figure 6B).
Burkholderia pseudomallei is an important cause of community acquired sepsis and death in endemic regions of SE Asia and Northern Australia and is listed as potential bioterrorism threat. Yet despite its current and potential impact on public health our understanding of immune defenses against this pathogen are incomplete. B. pseudomallei is capable of extensive extracellular growth and abscess formation, but is also genetically adapted to survive and replicate within host cells [6],[23]. It is killed by IFN-γ activated macrophages in vitro [24], making cell mediated immunity a potentially important component of resistance. Here, a total of 224 individuals living in the endemic area of NE Thailand of varying immunological and clinical history for exposure to B. pseudomallei were examined for the magnitude and characteristics of their IFN-γ responses following restimulation of whole blood with whole bacteria or B. pseudomallei derived antigens in vitro.
Northeast Thailand is the primary endemic focus of melioidosis in SE Asia and the majority of individuals show evidence of seroconversion from an early age [2]. To obtain an initial estimate of the diversity of the IFN-γ response in this setting, blood samples from 133 randomly selected individuals who had no clinical history of melioidosis were tested for reactivity to B. pseudomallei by IFN-γ ELISPOT. The majority showed clear induction of IFN-γ positive cells above that of medium alone controls in the presence of whole bacteria. Addition of cyclosporin A (CsA) which specifically inhibits T cell receptor-mediated but not cytokine mediated lymphocyte activation [9],[11],[25] showed this response was made up of both innate and adaptive IFN-γ responses. To further define the adaptive IFN-γ component, we compared the frequency of CsA sensitive IFN-γ producing cells against the antibody titer of each individual. Serological evidence of exposure to B. pseudomallei is clinically determined by an indirect hemagglutination (IHA) assay which mostly detects antibodies to conserved lipopolysaccharides and/or capsular polysaccharides and is useful in diagnosis of melioidosis particularly in non or low endemic areas [1]. Although the threshold IHA titer for serodiagnosis varies in different countries, in the Northeast Thai population; an IHA titer 1∶40 is considered to be indicative of previous exposure to B. pseudomallei in healthy individuals [15],[18]. The frequency of specific, B. pseudomallei induced IFN-γ cells closely correlated with the serological status of the donor, whereas no such correlation was observed with control antigens derived from viruses known to be prevalent in the Thai population. Thus environmental exposure to B. pseudomallei induces concordant adaptive T and B cell responses as also seen in other examples of infection or vaccination [26],[27].
In mice, IFN-γ is critical for survival of the infected host and NK cells, as well as both CD4+ and CD8+ T cells contribute to its production [9],[10]. Using intracellular cytokine staining and specific cell depletion we found a similar situation in humans in which all three cell types produced IFN-γ in vitro, with their relative contribution differing according to the serological status of the host. An initial finding was that human NK cells were prominent producers of IFN-γ in vitro, providing some 80% of the IFN-γ positive cells in the first few hours of the culture period. This response was observed in both seronegative and seropositive individuals, was not inhibited by CsA and most likely represents an innate response to the bacterium presumably driven via the generation of IFN-γ inducing cytokines such as IL-12, IL-15 and IL-18 in culture [9],[28]. This observation may also explain the previous findings by Lauw et al of a rapid IFN-γ dependent induction of the chemokines Mig and IP-10 in whole blood cultures of healthy individuals in the presence of dead B. pseudomallei [29].
In seropositive individuals, this innate response was supplemented by the presence of IFN-γ positive CD4+ T cells and CD8+ T cells in both recovered melioidosis patients and asymptomatic healthy control subjects. A predominance of CD4+ T cells was observed from the peripheral blood of recovered melioidosis subjects. Haque A, et al. also reported that antigen-specific CD4+ T cells were important for the resistance against B. pseudomallei during the later phase of primary infection [10]. Clear evidence of priming of CD8+ T cells was also observed, presumably reflecting the cytoplasmic habitat of the bacterium within host cells [14]. These antigen specific T cells provided the majority of the total IFN-γ generated in culture as evidenced by their higher mean fluorescent intensities (MFI) of IFN-γ staining (Figure S3), larger ELISPOT sizes (data not shown) and by the significant effect of T cell depletion on the IFN-γ ELISPOT response. Of note, we have compared T cell responses by the production of IFN-γ vs. granzyme B by ELISPOT and the results showed high correlation of these 2 indicators in response to B. pseudomallei suggesting the importance of cytotoxic T cell response in melioidosis (Figure S4). However, the role of these cells to combat this intracellular pathogen requires further studies.
We then used differential expression of CD45RA and CCR7 to characterize the IFN-γ producing T cells as either central memory (CM), effector memory (EM) or a more recently described effector memory RA (EMRA) populations [30]–[32]. By these criteria, >80% of IFN-γ+ CD4+ T cells and >90% of IFN-γ+ CD8+ T cells reacting to B. pseudomallei were TEMRA cells, with the remaining minority being TCM and TEM cells. Thus, B. pseudomallei predominately induces ‘effector memory RA’ T cells in the peripheral blood that respond rapidly to repeated exposure to the microorganism as also reported with other pathogens such as human cytomegalovirus and human immunodeficiency virus [31],[33]. There was a trend towards a greater contribution of TEM and TCM cells in patients with a history of septicemia compared to localized melioidosis but this did not attain statistical significance and further studies using larger cohort sizes are needed to confirm this.
Several earlier reports established that exposure to B. pseudomallei primed human T cells for proliferation and secretion of the macrophage activating cytokine IFN-γ in vitro. However, these studies were restricted to small numbers of individuals in Northern Australia and Papua New Guinea and did not define the frequencies, memory phenotypes of the responding cell populations or the antigen specificity of these responses. The results presented here confirm and extend these findings to a larger sample size in the endemic region of Thailand. A consistent finding in all studies are that T cell responses were greater in seropositive versus seronegative individuals. With the larger group sizes provided in the Thai population we can go further and show that this also correlates with antibody titer, and not simply between antibody positive versus negative status. What is less clear is the relative strength of the cell mediated responses between seropositive healthy donors and recovered patients. Barnes et al found that lymphocyte proliferation and IFN-γ production was greater in seropositive healthy donors (n = 8) than those recovered from infection (n = 5), arguing as in the case of tuberculosis, of impaired immunity in those who experienced clinical disease [13]. However, our results in the Thai population showed no difference in the frequencies of IFN-γ producing cells in the recovered melioidosis group versus seropositive healthy donors, although both were clearly greater than seronegative control subjects. In contrast, the amount of IFN-γ secreted (as determined by ELISA) and the frequency of high IFN-γ responders was greater in the recovered group suggesting an increased immune priming following a significant bacterial burden compared to healthy exposed control subjects. Of note, even with the larger sample sizes used here, IFN-γ responses were similar between individuals with and without diabetes, in patients with septicemic versus localized disease or in cases of recurrent versus single episodes of disease [4].
Given the high mortality of acute melioidosis and the problems of treatment, the development of an effective vaccine is an important but difficult task. This is needed to protect individuals living in endemic areas as well as in situations of accidental or purposeful exposure following a bioterrorism scenario. Experimental strategies using wild type bacteria of reduced virulence [34], live attenuated mutants of B. pseudomallei [23],[35] and killed whole cells [12] have all been attempted with varying success. However, one important approach requires identification of individual B. pseudomallei specific proteins, which are both immunogenic and protective, for inclusion in protein and/or polysaccharide sub-unit based vaccines [36]–[38]. To date, the number of B. pseudomallei proteins which have been defined as T cell immunogens in mice or humans is very limited [38]–[40]. In other pathogenic bacteria, ABC transporter proteins have roles in bacterial survival, virulence and pathogenicity, are immunogenic in humans and an increasing number are being considered as candidate vaccine antigens [41]–[43]. We have previously shown that three members of the bacterial ABC transporter family, LolC, PotF and OppA are immunogenic in mice and particularly in the case of LolC provide at least partial protection against lethal challenge with B. pseudomallei following immunization in adjuvant [17]. We now show that all three proteins are recognized by T cells from seropositive individuals and could be considered for future vaccine development. No T cell response was observed in B. pseudomallei seronegative individuals, arguing that these antigens are relatively specific for B. pseudomallei and priming is not the result of cross reactivity against other common bacterial infections in the community.
In conclusion, we provide here the most extensive study to date of the human cell mediated immune response to B. pseudomallei and the first to define this aspect of immunity in Thailand, the major endemic focus of melioidosis in the world. Our data demonstrate that B. pseudomallei specific CD4+ T cells secreting IFN-γ are generated following exposure to the bacterium in the environment and the magnitude of this cellular response correlates with the serological status of the individual. Our findings that NK cells and CD8+ T cells also provide a potential source of IFN-γ, may help to explain the apparent lack of impact of HIV/AIDS on the incidence of melioidosis in Thailand. Our ability to detect specific T cell responses to defined B. pseudomallei proteins in large numbers of individuals now provides the opportunity to screen candidate antigens for inclusion in protein or protein-polysaccharide conjugate subunit vaccines against this important and emerging infection.
|
10.1371/journal.pntd.0002719 | Deliberate Attenuation of Chikungunya Virus by Adaptation to Heparan Sulfate-Dependent Infectivity: A Model for Rational Arboviral Vaccine Design | Mosquito-borne chikungunya virus (CHIKV) is a positive-sense, single-stranded RNA virus from the genus Alphavirus, family Togaviridae, which causes fever, rash and severe persistent polyarthralgia in humans. Since there are currently no FDA licensed vaccines or antiviral therapies for CHIKV, the development of vaccine candidates is of critical importance. Historically, live-attenuated vaccines (LAVs) for protection against arthropod-borne viruses have been created by blind cell culture passage leading to attenuation of disease, while maintaining immunogenicity. Attenuation may occur via multiple mechanisms. However, all examined arbovirus LAVs have in common the acquisition of positively charged amino acid substitutions in cell-surface attachment proteins that render virus infection partially dependent upon heparan sulfate (HS), a ubiquitously expressed sulfated polysaccharide, and appear to attenuate by retarding dissemination of virus particles in vivo. We previously reported that, like other wild-type Old World alphaviruses, CHIKV strain, La Réunion, (CHIKV-LR), does not depend upon HS for infectivity. To deliberately identify CHIKV attachment protein mutations that could be combined with other attenuating processes in a LAV candidate, we passaged CHIKV-LR on evolutionarily divergent cell-types. A panel of single amino acid substitutions was identified in the E2 glycoprotein of passaged virus populations that were predicted to increase electrostatic potential. Each of these substitutions was made in the CHIKV-LR cDNA clone and comparisons of the mutant viruses revealed surface exposure of the mutated residue on the spike and sensitivity to competition with the HS analog, heparin, to be primary correlates of attenuation in vivo. Furthermore, we have identified a mutation at E2 position 79 as a promising candidate for inclusion in a CHIKV LAV.
| With the adaptation of chikungunya virus (CHIKV) to transmission by the Aedes albopictus mosquito, a pandemic has occurred resulting in four to six million human infections, and the virus continues to become endemic in new regions, most recently in the Caribbean. CHIKV can cause debilitating polyarthralgia, lasting for weeks to years, and there are currently no licensed vaccines or antiviral therapies available. While an investigational live-attenuated vaccine (LAV) exists, problems with reactogenicity have precluded its licensure. The purpose of the current study was to: i) devise an in vitro passage procedure that reliably generates a panel of CHIKV envelope glycoprotein mutations for screening as vaccine candidates; ii) determine the position of the mutations in the three-dimensional structure of the alphavirus spike complex and their effect on electrostatic potential; iii) determine the attenuation characteristics of each mutation in a murine model of CHIKV musculoskeletal disease; and iv) to identify in vitro assays examining the dependency of infection upon HS that correlate with attenuation and localization in the glycoprotein spike. This approach provides a paradigm for the rational design of future LAVs for CHIKV and other mosquito-borne viruses, by deliberately selecting and combining attenuating processes.
| In the last few years, considerable attention has been focused upon mosquito-borne chikungunya virus (CHIKV); once a relatively obscure member of the Alphavirus genus in the Togaviridae family of enveloped, positive-sense RNA viruses [1]–[3]. In 2005, an East African clade CHIKV strain emerged on the Indian Ocean island of La Réunion that was maintained in a human-mosquito-human transmission cycle and caused a massive outbreak of CHIK fever [4]. Spread and/or re-emergence of CHIKV in Indian Ocean areas, Asia, southern Europe, and most recently in the Caribbean, in the following years has resulted in an estimated four to six million cases of CHIK fever with painful, often chronic, arthritides and an ongoing worldwide public health problem [3], [5]. The future occurrence of autochthonous cases in the mainland Americas seems inevitable with frequent travel-associated virus introduction, and the likelihood of a resulting outbreak is predicted to be high [6], [7]. Thus, the need to develop therapeutics and vaccine candidates for protection against this virus is ever more urgent. A cell culture-adapted LAV (181/25) is available as an investigational drug to at-risk researchers [8], [9]. However, insufficient attenuation and consequent reactogenicity problems have precluded its licensure for general use [8]. New strategies are required to develop CHIKV LAVs combining a more refined balance between attenuation and immunogenicity [10].
Like all members of the genus, CHIKV has an infectious, single-stranded RNA genome of ∼11 kb with a m7G 5′ cap structure and a 3′ polyadenylated tail (reviewed by [11]). Within this genome are encoded four non-structural (nsP1-4) and three structural (C, 6K/E1 and pE2) proteins from two open reading frames in the genomic and subgenomic RNAs, respectively [12], [13]. CHIKV virions have icosahedral symmetry, with a glycoprotein shell enclosing the viral membrane and nucleocapsid [14]. pE2 and E1 glycoproteins insert into the cell's endoplasmic reticulum as they are synthesized, forming heterodimers that trimerize into the viral ‘spikes’ and envelop nucleocapsids, budding as virus particles from the cell's plasma membrane [11]. pE2 is cleaved by host furin into E2 and a released E3 fragment during egress to produce the mature, fusion-competent virions in preparation for the next round of infection.
CHIKV isolates bind to cell surface receptors via E2 and fuse with cell membranes by clathrin-independent, Eps15-dependent, endocytosis via E1 [15]. The relationship between alphaviruses and their receptors is a complex one with many questions still unanswered and the identity of the receptor(s) utilized by wild-type isolates is still elusive with the exception of C-type lections, DC-SIGN and L-SIGN [16]. For many years, receptor identification was clouded by the use of strains that had adapted to growth in cultured cells. Fifteen years ago, we demonstrated that in vitro passage of the prototypic Old World alphavirus, Sindbis (SINV), in different laboratories had resulted in the accumulation of positively charged mutations in the E2 glycoprotein, which dramatically improved the virus-cell surface receptor interaction in vitro [17], [18]. In the converse experiment, Griffin and coworkers showed that in vivo passage in immune-deficient mice of a laboratory SINV strain could select for acquisition of negative charge and reduced heparan sulfate (HS)-dependence in vitro [19], [20]. Amino acid substitutions that increased net positive charge in certain E2 regions could dramatically increase per particle infectivity in cultured cells, dependent upon ionic interaction with negatively charged, cell-surface HS chains [17], [19], [21]. The following observations are of profound importance to the current study and to the biology of Old World alphaviruses in general: i) the substitution for positively charged residues in E2 that confer enhanced, HS-dependent infectivity in vitro is a common phenomenon amongst cell culture-passaged strains of SINV [17]–[19], [21], Ross River (RRV; [22]) and Semliki Forest (SFV; [23]) viruses; ii) these mutations can be selected within only a few serial passages in vitro [17]; and iii) viruses whose in vitro infectivity is enhanced by artificial HS attachment/entry are typically attenuated/avirulent in vivo from the natural infection route, at least in part due to reduced level and/or duration of viremia [20], [24].
We demonstrated recently that a wild-type CHIKV strain (LR2006 OPY1), isolated during the La Réunion outbreak and sequence-stabilized in cDNA clone form (CHIKV-LR; [25], [26]), exhibited no significant dependence upon HSPGs or other glycosaminoglycans (GAGs) for infectivity in vitro [27]. In contrast, the 181/25 CHIKV LAV candidate, derived by 18 serial passages of wild-type CHIKV (strain 15561) in MRC-5 fibroblasts to achieve attenuation [9], was highly dependent upon ionic interaction with HS for infectivity [27]. We predicted that this was due to an amino acid substitution at E2 position 82 [27], which was subsequently shown by Weaver and coworkers to attenuate both CHIKV-15561 and CHIKV-LR in vivo [28].
Here, we have exploited existing knowledge to deliberately select for and identify a set of E2 mutations that confer HS-dependence for infectivity by serial passage of wild-type CHIKV-LR on different cell-types in vitro. Single amino acid mutations that became predominant in the virus population within only five to ten passages through mammalian or mosquito cells were predicted by computational modeling to alter the electrostatic profile of the E2 glycoprotein and increase net positive charge in two exposed regions. By individual introduction of these mutations into CHIKV-LR, we have identified a panel of E2 mutations that confer reduced virulence in a murine model of musculoskeletal disease (MSD) and associated these with particular aspects of dependence upon HS for attachment/infectivity in vitro. In particular, we identified a novel mutation at E2 position 79 that increased attenuation over the E2-82R mutation present in the 181/25 LAV but did not diminish immunogenicity or protective efficacy. The positions of the attenuating mutations are clustered within two regions in the three-dimensional structure of the alphavirus trimer-heterodimer with the most attenuating mutated residues prominently exposed to the exterior of the spike. Furthermore, mutations conferring the greatest attenuation were associated with very small plaque size and sensitivity to competition with the HS analog, heparin. We propose this approach as an informed means to create mutant viruses and utilize in vitro HS interaction phenotypes and structural modeling to identify promising candidates for inclusion in CHIKV and other arboviral LAVs.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animal procedures were performed according to a protocol approved by the Institutional Animal Care and Use Committee of the University of Pittsburgh (Protocol 1001073).
Pregnant and 21 d old CD-1 (Charles River Laboratories), and 8 wk old STAT129 (Taconic Laboratories) mice were housed under specific pathogen free conditions and all experiments were conducted at ABSL-3.
BHK-21, CHOK1, pgsA745, pgsD677, and RAW264.7 cells were cultured as previously described [27], [29]. MC3T3-E1 osteoblasts were maintained in alpha minimum essential medium (AMEM) supplemented with 10% fetal bovine serum (FBS), 1 mM sodium pyruvate and 0.05 g/mL L-glutamine.
CHIKV-LR virus stock was serially passaged 10 times in triplicate series on CHOK1, pgsA745, or C6/36 cells, with a 1∶100 dilution of progeny virions between passages. At P5 and P10 supernatant from infected cells was placed in Tri Reagent-LS (MRC) containing 5 µg of tRNA carrier, and total RNA was extracted as per manufacture instructions. To sequence mutations in E2, cDNA was generated using RT-PCR (Roche) with a specific primer in the E1 gene, immediately downstream of the E2 gene 3′ terminus (GCAGCCTCTTGGTATGTGGC), and the entire pE2 gene was PCR amplified (S-CTAATGAAGGAGCCCGTACA; AS-GCAGCCTCTTGGTATGTGGC) using Deep Vent polymerase (NEB). The PCR fragment was either directly sequenced (Retrogen) or cloned into pCR-Blunt (Invitrogen) and sequenced.
pE2 gene mutations were introduced into the cDNA clone of CHIKV-LR using the Quick Change II XL mutagenesis kit (Stratagene). CHIKV-LR reporter viruses were created by inserting a cleavable in-frame fusion between capsid and E3 using Quick Change II XL mutagenesis to insert a PCR fragment at the capsid/E3 junction that encodes the first five amino acids of E3 fused in-frame with firefly luciferase (fLuc) followed by the 2A-like protease of Thosea asigna virus. Stocks of CHIKV-LR, E2 mutant viruses, and reporter viruses were generated from cDNA clones as previously described [27]. Briefly, cDNA was linearized and in vitro transcribed (mMessage mMachine, Ambion) to generate infectious, capped viral RNA genomes. Viral particles were harvested from supernatant of BHK cells 18–24 h post-electroporation with these RNAs. For all virus stocks, supernatant was clarified by centrifugation and single-use aliquots were stored at −80°C.
Virus stocks (200 µL) were treated with 40 U of RNase ONE (Promega) for 1 h at 37°C to remove any contaminating RNA that was not encapsulated in the virion and added to Tri Reagent-LS (MRC) along with 5 µg of carrier tRNA, before RNA was extracted per manufacture instructions. Equal total RNA concentration was used for reverse transcription (RT) with a primer complementary to sequence in the nsP2 gene and tagged with T7 to reduce background (5′-CGTAATACGACTCACTATAAGTACGTTGACGTGCTCTGACGTT-3′). Equal volumes of cDNA were used for real-time (q) PCR using SYBR green (Fermentas) to detect nsP2 on the positive strand (nsP2-TCfGTGTTAACGTGCTTCAGAGGGT; T7-GCGTAATACGACTCACTATA). A standard curve for viral genome number by qRT-PCR of RNA in vitro-transcribed from the CHIKV-LR cDNA clone. qRT-PCR results were analyzed using CFX Manager software (BioRad).
The models of the E1/E2 trimer of the 05-115 CHIKV strain and mutants, including the E2 N-terminal tail that is missing in the CHIKV template trimer structure (pdb: 2XFB) [14], were constructed by using Modeller version 9v8 [30]. The presence of the N-terminal tail was found to have significant influence on the electrostatic potential in the areas of interest in the current studies. Charge of individual atoms and their radius parameters based on an amber force field [31] were generated by pdb2pqr program [32]. Electrostatic potential was generated by Adaptive Poisson-Boltzmann Solver (APBS) package [33]. A linearized Poisson-Boltzmann equation was applied with dielectric constant 2.0 for protein and 78.0 for solvent. Electrostatic potential on solvent accessible surface in the range from -5 kT to 5 kT and solvent radius 1.4 A was visualized with the PyMol Molecular Graphics System (PyMol Molecular Graphics System, Schrodinger, LLC).
1 d old CD-1 mice were inoculated subcutaneously in the ventral thorax, 21 d CD-1 and 8 wk old STAT129 mice were inoculated subcutaneously in the hind footpad with either 10 µl of 105 genomes or 103 plaque forming units (PFU) of CHIKV viruses diluted in Optimem (Invitrogen). Mice inoculated with equal genome equivalents (GE) were inoculated with 105 genomes (equivalent to ∼103 PFU of CHIKV-LR). Mice were weighed and monitored daily for clinical signs of disease and mice showing severe signs of disease were monitored twice a day. The width and height of the metatarsal region of the rear footpad of CD-1 and STAT129 mice were measured daily with a caliper. Changes in footpad swelling were expressed as fold change in area (width×height) compared to pre-inoculation area. AST and percent mortality were calculated. Surviving mice were bled and challenged with 103 PFU of CHIKV-LR 21 d post primary infection.
At 48 h post-infection (p.i.), groups of three mice were euthanized with isofluorane and then exsanguinated by cardiac puncture to collect blood. The serum was separated from the blood using Microtainer tubes (Becton-Dixon). Mice were then perfused with PBS-1% DBS virus diluent (VD). Tissues collected were homogenized in VD by mechanical disruption. Virus titer was assessed in supernatants from homogenized tissue by standard plaque assay on BHK cells and titers were expressed as PFU/g, mL or draining lymph node (DLN). Serum cytokine concentrations were measured using a mouse cytokine 20-plex kit (Invitrogen) per manufacture instructions and analyzed using the BioRad Bioplex 200.
Serum was diluted at a final concentration of 1∶20 in VD containing the CHIKV-LR reporter virus expressing fLuc. Serum and virus were incubated together at room temperature for 30 min before being used to infect a 96-well plate of BHK cells for 1 h at 37°C. Cells were washed twice with VD before culture media was added and the cells were then incubated for 16 h before cell lysates were harvested using passive lysis buffer (Promega). fLuc substrate (Promega) was added and relative light units (RLUs) were determined by microplate luminometer (Orion).
Virus was diluted in either RPMI1640 (has a basal salt level of 103 mM NaCl) or RPMI1640 containing different concentrations of NaCl and used to infect MC3T3-E1 cells for 1 h at 37°C before overlaying with immunodiffusion-grade agarose.
For the footpad swelling, the area of under the curve was determined (GraphPad PRISM software) to assess the differences in swelling during the entire course of infection and then Student's t test was used to determine significance. Student's t test was used for all other experiments.
As a proof of concept, we began these studies by passaging CHIKV on cells from two evolutionarily divergent organisms, Chinese hamster ovary (CHO) K1 fibroblasts and C6/36 Aedes albopictus mosquito cells. A population of wild-type strain CHIKV-LR particles with maximal genome sequence homogeneity was generated by transfection of in vitro-transcribed, full-length, infectious viral RNA genomes into BHK-21 fibroblasts. This virus population was then subjected to a positive selection pressure for rapid growth on CHOK1 or C6/36 cells by performing ten sequential passages in triplicate parallel series. Infectious virion yields in supernatants, harvested after each amplification, declined slightly on both cell-types in the first two passages (data not shown) but then gradually increased to surpass passage 1 (P1) yields by ∼10 or 1,000-fold on mammalian or mosquito cells, respectively (Fig. 1). These data gave an initial indication of adaptation to cell culture within five to 10 passages. Importantly, infectious virion yields were not significantly enhanced by passage on CHOK1-derivative pgsA745 cells, deficient in the synthesis of GAG chains, suggestive that GAG-dependent changes underlay cell culture adaptation.
The sequence encoding the entire pE2 protein (E3 and E2) was analyzed from virus populations at P5 and P10 to identify mutations that might have accumulated during cell culture passage (Table 1). Each population of RT-PCR products was sequenced to reveal mutations present in the majority of packaged genomes. In addition, several individually cloned RT-PCR products from each P10 population were sequenced to determine whether amino acid substitutions were occurring alone or together, and to identify mutations occurring at lower frequency in the population. Passage on C6/36 mosquito cells selected strongly for an E to K substitution at E2 position 79, as this variant was the majority (or consensus) sequence for progeny virus genomes by P5 in three parallel passage series, and was retained through P10. In two of five individually cloned pE2 sequences from one passage series, we also identified a deletion of the codon for negatively charged E2-166E in mosquito cell-passaged virus populations. This mutation arose independently of E2-79K, which was found in the other three sequenced clones for this passage series, and in all sequenced clones for the two other passage series. Surprisingly, no other nucleotide mutations were detectable at this depth. Passage on CHOK1 fibroblasts selected strongly for an S to R substitution at E2-159 in the majority sequence in each of the three parallel passage series within five passages, and was retained as the dominant mutation in two passage series through P10. Interestingly, however, by P10 in the third passage series the dominant mutation in the population, and in four of six individual clones, was the E2-79 E to K substitution also selected on C6/36 cells. Three additional pE2 mutations were revealed in individual virion sequences. An E2-55 G to R substitution occurred at two out of five frequency in the second passage series on CHOK1 cells, while the combination of E2-99 H to Y/E2-168 E to K was present in two out of six of the population of the third passage series. After serial passage on GAG-deficient CHO cells, a dominant E2-264 V to A substitution was detected in one of three passage series at P5 and two of three at P10.
In total, seven amino acid changes were identified in the E2 proteins of the cell culture-passaged virus populations, four of which substituted a neutrally, or negatively, charged residue with a positively charged one (G to R at E2-55; E to K at E2-79; S to R at E2-159 and E to K at E2-168), while a fifth deleted a negatively charged amino acid (E2-Δ166E). Neither the V to A substitution at E2-264 nor the H to Y substitution at E2-99 would be anticipated to alter the net charge.
Based upon the X-ray crystallographic E1/E2 heterotrimeric structure of the CHIKV clinical isolate 05-115 [14], [34], we generated a 3D structural model of the trimeric envelope glycoprotein heterodimer (Fig. 2A) and used Adaptive Poisson-Boltzmann Solver (APBS; [33]) to determine the electrostatic potentials for the trimer-heterodimers of the wild-type CHIKV-LR strain and the identified mutations (Fig. 2B–D). The cell culture passage-selected E2 mutations mapped to two regions previously defined by Voss et al. [14]. Residues E2-55 and E2-79 mapped to the “wing” portion of Domain A in the region of insertion strands i3, i5 and i6 into Ig-like domains (Fig. 2C). The E2-82R mutation selected during 18 serial passages of the 15561 wild-type CHIKV strain on MRC-5 fibroblasts to produce the 181/25 CHIKV LAV candidate [35], and previously shown to attenuate both 15661 and CHIKV-LR [28], also mapped within this region. When viewed from above the three-fold axis of symmetry (Fig. 2B), the E2-55 residue resides in a cleft, overhung by other portions of Domain A, but facing outward from the spike interior adjacent to the β ribbon connector, whereas the E2-79 and E2-82 residues lie toward the apical surface of the protein, facing the solvent-exposed interior of the E1/E2 heterotrimer, with E2-79 more exposed than E2-82.
The E2-159R, E2-Δ166E and E2-168K mutations mapped to the acid-sensitive region in arch 1 of the β ribbon connector between Domains A and B (Fig. 2D). Interestingly, an E to K substitution at E2-166 was previously selected during passage of CHIKV-06.049 on human epithelial carcinoma HeLa cells [36]. The E2-159, E2-166 and E2-168 residues lie within the β ribbon connector facing outward from the heterotrimeric spike interior essentially on the opposite side of the E1/E2 heterodimer from E2-79 and E2-82 residues. E2-166 and E2-168 residues are close to areas of contact between the β connector and E1, although residue E2-166 is located more towards the heterotrimer exterior than E2-168 which is deeper within an invagination of the spike between the β connector and E1. The E2-159 residue lies higher and more towards the cleft in Domain A, where E2-55 is located.
For ease of viewing, regions affected by the changes in electrostatic potential created by each mutation are shown magnified on only one of the E2 molecules in the heterotrimer (Fig. 2E & F). Those amino acid mutations that were predicted to increase net positive charge also produced localized increases in the computer-predicted positive electrostatic potential on the surface of the E2 protein. In contrast, the E2-Δ166E deletion mutation appeared to affect a broader area and, at the resolution of the model, alter contact regions between the β connector and E1. The prominent exposure of the positive-charge shift conferred by the highly attenuating E2-79K mutation in comparison with the similarly located E2-55R and E2-82R is particularly apparent in the top view of the spike three-fold axis of symmetry (Figs. 2H–J).
Each amino acid substitution or deletion discussed above, including E2-82R and E2-166K, was introduced separately into the CHIKV-LR cDNA clone. Stocks of CHIKV-LR and the E2 mutant viruses were generated by transfection of in vitro-transcribed, capped genomes, and not passaged further. Reasoning that the positive electrostatic potential increases in E2 would result in a dependency upon cell surface HS for infectivity, we compared particle infectivity on CHOK1 cells versus derivatives that lack the ability to synthesize either all GAG chains (pgsA745) or just HS (pgsD677; [37]). Like other Old World alphaviruses, the infectivity of wild-type CHIKV-LR did not depend upon the presence of these sulfated glycans (Supplemental Fig. S1; [27], [38]). In contrast, the infectivities of LR-55R, LR-79K, LR-82R, LR-159R, LR-166K, LR-Δ166E and LR-168K for CHO cells all exhibited significant dependence upon GAGs, and this phenotype was almost completely conferred by the absence of HS alone. Neither E2-99Y, nor E2-264A mutation, exhibited significant dependence upon GAGs for infectivity (data not shown). In our experience the usefulness of the pgsD677 and pgsA745 CHO cells is limited to determining whether or not viral infectivity is affected by the absence of HS or GAGs but does not accurately determine relative degrees of dependency. However, these data indicate that predicted increase in exposed positive charge on E2 in the trimer-heterodimeric spike correlates with a significant dependence upon HS for infection of CHO cells. Most of the mutations increased per-particle infectivity, with the notable exceptions of E2-Δ166E and E2-159R (data not shown), which reduced infectivity unless HS was present, indicating that the latter mutations may compromise attachment/entry via another receptor(s) pathway used by CHIKV-LR.
Focusing upon those mutations that increase electrostatic potential on the E2 surface positive charge compared to wild-type CHIKV-LR (Fig. 2), and conferred the ability to bind HS (Supplemental Fig. S1), the virulence of each mutant relative to wild-type CHIKV-LR was assessed in a murine model of MSD with edema/inflammation, by measuring hind-limb swelling across the metatarsal region after subcutaneous inoculation of virus into the footpad, as described previously [27], [39], [40].
Two waves of limb swelling were consistently observed in mice infected with CHIKV-LR, the first peaking 1–2 d p.i., and the second 6–7 d p.i., with complete resolution by 12–14 d p.i. (Fig. 3A). Varying degrees and patterns of attenuation compared to the wild-type virus were observed for the mutant viruses, which we grouped into three categories for clarity. In the first category (purple), little or no attenuation was observed for LR-Δ166E infection (Fig. 3B) compared to wild-type CHIKV-LR, although the onset of clinical signs was delayed by ∼24 h and the duration of disease was longer for some animals. In the second category (blue), partial attenuation was observed for several virus mutants (Fig. 3C–E). LR-168K (Fig. 3C) was ∼24 h delayed and attenuated at 1–2 d p.i. but produced wild-type levels of swelling in some cases in the second phase. Interestingly, some animals infected with LR-168K also exhibited a more prolonged swelling than we observed for the wild-type virus infection. LR-55R (Fig. 3D) and LR-159R (Fig. 3E) were not delayed but demonstrated significantly reduced swelling compared with CHIKV-LR in both waves. In the third category (green), three virus mutants were highly attenuated, causing little or no hind-limb swelling (Fig. 3F–H). LR-166K (Fig. 3F) caused a transient, mild swelling only in the second wave, whereas no evidence of swelling was detectable for LR-79K (Fig. 3G) or LR-82R (Fig. 3H). This categorization and color scheme is used for subsequent figures to demonstrate prominent associations of genotype and phenotype.
A number of chemokines and cytokines have been associated with the acute phase of disease in humans, including MIG (CXCL9), MCP-1 (CCL2) and IP-10 (CXCL10) [41]–[45]. Furthermore, in a murine model similar to the one used here, MCP-1 has been shown to play an important role in pathogenesis of CHIKV-induced MSD [39], [41]. Interestingly, comparison of inflammatory responses to CHIKV-LR and the E2 mutant viruses revealed that higher induction of two cytokines (IL-12 p35/p40 and IL-5) and three chemokines (MCP-1, MIG and IP-10) correlated well with the severity of MSD (Fig. 4). On the other hand, IL-1α, IFN-γ and IL-2 were significantly induced in all virus-infected animals at 48 h p.i. versus mock-infected counterparts but no association between disease severity and these cytokine levels was observed (Fig. 4 and Supplemental Table S1). No significant elevation of other measured cytokines was observed over mock-infected animals for any of the viruses during this acute phase of infection (Supplemental Table S1). These data not only indicate that the levels of certain inflammatory molecules provide early biomarkers of MSD severity or attenuation in mice, but further validate the murine model of human infection.
In an effort to identify particular in vitro phenotypes that could be used to identify promising mutations for a LAV, we compared multiple characteristics of the dependence of infectivity of each mutant upon HS. By calculating plaques per GE (specific infectivity), we estimated that under plaque assay conditions, ∼1% of wild-type CHIKV-LR virus particles initiated infection on BHK-21 fibroblasts, 0.1% on Vero cells, 0.05% on CHOK1 cells and 0.01% on MC3T3-E1 osteoblasts (Fig. 5A). These susceptibility differences may be due to reduced attachment, entry and/or to subsequent steps in propagation. The LR-Δ166E mutant exhibited minimal changes in specific infectivity on the four cell-types compared to wild-type CHIKV-LR. Otherwise, with a few exceptions (E2-55R on Vero and MC3T3-E1, and E2-159R on CHOK1 and MC3T3-E1), the cell culture passage-derived E2 mutations tended to increase the infectivity of CHIKV particles by up to 70-fold (Fig. 5B) and thus provided an advantage to virions in culture conditions. However, these data also indicated that the virus-host receptor interactions differ between cell-types, as the hierarchies of cell susceptibility and virion infectivity were not consistently maintained.
To further explore in vitro phenotypes we focused upon the MC3T3-E1 osteoblasts as a cell-type with relevance to CHIKV-induced MSD in vivo [46] and known to secrete extracellular matrix components including HSPGs [47]. Virus inocula were prepared in media with a range of salt concentrations to disrupt potential ionic interactions between the virus particles and cell-surfaces during the infection period. As expected from our prior studies, the infectivities of wild-type SINV (SINV-TR339) and CHIKV-LR particles exhibited no significant dependence upon ionic interaction, whereas an E2-70K mutation in SINV-TR339 (SINV-K70) makes this virus highly sensitive to ionic interactions [17], [21], [27]. Interestingly, although the electrostatic model of the E2 protein predicted that the deletion of residue E2-166E would increase net-positive charge, the infectivity of LR-Δ166E virus for MC3T3-E1 osteoblasts was insensitive to even the highest salt concentration (350 mM; Fig. 5C), and indistinguishable from wild-type CHIKV-LR by this assay. The infectivity of the other viruses with positively charged E2 amino acid mutations was significantly disrupted at 200–250 mM salt concentration (Fig. 5C), indicating that the vast majority of virions now relied upon an ionic interaction for infectivity.
When we examined the ability of a high concentration of soluble heparin (200 µg/mL) to block infection of MC3T3-E1 osteoblasts (Fig. 6A), we were surprised to discover that only LR-79K, LR-82R and LR-166K were effectively competed versus BSA-treated controls and this treatment reduced infectivity by >90%. However, heparin is a uniformly highly sulfated GAG, lacking the diversity of HS chain charge distributions and therefore does not necessarily block all HS-ligand interactions. It should also be noted that, because the soluble heparin treatment unexpectedly increased the infectivity of wild-type CHIKV-LR on MC3T3-E1 cells by ∼15%, all of the mutations except for E2-Δ166E significantly increased sensitivity to heparin blocking even if only by a small fraction. Compared to previous observations for other alphaviruses that ionic interactions were almost completely HS-mediated and competed by soluble heparin [17], [19], [20], [29], [48], these findings suggest a more complicated interaction for these CHIKV mutants.
While performing the above experiments, it was noted that plaque sizes on MC3T3-E1 osteoblasts were highly variable between mutants (Fig. 6B). When quantified and compared with CHIKV-LR (Fig. 6C), it was revealed that the plaques formed by LR-166K, LR-82R and LR-79K were extremely small, LR-168K plaques were of intermediate size, while the plaque sizes of LR-Δ166E, LR-55R and LR-159R did not differ significantly from wild-type virus. The heparin blocking assay measures only the ability of the virus to initiate infection in the presence of varying amounts of heparin. As an extension to this assay, we determined whether or not the addition of heparin to the overlay impacted virus plaque size phenotypes as a more sensitive measurement of HS-dependence. As expected the plaque size of the wild-type virus was not affected (Fig. 6D). The three viruses exhibiting the smallest plaque size were also unaffected by added heparin, presumably because cell-surface or extracellular matrix HS already inhibits the spread of these highly sensitive mutants. However, heparin reduced the plaque sizes of LR-55R and LR-168K to a size comparable with LR-79K, LR-82R and LR-166K, and reduced the plaque size of LR-159R and LR-Δ166E, significantly. The reduced plaque size of LR-Δ166E under heparin overlay along with the reduced infectivity on HS-deficient CHO cells (Supplemental Fig. S1) suggested that this virus has a slight dependence on HS depending upon culture conditions, which fit with the prediction of an alteration of the positive electrostatic potential.
We examined the infectivity of each virus for MC3T3-E1 cells digested with microbial heparinases, (Hep) I, II and III, which digest cell-surface HS chains (recently reviewed in [49]). Hep II acts with little specificity, cleaving both HS chains and heparin regardless of sulfation. In contrast, Hep I cleaves primarily heparin and highly sulfated HS regions, while Hep III primarily cleaves less-sulfated regions of HS chains. Thus, the HS structures remaining on the cell surface after digestion differ between heparinases. Successful digestion of HS was confirmed by the greatly reduced infectivity of SINV-K70. As expected, viruses with little infectivity dependence upon HS (SINV-TR339, CHIKV-LR and LR-Δ166E) were unaffected by the digestion of HS chains with Hep I, II or III, even when high concentrations of enzyme were used (Fig. 7A–C). With the exception of LR-55R, heparin blocking (Fig. 6A) correlated with >90% reduction in infectivity on MC3T3-E1 cells digested with Hep II. However, the infectivities of these viruses were only reduced ∼50% by digestion with Hep I or III suggesting that they were able to utilize residual chains for infection regardless of their sulfation level, unlike SINV-K70. The infectivity of LR-55R was reproducibly reduced by only ∼50% following digestion with Hep I, II or III. Overall, LR-79K infectivity appeared to be the most sensitive to the digestion of HS, followed by LR-166K.
To better understand the reasons for attenuation of the HS-binding E2 mutants in vivo, especially LR-79K and LR-82R, we measured viral load in various tissues at 48 h p.i. (Fig. 8). This time-point was chosen to represent the first peak of swelling, and the previously observed peak of CHIKV-LR replication in this model [27].
Based upon the studies above, we proposed that two of the mutant viruses were sufficiently attenuated for inclusion in LAV vaccine formulations: LR-82R and LR-79K. To stringently determine the degree and stability of their attenuated phenotypes, we infected more susceptible animals. We have previously shown that mice deficient in STAT1-dependent type I IFN signaling pathways had exacerbated MSD, whereas CHIKV-181/25 remained partially attenuated in these animals [27]. STAT1-deficient mice infected with LR-82R exhibited exacerbated hind limb swelling compared to 129/Sv control animals, similar to that caused by wild-type CHIKV-LR (Supplemental Fig. S2). However, LR-79K infection of the STAT1-deficient mice caused only mild MSD compared to CHIKV-LR and LR-82R, from which we infer that LR-79K is substantially more attenuated for ability to cause MSD than LR-82R or even the 181/25 LAV. Both LR-82R and LR-79K infections were lethal to the STAT1-deficient mice, however, supporting our prior contention that MSD and fatality are not closely linked.
Finally, we immunized CD-1 mice with LR-82R or LR-79K, and challenged with CHIKV-LR three weeks p.i., to determine whether or not these attenuated E2 mutants could efficiently elicit a protective adaptive immune response. All of the immunized mice were completely protected from development of MSD upon CHIKV-LR challenge, whereas mock-immunized mice developed mild MSD (Fig. 9A). Only the second wave of swelling was observed in these control animals most likely due to the ongoing age-dependent attenuation of CHIK-LR-induced MSD in this model. Protection from disease was coincident with the presence of neutralizing antibodies prior to challenge (Fig. 9B) and, although we have not shown directly that this confers the immunity, antibody responses have been shown to protect [50], [51]. Interestingly, all of the E2 mutants elicited levels of neutralizing antibody not dissimilar to the wild-type infection by 21 d p.i. (Fig. 9B and data not shown), despite there being highly significant differences in dissemination, replication and disease.
The gold standards for a successful LAV against mosquito-borne virus infection are the 17D strains of yellow fever virus (YFV), currently used for routine immunization where YFV is endemic and in regional mass vaccination campaigns at the first sign of an outbreak [52]. The 17D vaccines immunize 99% of vaccinees with only one inoculum dose, provide immune memory for decades, and very rarely cause vaccine-associated adverse events. The original 17D virus was fortuitously selected by extensive blind cell culture-passage in the 1930's [53]. Although, the molecular mechanisms underlying the consistent immunogenicity and stable attenuation of 17D are largely unknown and cannot yet be intentionally duplicated for viruses or even for YFV, increased HS interaction has been identified as significantly contributing to the attenuation of 17D [54]. Alphavirus LAV candidates have been generated for VEEV (TC83; [55]) and CHIKV (181/25; [9]) by serial passage in cell culture but both have encountered problems during clinical trial for reasons of inadequate immunogenicity in some vaccinees and residual virulence in others. Yet, both of these vaccines are attenuated, at least in part due to more efficient interaction with HS than their wild-type counterparts [27], [48]. Attaining the ideal balance between attenuation of disease and immunogenicity is extremely challenging, and to do so by design will require knowledge of virus biology and the optimization of rational attenuation strategies that can be combined to produce a safe and effective vaccine. Thus, one characteristic of most, if not all, arbovirus LAV is increased HS interaction over the wild-type virus, yet these mutations have previously been selected at random as a component of extensive blind passages without knowledge of their contribution to attenuation of the vaccine.
Our current studies represent the first attempt to evaluate systematically selected mutations that confer interaction with HS and attenuate CHIKV in vivo. We began by assuming that a rational method for identification of a panel of E2 glycoprotein mutants highly enriched in HS binding mutations would be serial passage of CHIKV virus on two evolutionarily divergent cell types. Within an organism, HS biosynthesis is primarily regulated by domain distribution and degree of sulfation (i.e., the distribution of N-substituents and the levels of 2-O- and 6-O-sulfation) resulting in different composition of HS species from different cellular or tissue sources [56], [57]. Furthermore, invertebrate HS has a number of unique properties, including unusually low O-sulfation, yielding different domain structures from vertebrate HS [58], [59]. This diversity most likely results in the duality of both non-specific ionic interactions occurring between HS and positively charged ligands versus highly specific interactions of particular ligands with certain HS species [60]–[63]. Specificity for particular HS structures has been documented for several virus-HS interactions [38], [64]. Therefore, it was possible that mutations affecting the charge balance of E2 selected on different cell types might exhibit different types of HS interaction. While the particular contribution of HS binding to alphavirus infection is not fully characterized beyond increases in virus association with cells [17], [21], [24], it is possible that structural differences in HS chains, their attachment to core proteins or their associations with other cell surface factors may be involved in the selective advantage provided by a particular virus mutation in vitro.
Comparing CHO and C6/36 cell passage series, we obtained one common mutation altering charge (E2-79 E-K), three CHO-only mutations (E255 G-R, E2-159 S-R and E2-168 E-K) and one C6/36-only mutation (deletion of E2-166E). Neither the V to A substitution at E2-264, nor the H to Y substitution at E2-99, would be anticipated to have considerable impact upon electrostatic potential; therefore, these were not considered. Interestingly, the E2-159R selected in these studies was previously selected during passage of CHIKV on Vero primate kidney fibroblast cells [65], the E2-166K was selected on HeLa human adenocarcinoma cells [36], and the E2-82R in the vaccine strain was selected on MRC-5 human fetal fibroblast cells [9], although none of these studies deliberately selected for increased infectivity or HS interaction. Therefore, common mutations can be selected between evolutionarily divergent hosts (E2-79 E-K in hamster and mosquito, and E2-159 S-R in monkey and mouse) while it is possible that others are unique to particular cells or species. While not considered in our studies, it is possible that the number of different mutations selected would be increased by using cells from different tissue types reflecting the diversity of tissue specific HS structure described above.
To determine whether or not the mutations identified were unique to cell culture-adapted virus populations, we examined these positions in an alignment of 200 pE2 sequences available in GenBank and believed to be from isolates minimally or unpassaged before sequencing (Gardner et al., unpublished observations). Only two of the mutations selected in our study were present in any of these isolates: four isolates had E2-55R (ADV31296, AEE60792, AEE60797 and BAH97931) and 10 isolates had the E2-264A (AEX25348, AEX25344, AEX25346, AEE60795, AEE60793, AEE60796, AEE60794, AEE60792, CCA61129 and CCA61128). Whether or not these residues are present in naturally circulating viruses or represent cell culture adaptive mutations in the minimally passaged strains remains to be determined.
With the exception of LR-Δ166E, all of the E2 mutants were attenuated for the ability to cause MSD in mice. Although all of the viruses were clearly able to replicate at the site of inoculation in the footpad, in general their replication in MST was reduced. The virus mutants fell into three groups of disease severity with LR-Δ166E disease indistinguishable from CHIKV-LR; LR-168K, LR-55R and LR-159R eliciting a biphasic MSD similar to CHIKV-LR but with reduced disease; and LR-166K, LR-82R and LR-79K eliciting either a minor monophasic MSD (LR-166K) or no detectable MSD (LR-82R and LR-79K). As suggested by studies with other alphaviruses [17], [20], [48], the dependence upon HS for infectivity prevents efficient spread of CHIKV in vivo. In the current studies, this led to decreased levels of inflammatory chemokines such as MCP-1, MIG and IP-10 that are associated with CHIKV-induced disease. Further, the most attenuated viruses elicited the lowest levels of these factors. It was not surprising that the levels of IP-10 and MCP-1 correlated with attenuation, as these cytokines are chemoattractants for monocytes/macrophages which have been shown to be important for disease mice [39], [41] and humans, especially during the acute phase of disease [42]–[44], [66], and can be associated with higher CHIKV loads [44]. MIG and IL-12 are involved in T cell recruitment and differentiation respectively and have been shown in some human cohorts to be elevated during infection [42]–[45], [66], and CD4+ T cells have been shown to contribute to hind limb swelling in mice [67]. IL-1α, IFN-γ and IL-2 were increased in the serum of mutant viruses similarly to wild-type CHIKV-LR and several of these cytokines were also shown to be upregulated during human infections [41], [43]–[45], [66]. It remains to be determined if elevation of these cytokines is associated with the attenuated phenotype and/or the induction of the protective immune response. Taken together, the cytokine profiles in mice correlated with virulence and for many of the cytokines are similar to profiles in humans following CHIKV-LR infection.
We attempted to associate in vitro measurements of HS infection dependence with virulence, examining virus specific infectivity, infection efficiency in the absence of HS or all GAGs, plaque size, resistance of infectivity to competition with increasing salt concentrations, sensitivity to competition with heparin, and infection sensitivity to digestion of cell surfaces with heparinases. Each of these assays has previously been used to compare HS infectivity dependence of alphaviruses (e.g., [17], [19], [21], [22], [48], [68]). All viruses attenuated in comparison with CHIKV-LR did exhibit increased specific infectivity on at least one of the four cell types tested and this was associated with similar infectivity diminution in cells genetically deficient in GAG synthesis. LR-Δ166E dependence on HS was significantly less than the other mutants but much greater than CHIKV-LR indicating that at least partial dependence upon HS for infectivity is not invariably associated with attenuation. Interestingly, in contrast with the SINV-K70 mutant whose infectivity appeared to depend solely upon HS among GAGs, all CHIKV mutants exhibited a minor dependence of infectivity upon GAGs other than HS evidenced by slightly increased infectivity on HS-deficient pgsD-677 cells versus GAG-deficient pgsA745 cells. This suggests the possibility of differences in attachment/entry mechanisms of SINV versus CHIKV.
Disruption of infection with salt did not distinguish between attenuated mutants with each exhibiting moderate decrease at 200 mM and ∼90% decrease in infectivity at 250 mM and above. Interestingly, LR-Δ166E, although sensitive to genetic deficiency in HS and all GAGs, was insensitive to all salt concentrations used, similar to CHIKV-LR. Furthermore, digestion of HS chains with Hep I, II or III did not distinguish the three groups of attenuation phenotypes. Likely reflecting the different substrate specificities of the three enzymes, differential sensitivity to the three heparinases can distinguish between virulence phenotypes with other alphaviruses with genetic differences in HS binding domains [17], [29]. In these studies, CHIKV-LR was insensitive to all three heparinases, while LR-166K, LR-168K, LR-55R, LR-159R and LR-82R were partially sensitive to Hep I, highly sensitive to Hep II and partially sensitive to Hep III. Notably, LR-79K was significantly more sensitive to Hep I and Hep III than the other E2 mutants. LR-Δ166E, again, exhibited an intermediate dependence phenotype with all three heparinases. SINV-K70, in comparison, was highly sensitive to all three heparinases, again suggesting differences in GAG interactions between SINV and CHIKV mutants. In contrast with these assays, blocking of infectivity by reaction of virus particles with soluble heparin clearly distinguished LR-166K, LR-82R and LR-79K from the other viruses. Small plaque size and lack of change in size with addition of heparin to the overlay provided a similar distinction.
Comparison of location in an electrostatic map of the CHIKV heterotrimeric spike complex revealed that mutations that were more attenuating to disease in mice appeared to create additional positive charge that was more exposed to the exterior of the spike (e.g., E2-79K, E2-82R and E2-166K as opposed to E2-55R, E2-Δ166E or E2-159R). The lack of any attenuating phenotype with E2- E2-Δ166E was surprising, especially since the electrostatic model of this mutant suggested that the deletion should lead to increased electrostatic potential of E2 similar to that of the other E2 amino acid substitution mutants listed in Table 1 which demonstrated various degrees of attenuation. Clearly, the interpretation of the electrostatic potential maps of particular E2 mutant proteins will depend upon confirmation with in vitro assessment of GAG dependency.
Our data for the E2-166K mutation, which we tested because of its identical location to the selected E2-Δ166E mutation, indicate that it confers efficient attachment to HS. This mutation was selected by passage in a context where the authors concluded that the mutation increased virus resistance to the OAS3 antiviral protein, possibly by conferring a rapid entry phenotype [36]. Many of the HS-binding E2 mutants for SINV and VEEV were originally selected and identified as rapid entry mutants and only later shown to confer efficient HS binding [17], [24], [69]–[72]. The relationship of HS binding, rapid entry and antiviral resistance is unclear. However, if the LR-166K virus effectively antagonized the antiviral response as suggested by Henrik Gad et al. [36], one would expect the virus to be more virulent in vivo instead of more attenuated compared to CHIKV-LR and thus this phenotype may be an in vitro artifact or a localized phenomenon.
Overall, our data suggest that positively charged amino acid substitutions in CHIKV E2 that result in small plaques and efficient competition with heparin will be highly attenuated in the adult mouse model of MSD. The LR-166K, LR-82R and LR-79K viruses were very similar in terms of the primary in vitro correlates of attenuation: plaque size and heparin sensitivity. In addition, LR-79K was dramatically more attenuated for MSD than LR-82R in the severely immunocompromised STAT1-deficient mice. Testing in this model will refine choices between mutants with similar in vitro characteristics and sensitivity to all three heparinases may indicate the highest degree of in vivo attenuation within this group. Furthermore, combinations of such mutations can be tested to improve immunogenicity and/or stabilize attenuation. One question raised by these studies is whether or not this paradigm for deliberate mutation selection will be applicable to other alphaviruses and/or other arboviruses. We have detected differences between SINV and CHIKV mutants in dependency upon HS versus other GAGs for infectivity and it is unclear how this phenotype correlates with attenuation in vivo especially when evaluating the results of assays that are not specific to a particular GAG. It is likely that the specific combination of in vitro assays most correlated with attenuation must be determined for each virus type.
In summation, these studies demonstrate that very limited passage of CHIKV is sufficient to generate HS binding mutations that are highly attenuating but retain immunogenicity, and would make satisfactory candidates for testing in a LAV. Furthermore, using this approach, we have identified E2-79K as a single site mutation that is highly attenuating for MSD and stable even in a severely immunocompromised model of disease. It is possible that attenuating mutations present in other areas of vaccine genomes (e.g., nonstructural or non-translated region mutations in the YFV 17D vaccine or the VEEV TC83 vaccine [73]) would require further passage to accrue. Therefore, rational LAV creation may benefit from multiple selection strategies with an initial screen identifying mutants meeting the criteria we have outlined for HS binding-mediated attenuation followed by additional passage on the same cell type or others or introduction of mutations identified with other alphaviruses that could be reasonably transferred to CHIKV.
|
10.1371/journal.pgen.1000246 | Mutations in the SLC2A9 Gene Cause Hyperuricosuria and Hyperuricemia in the Dog | Allantoin is the end product of purine catabolism in all mammals except humans, great apes, and one breed of dog, the Dalmatian. Humans and Dalmatian dogs produce uric acid during purine degradation, which leads to elevated levels of uric acid in blood and urine and can result in significant diseases in both species. The defect in Dalmatians results from inefficient transport of uric acid in both the liver and renal proximal tubules. Hyperuricosuria and hyperuricemia (huu) is a simple autosomal recessive trait for which all Dalmatian dogs are homozygous. Therefore, in order to map the locus, an interbreed backcross was used. Linkage mapping localized the huu trait to CFA03, which excluded the obvious urate transporter 1 gene, SLC22A12. Positional cloning placed the locus in a minimal interval of 2.5 Mb with a LOD score of 17.45. A critical interval of 333 kb containing only four genes was homozygous in all Dalmatians. Sequence and expression analyses of the SLC2A9 gene indicated three possible mutations, a missense mutation (G616T;C188F) and two promoter mutations that together appear to reduce the expression levels of one of the isoforms. The missense mutation is associated with hyperuricosuria in the Dalmatian, while the promoter SNPs occur in other unaffected breeds of dog. Verification of the causative nature of these changes was obtained when hyperuricosuric dogs from several other breeds were found to possess the same combination of mutations as found in the Dalmatian. The Dalmatian dog model of hyperuricosuria and hyperuricemia underscores the importance of SLC2A9 for uric acid transport in mammals.
| Animals excrete waste products in their urine. When most mammals metabolize compounds, called purines, they produce allantoin as one waste product in their urine. Humans, great apes, and Dalmatian dogs produce a different breakdown product, uric acid. This leads to high levels of uric acid in the urine and blood. In humans, this can result in diseases such as kidney stones and gout and may cause hypertension. In Dalmatians, high uric acid levels result in bladder stones that often have to be removed surgically. The cause of high uric acid levels in humans and great apes is not the same as in the Dalmatian dog. Here we report the genetic cause of the Dalmatian condition. This change is shared by dogs from unrelated breeds, indicating that it predates the separation of dog breeds. The gene that causes excretion of uric acid in Dalmatians is important for controlling the amount of uric acid in human blood and is therefore important for human diseases. It is not clear why humans and great apes have evolved to excrete uric acid, but it appears that some dogs have developed a different mechanism that leads to the same result: elevations in urine and blood uric acid levels.
| Uric acid is the predominant product of purine metabolism in humans, great apes and one breed of dog, the Dalmatian; all other mammals excrete allantoin. During primate evolution, urate oxidase (UOX), which catalyzes the oxidation of uric acid into allantoin, accumulated several independent nonsense mutations that led to its silencing and resulted in high serum and urine uric acid levels in humans and great apes [1],[2]. Huu in the Dalmatian results from a different cause [3],[4]. Uric acid freely circulates in the form of urate, the salt of uric acid, in the plasma where it serves as a free-radical scavenger. Although uric acid has evolved in humans to be the main product of purine metabolism, this change has had some negative effects. High levels of urate predispose humans to gout [5],[6]. In addition, uric acid levels have been correlated with hypertension, vascular disease and metabolic syndrome although it is unclear whether hyperuricemia is primary or secondary in these cases [7]–[10].
As in humans, all Dalmatian dogs have a defect in urinary metabolism that leads to excretion of uric acid rather than allantoin [11]. As a result, Dalmatians are predisposed to form urinary calculi composed of urate (Figure 1B). Hyperuricosuria in the Dalmatian is relatively easy to identify since Dalmatian urine forms a crystallized precipitate when cooled (Figure 1C). This trait was probably fixed in the breed through selection for a more distinctive spotting pattern [12],[13]. Dalmatian coat pattern involves mutations in at least three different spotting genes (Figure 1A). Dalmatians have a mutation for extreme white in the MITF gene [14] that leads to an all white coat. A dominant mutation, called T for ticking [15], is responsible for adding the pigmented spots to the white coat. Based on segregation analysis, the huu locus appears to be closely linked to a modifier of spot size [16].
In mammals that produce uric acid rather than allantoin, the level of uric acid in the blood is controlled by differences in production as well as differences in the amount that is excreted in the urine. In the kidney, uric acid is filtered by the glomerulus and then a portion is reabsorbed in the proximal tubules where it re-enters circulation. There are species-specific differences in the production of uric acid versus allantoin and the relative amounts of reabsorption and secretion in the proximal tubules, making the use of animal models in this area of research challenging [17].
Dogs and humans, unlike many other mammals, undergo bidirectional transport of urate along the nephron, which results in net reabsorption of urate from the glomerular filtrate. In Dalmatians, reabsorption is lost entirely and urate excretion equals or exceeds the glomerular filtration rate [18]. This change in uric acid excretion by the Dalmatian kidney is not secondary to hyperuricemia since non-Dalmatian dogs with artificially raised serum uric acid levels can only clear uric acid at ∼1/3 the rate of the Dalmatian [18]. Free-flow micropuncture experiments were used to demonstrate that in Dalmatian kidneys there is a deficiency of proximal tubular reabsorption of urate [19].
Although findings stated above implicate the kidney in Dalmatian huu, reciprocal liver and kidney transplant experiments between Dalmatian and non-Dalmatian dogs demonstrate that the liver is also important for the phenotype. Kidney transplants between normal dogs and Dalmatians only partially ameliorated the hyperuricosuria phenotype [20],[21]. However, Dalmatian hepatocyte transplants can correct the phenotype [20]. Therefore, a logical cause for the Dalmatian phenotype is a mutation in urate oxidase, similar to humans. However, Dalmatian liver homogenates are capable of oxidizing uric acid to allantoin and the urate oxidase gene was excluded genetically [3]. The Dalmatian phenotype could also be explained by a generalized defect of urate transport since liver slices were not capable of converting uric acid to allantoin [4]. Dalmatian dog erythrocytes have been shown to transport urate normally, demonstrating that Dalmatians do not have a generalized defect in urate transport [22]. Although Dalmatians have functional urate oxidase activity in their livers, they effectively have a similar phenotype to humans and great apes since they cannot transport urate into the liver for degradation. The Dalmatian phenotype can be summarized as a hepatic and renal urate transport defect which leads to hyperuricosuria and relative (compared to other dogs) hyperuricemia.
The discovery of various proteins that transport urate has shed some light on the control of serum and urine uric acid levels [23]. In the kidney, urate is transported across the apical membrane of the proximal tubules and then across the basolateral membrane before re-entering circulation. In humans, the transporter that reabsorbs urate across the apical membrane in the proximal tubule is Urat1, or SLC22A12 [24], which is expressed exclusively in the kidney. Mutations in SLC22A12 are thought to be the major cause of idiopathic renal hypouricemia in humans [24],[25], which is also called “Dalmatian hypouricemia” since people with this disorder spill uric acid into their urine resulting in a phenotype similar to the Dalmatian dogs [26]. In addition, recent work has shown an association of SLC2A9 with serum uric acid levels in several different populations [27],[28]. Variants in the non-coding region of SLC2A9 are associated with gout and uric acid levels in several human populations [29]–[31]. SLC2A9 has been shown by Xenopus oocyte experiments to transport uric acid [30]. In humans and mice it is expressed in liver and kidney [32],[33]. In particular, in humans SLC2A9 isoforms have been localized to both the apical and basolateral membrane of the proximal tubules, allowing the possibility that SLC2A9 influences serum uric acid levels by transport in the kidney [34].
All Dalmatians are homozygous for huu, which is inherited as a simple autosomal recessive trait as demonstrated by crosses performed between Dalmatians and other breeds of dogs [12],[13]. In order to identify the causative gene, an interbreed backcross (Dalmatian×Pointer) was developed which introduced the wildtype version of the huu gene into the Dalmatian breed while maintaining the breed characteristics of the Dalmatian. Based on linkage analysis using this cross (LOD 6.55), huu localizes to CFA03, excluding SLC22A12 as a candidate [35]. Using recombination breakpoints in the interbreed backcross and taking advantage of the homozygosity within the Dalmatian breed for huu, a critical interval containing four candidate genes was defined. Microsatellite markers within this interval gave LOD scores over 17 for linkage to huu. A candidate causal missense mutation (C188F) was identified within a highly conserved transmembrane (TM5) of the 12 transmembrane transporter protein, SLC2A9. The missense mutation is homozygous in all Dalmatians tested (247) as well as in hyperuricosuric dogs of other breeds.
Linkage analysis using a Dalmatian×Pointer backcross family localized Dalmatian huu to CFA03. Haplotype analysis defined a 3.3 Mb critical interval (CFA03 72,063,073–75,355,028 Mb), estimated to contain 24 candidate genes [35]. Additional backcross dogs were genotyped with the microsatellites used for haplotype analysis and with new microsatellites mined from the critical region on CFA03. The full pedigree of all the dogs used for this analysis is shown in Figure S1. Urine uric acid/creatinine ratios were used to categorize the dogs' genotypes at huu. LOD scores were determined for a subset of these markers. Two markers within the critical interval defined by recombination breakpoints gave LOD scores over 17, further confirming the linkage to this region. Recombination events in two dogs narrowed the critical interval to a 2.5 Mb region containing 19 candidate genes (CFA03; 71,796,048–74,348,350 Mb) (Figure 2).
Since Dalmatians are fixed for huu, it was expected that an area of homozygosity around the huu locus would be identified. Blocks of linkage disequilibrium (LD) in purebred dogs extend between several megabases in rare breeds with small population sizes to several kilobases in popular breeds with larger population sizes [14],[36]. Dalmatians have a moderate population size, so the homozygous region surrounding huu should be smaller than 2.5 Mb. Microsatellites mined from the canine genome and located ∼100 Kb apart were genotyped in the backcross dogs. Markers that were homozygous in huu/huu individuals and heterozygous in huu/+ individuals were typed in 24 unrelated Dalmatians to verify that the region of homozygosity is not the result of familial linkage disequilibrium. Haplotypes were constructed to rule out the possibility of more than a single ancestral mutation. Regions of homozygosity were found between ms173 and ms9 (CFA03; 71,796,048–72,363,187) and ms187 and ms128 (CFA03; 74,183,928–74,245,743). Eighteen SNPs, extending across 2.5 Mb, were genotyped in 24 Dalmatians, a wild-type Labrador Retriever and a huu/+ backcross dog. The results excluded the area between ms187 and ms128 and a total of 13 SNPs spanning ∼333 Kb confirmed the homozygous region between ms173 and ms9 in the 24 Dalmatians (Figure 3). These SNPs are also heterozygous in the huu/+ backcross dog, polymorphic in two unaffected dogs (ND1 and ND2) and the Boxer genome assembly sequence.
Four candidate genes were identified in the region of homozygosity, LOC488823 (similar to mast cell immunoreceptor signal transducer – MIST), LOC479092 (zinc finger protein 518B, KIAA1729), LOC611070 (similar to WD-repeat protein 1 – WDR1) and LOC479093 (similar to solute carrier family 2, member 9 protein, isoform 1 – SLC2A9) (Figure 3). RT-PCR established that all four genes were expressed in canine liver and kidney. Since all the candidates were expressed in the appropriate organs, all four genes were sequenced from Dalmatian and non-Dalmatian liver cDNA samples and the untranslated regions (UTRs) were determined by 5′ and 3′ RACE (Genbank EU371511–EU371515).
The 5′UTR and exons 1–7 of MIST were outside of the LD region and were not pursued. A single silent mutation was identified in exon 16 of the MIST gene (T951C). Three silent mutations were found in the single exon of the canine KIAA1729 gene (T932C, C2480T, G3092A) and a 33 bp insertion/deletion was discovered in the 3′UTR. Both alleles of these three SNPs and the insertion/deletion were seen in unaffected non-Dalmatian dogs. WDR1 was sequenced in liver cDNA and genomic DNA from a Dalmatian and a non-Dalmatian. A single SNP was found in intron 9 that does not affect a splice site. Although SLC2A9 did not have an assigned function related to uric acid metabolism at the time this work was performed, it was considered to be the most promising candidate since it is a transporter protein.
A discrepancy in SLC2A9 exon annotation between NCBI and the UC Santa Cruz genome browser was addressed by 5′ RACE. Two SLC2A9 exon 1 variants were found in a Dalmatian and a wild-type Golden Retriever, variant O (CFA03:72,222,637–72,416,753) and variant N (CFA03: 72,227,605–72,416,753). The difference between the two variants lies in the first 28 amino acids of the N terminus, similar to known mouse and human variants. Both transcripts were shown to be expressed in canine liver and kidney by RT-PCR. Expression differences were observed for variant O between Dalmatian and non-Dalmatian in both liver and kidney (equivalent to isoform 2 in human) but not for variant N (Figure 4A). Expression in the Dalmatian samples was ∼50% of non-Dalmatian levels in both tissues. SLC2A9 expression was further evaluated by RT-PCR using primers that amplify both transcripts in 11 different tissues from an unaffected Beagle. The highest levels of expression were observed in the kidney and liver (Figure 4B).
Sequencing of the SLC2A9 gene was performed on canine liver cDNA as well as genomic DNA so that intron/exon boundaries and promoter regions could be evaluated. SLC2A9 coding exons 2–12, as well as the exon-intron boundaries, were sequenced in a Dalmatian and a Labrador Retriever. Six SNPs were discovered in the SLC2A9 sequence. Two are exonic; exon 5 G563T;Cys188Phe (nucleotide and amino acid numbering are reported with reference to variant N) and exon 11 G1303A;Val435Ile (nucleotide and amino acid numbering are reported with reference to variant N), two are located 99 and 101 bp 5′ to the start codon of variant O. SNPs were also identified in introns 1 and 10. None of the intronic SNPs are within or near conserved splice-site elements. The SLC2A9 exon 11 G1303A SNP is polymorphic in the 24 unrelated Dalmatians tested, consistent with its genomic location outside of the region homozygous in Dalmatians (SNP17, Figure 3). The polymorphisms located 5′ to the start codon were tested in a panel of DNA samples from 15 unaffected dogs from various different breeds. Both SNPs are polymorphic in unaffected non-Dalmatian dogs, displaying both the Dalmatian haplotype (A–C) and other combinations. These SNPs are always fixed in affected Dalmatian dogs (SNP9-10, Figure 3). Primary sequence data from the exon 5 SNP is shown in Figure 5A.
The two non-synonymous SNPs found in SLC2A9 coding sequence were tested using the PANTHER program (http://www.pantherdb.org/tools/), which assigned the Cys188Phe substitution a score of −4.047 (scores range between 0 to −10, −10 indicating the most deleterious) with a probability that it is a deleterious substitution of 0.74 [37]. The Val435Ile variation was not given a prediction since the substitution is not conserved in other species and unlikely to be deleterious. The SIFT program also scored the Cys188Phe missense mutation as deleterious with a probability of 0.01, where any probability below 0.05 is considered deleterious [38]. In order to evaluate the degree of protein conservation in the region flanking the candidate SNP, SLC2A9 protein sequences were compared between chicken, mouse, rat, human, Boxer and Dalmatian (Figure 4B). This region of the protein has a high degree of identity across mammals. The variant O promoter SNPs (located 99 and 101 base pairs upstream of the initiator methionine) were analyzed for transcription factor binding site differences and the SNPs are predicted to disrupt binding sites for DeltaE, AML-1a, S8 and Cre-BP in the Dalmatian version.
The SLC2A9 exon 5 C188F amino acid substitution was further tested in 247 Dalmatians and 378 non-Dalmatian dogs from 58 different breeds. It was homozygous in all Dalmatians tested and only the wildtype allele was present in the non-Dalmatian dogs.
Individual dogs from non-Dalmatian breeds are known to form urate calculi and have been diagnosed with huu [39],[40]. Two Dogs from two of these breeds, Bulldogs and Black Russian Terriers, that had formed urinary urate calculi were tested for the SLC2A9 exon 5 G563T;Cys188Phe missense mutation (SNP 15 in Figure 3) and the variant O promoter SNPS (SNP 9, 10 in Figure 3) and found to be homozygous. These four dogs shared the Dalmatian haplotype across the homozygous interval defined during the positional cloning of the huu locus (Figure 3).
The Dalmatian dog exhibits hyperuricosuria and relative hyperuricemia due to a defect of urate transport in the liver and kidney. Positional cloning of the huu locus using an interbreed backcross, as well as homozygosity within the Dalmatian breed, has identified SLC2A9 as the cause of the change in uric acid handling by this dog breed. Unrelated breeds of dog with hyperuricosuria or urate stone disease share the same haplotype as Dalmatian dogs, providing compelling evidence that this is the gene responsible. This result was somewhat unexpected because SLC2A9 was classified as a member of the large glucose transporter family and did not have an assigned function with respect to urate transport until recently.
SLC2A9 is classified as a part of the large glucose transporter family based on amino acid sequence identity of 44% and 38% to Glut5 and Glut1, respectively [32]. SLC2A9 has been localized to the cell surface in humans [34]. It contains 12 transmembrane domains and, based on homology to other glucose transporters, the central channel is essentially formed by helices 2, 4, 5, 7, 8, and 10 [41]. The Cys188Phe amino acid change occurs within a highly conserved residue located within the fifth transmembrane domain of the protein. Although a cysteine to phenylalanine amino acid change is not expected to disrupt the localization of the alpha helix to the transmembrane, this change could disrupt the proper function of the protein by altering the pore. Precedent for single amino acid changes altering substrate specificity has been shown for a number of the SLC2A transporters [42].
Expression differences between Dalmatian and non-Dalmatian samples were observed for one of the isoforms (O) of the gene (Figure 4). SNPs in the promoter region of this variant were identified and were fixed in the Dalmatian breed. These same SNPs were also identified in non-Dalmatians without hyperuricosuria. Therefore, these SNPs alone are not sufficient to confer the phenotype but along with the missense mutation they may contribute to the expression of the phenotype. It remains to be determined if there are variant O transcript differences in the general dog population and if the SNPs identified in this work are responsible for the differences in the level of variant O expression. In humans, the equivalent isoform to variant O is expressed on the apical surface of the proximal tubules [34]. Reduction in the amount of the protein combined with a deficiency in the protein itself may contribute to the decreased conversion to allantoin by the liver as well as the increased excretion of uric acid in these dogs. SLC2A9 tissue expression in dogs is similar to that observed in humans and mice. Expression in humans is highest in kidney, liver and placenta [32]. In humans, isoform 1 has been localized within polarized canine kidney cells to the basolateral membrane of the proximal tubules while isoform 2 is localized to the apical side [34]. The tissue and cellular localization of SLC2A9 is consistent with the Dalmatian huu phenotype and with SLC2A9 having a role in urate transport in both the liver and kidney.
Mutations in SLC2A9 will likely have important consequences for a number of different disorders of uric acid homeostasis in people. Significant alterations of SLC2A9 could cause primary renal hypouricemia in people similar to SLC22A12 mutations since genetic heterogeneity exists for this disorder [43]–[45]. Mutations in SLC2A9 could also cause cases of primary gout by significantly altering the serum uric acid concentration. In addition, more subtle changes could alter serum uric acid levels by changing the amount of uric acid excreted in the urine. There is evidence that polymorphisms in SLC22A12 are capable of increasing serum uric acid levels in Japanese populations [46]. A significant association of SLC2A9 to serum uric acid levels was recently reported among Caucasian individuals [28] and among Sardinian and Chianti cohorts [27]. Two other papers were recently published on the importance of the SLC2A9 gene to urate transport in humans. The first demonstrated that human SLC2A9 is a high capacity, low affinity uric acid transporter and that genetic variants are associated with gout; however causative mutations were not determined [30]. The second paper documented a strong association between SLC2A9 and uric acid levels in cohorts of German and Austrian nationalities as well as a correlation between RNA expression levels of this gene and serum uric acid levels [29]. Thus, strong evidence exists that SLC2A9 functions as a urate transporter in humans, and likely in dogs.
In addition to the potential influence with respect to human uric acid disorders, the present studies also impact our understanding of the history of this defect in dogs. The story of hyperuricosuria in dogs started in 1916 when Benedict first recognized the similarity of uric acid defects in the Dalmatian dog and people [11]. However, the origin of this defect may have preceded the development of the Dalmatian breed. The present studies show that the same genetic mutation is present in Bulldogs and Black Russian Terriers, breeds that are not known to be closely related to the Dalmatian. It appears that affected individuals from these breeds share the same haplotype as the Dalmatian, indicating that the mutation is identical by descent between these breeds. The mutation must be quite old since it would have to predate breed formation; however, additional evaluation of the extended haplotype in all three breeds is necessary to estimate an actual age. Alternatively, although unlikely, the mutation could have been introduced into these other breeds by crosses between breeds. Although Dalmatians are fixed for hyperuricosuria, this is not true of the other breeds. Therefore, genetic testing and selection in those breeds can eliminate the disease. Within the Dalmatian breed, the possibility exists for correction of this defect by the introduction of unaffected Dalmatian×Pointer backcross dogs into the purebred gene pool. These dogs are currently registered with the United Kennel Club in the United States. The disease allele probably became homozygous in modern Dalmatians through selection for more distinctive spots. However, most low uric acid excreting backcross dogs have acceptably sized spots (according to the breed standard), allowing breeders the unique opportunity to correct a fixed genetic defect while maintaining the breed characteristic that may ultimately be responsible for its fixation.
Two independent lines of evidence from different species point to the key role of SLC2A9 in urate transport; the Dalmatian uric acid phenotype itself and genome wide association studies linking SLC2A9 to uric acid levels in people along with direct uric acid transport data. There are many questions to be answered about the role of SLC2A9 in urate homeostasis and its other transport functions. In dogs and other mammals with endogenously low serum uric acid, SLC2A9 may normally play a different transport role than in people where it appears to have an important function in uric acid transport. Since degradation of uric acid to allantoin does not occur in humans and great apes, SLC2A9 may play a different role in transport in the liver as compared to those species that excrete allantoin. The positional cloning of the hyperuricosuria locus in the Dalmatian dog has provided a compelling new avenue of investigation toward a better understanding of urate transport in mammals and successfully completes a story started in 1916 when Benedict first recognized the similarity of the uric acid defect in the Dalmatian dog and people.
DNA samples from backcross dogs were acquired as previously described [35]. The full pedigree of the dogs used in this analysis is shown in Figure S1. Blood, buccal swab and DNA samples from Dalmatians and non-Dalmatian dogs were obtained from patients of the Veterinary Medical Teaching Hospital at UC Davis, the UC Davis Veterinary Genetics Laboratory, Dr. Gary Johnson at the University of Missouri, Columbia, and from private owners. The use of animals in this research was approved by the University of California, Davis animal care and use committee (protocol #11962). Urine uric acid and creatinine was measured in 3–7 week old puppies as previously described [13].
LOD scores were calculated as previously described [35]. Primers for microsatellites spaced on average 100 Kb apart in the 3.3 Mb candidate region of CFA03 were identified using the May 2005 CanFam2.0 sequence assembly on the UCSC genome browser and designed within sequence flanking each repeat using the Primer3 program (Table S1) [47].
Eight SNPs, which are part of the Affymetrix canine SNP array, were chosen because they are polymorphic in the general dog population. Additional SNPs were mined from the canine genome sequence or identified during sequencing of the candidate genes. Primers for SNPs (Table S2) were designed as described above. SNP and microsatellite sequences were PCR amplified and genotyped in the backcross dogs and only informative markers for this family were then genotyped in 24 unrelated purebred Dalmatians.
Human mRNA sequences were obtained from GenBank for each of the candidate genes (MIST NM_052964.1, KIAA1729 NM_053042, WDR1 NM_005112.4, SLC2A9 AF210317). These sequences were compared to the UCSC Genome Browser annotation of the canine genome using the BLAT function to obtain canine sequence for each gene. The Primer3 program was used to design primers for the canine sequences (Table S3). PCR reactions, genotyping and sequencing were done as previously described [35]. Sequences were visualized using Chromas2 (Technelysium, Tewantin, QLD, Australia) and analyzed with Vector NTI software (Informax, Frederick, MD, USA). TFsearch Version 1.3 was used to analyze the Variant O promoter SNPs affect on transcription factor binding sites.
5′ and 3′ RACE were performed for 3 of the 4 candidate genes. RACE primers (Table S3) were designed as described above and RACE products amplified with the SMART RACE cDNA Amplification Kit (Clontech, Mountain View, CA, USA) and cloned using the TOPO TA Cloning kit (pCR2.1-TOPO vector) with One Shot TOP10 Chemically Competent E. coli (Invitrogen, Carlsbad, CA, USA). Products were isolated with the Qiaprep Spin Miniprep kit (Qiagen, Valencia, CA, USA) and sequenced as described above. Genbank accession numbers for the transcripts are EU371511–EU371515. All genomic locations given in the text are based on the May 2005 CanFAm2.0 genome assembly and are viewed using the UCSC genome browser.
RNA was isolated from kidney and liver samples with the Micro-FastTrack 2.0 mRNA isolation kit (Invitrogen, Carlsbad, CA, USA). cDNA was synthesized with the SuperScript III First-Strand Synthesis System for RT-PCR (Invitrogen, Carlsbad, CA, USA). Expression was evaluated for LOC479092, LOC611070 and both variants of LOC479093 from liver and kidney from Dalmatians and unaffected Newfoundland×Border Collie crosses using primers SLC5′UTREx1VarNF and SLCR1 for variant N and primers IntF-Ex1 and SLCR1 for variant O. The expression of LOC479093 was also evaluated in an array of tissues from an unaffected Beagle (cerebellum, cerebral cortex, heart, kidney, liver, skeletal muscle, skin, spinal cord, spleen, testis and thymus) using primers IntF5 and SLCdnaR. RNAs for these tissues were acquired from Zyagen (San Diego, CA, USA) and cDNA synthesized as described. GAPDH was amplified (F primer-5′AAGATTGTCAGCAATGCCTCC3′, R primer -5′CCAGGAAATGAGCTTGACAAA3′) in these tissues to ensure that equivalent amounts of cDNA were added.
SLC2A9 genotypes were determined by a restriction fragment length polymorphism (RFLP) assay. PCR products were produced as described previously using an unlabeled forward primer, 5′-TGCTTCTCTGAAATTTACCTCCA – 3′ and a fluorescently labeled reverse primer, 5′-6FAM-CGAGAGGATGGTATACGGTGA -3′ (Applied Biosystems, Foster City, CA). Products were then digested with the enzyme HpyCH4V (New England Biolabs, Ipswitch, MA) for 1 hour at 37°C. Digestions were analyzed on an ABI 3100 Genetic Analyzer with GeneScan 400HD Rox size standard. A 79 base pair unlabeled product is generated from the 440 bp product by a control cut site. The A allele produces a labeled 361 base pair product and the G allele produces a labeled 106 base pair product.
|
10.1371/journal.ppat.1006482 | KSHV encoded ORF59 modulates histone arginine methylation of the viral genome to promote viral reactivation | Kaposi’s sarcoma associated herpesvirus (KSHV) persists in a highly-ordered chromatin structure inside latently infected cells with the majority of the viral genome having repressive marks. However, upon reactivation the viral chromatin landscape changes into ‘open’ chromatin through the involvement of lysine demethylases and methyltransferases. Besides methylation of lysine residues of histone H3, arginine methylation of histone H4 plays an important role in controlling the compactness of the chromatin. Symmetric methylation of histone H4 at arginine 3 (H4R3me2s) negatively affects the methylation of histone H3 at lysine 4 (H3K4me3), an active epigenetic mark deposited on the viral chromatin during reactivation. We identified a novel binding partner to KSHV viral DNA processivity factor, ORF59-a protein arginine methyl transferase 5 (PRMT5). PRMT5 is an arginine methyltransferase that dimethylates arginine 3 (R3) of histone H4 in a symmetric manner, one hallmark of condensed chromatin. Our ChIP-seq data of symmetrically methylated H4 arginine 3 showed a significant decrease in H4R3me2s on the viral genome of reactivated cells as compared to the latent cells. Reduction in arginine methylation correlated with the binding of ORF59 on the viral chromatin and disruption of PRMT5 from its adapter protein, COPR5 (cooperator of PRMT5). Binding of PRMT5 through COPR5 is important for symmetric methylation of H4R3 and the expression of ORF59 competitively reduces the association of PRMT5 with COPR5, leading to a reduction in PRMT5 mediated arginine methylation. This ultimately resulted in a reduced level of symmetrically methylated H4R3 and increased levels of H3K4me3 marks, contributing to the formation of an open chromatin for transcription and DNA replication. Depletion of PRMT5 levels led to a decrease in symmetric methylation and increase in viral gene transcription confirming the role of PRMT5 in viral reactivation. In conclusion, ORF59 modulates histone-modifying enzymes to alter the chromatin structure during lytic reactivation.
| Kaposi’s sarcoma-associated herpesvirus (KSHV) must carefully regulate both phases of its lifecycle in order to persist and proliferate effectively in the infected cells. In this study, we show the importance of dynamic epigenetic modifications on the viral chromatin in dictating whether KSHV displays the latent or lytic phase of its life cycle. Various chromatin-modifying enzymes are responsible for adding activating or repressive ‘marks’ on chromatin, one of these is a PRMT5 (protein arginine methyltransferase 5), which symmetrically dimethylates arginine 3 of histone H4 (H4R3me2s) and associates with condensed chromatin leading to restricted gene expression. An early lytic protein of KSHV, ORF59 associates with PRMT5 to disrupt its binding with the chromatin leading to a loss of repressive, H4R3me2s mark and corresponding gain of activating H3K4me3 during lytic reactivation.
| Kaposi’s sarcoma-associated herpesvirus (KSHV), also known as human herpesvirus 8 (HHV8), is a member of the gammaherpesvirus family that is associated with Kaposi’s sarcoma (KS), Primary Effusion Lymphoma, a subset of Multicentric Castleman’s Disease, and (in HIV-co-infected patients) KSHV Inflammatory Cytokine Syndrome [1–4]. KSHV is a double-stranded DNA virus with a large genome that encodes for over 87 open reading frames (ORFs) including genes necessary for capsid, tegument, envelope, DNA replication and regulatory proteins. KSHV undergoes a bi-phasic lifecycle, common to other herpesviruses, that features both latent and lytic modes of infection. The virus persists indefinitely in the infected host in a latent form during which time only a small fraction of regulatory viral proteins are expressed, most notably the latency-associated nuclear antigen protein [5–7]. In the latent stage, LANA regulates latent genome replication and tethers the circular viral episomes to the host chromosomes to ensure the segregation of KSHV episomes to daughter cells upon cell division [8–11] Additionally, LANA modulates several signaling pathways to suppress the host immune antiviral responses to induce cell growth and survival [12–17].
During latency, the KSHV genome is maintained primarily in a heterochromatic conformation in which the genome is highly compact with restricted transcription of the viral genes [18, 19]. Specific ‘repressive’ epigenetic marks on the viral heterochromatin that contribute to the stability and tight regulation of gene expression include trimethylation of lysines 9 (H3K9me3) and 27 (H3K27me3) on histone H3, ubiquitination of lysine 119 of histone 2A (H2AK119Ub), and CpG-methylation [20]. The compactness of KSHV chromatin during latency was confirmed by sequencing the nucleosomal depleted DNA in FAIRE (Formaldehyde-Assisted Isolation of Regulatory Elements) assays, which revealed that only a small percentage of the viral genome, primarily the latency-associated regions, were in an active chromatin (euchromatin) state [18, 21, 22].
Latent viral genomes reactivate upon transcription of viral genes in a synchronized cascade of immediate early (IE), early (E), and late (L) genes, which leads to the production of infectious virion particles. Control of lytic reactivation is governed by the presence of both activating and repressive marks on the viral chromatin [19, 23, 24]. These are particularly important for certain regulatory regions of the KSHV genome with a bivalent chromatin structure because the balance between repressive and activating marks can tilt the scale for latency or reactivation [25, 26]. For example, the promoter of immediate early gene, RTA exists in a bivalent chromatin structure as it simultaneously possesses both activating, H3K4me3 and repressive, H3K27me3 marks [22]. Balance between these epigenetic marks is determined by multiple host cellular and viral factors [19–21, 25].
There is a growing list of epigenetic marks important for regulating chromatin structure and gene regulation [27]. One of the less studied epigenetic marks is arginine methylation, which is carried out by either type I or type II protein arginine methyltransferases (PRMTs) [28, 29]. PRMT5 is a type II methyltransferase that symmetrically dimethylates arginine 3 of histone H4, H4R3me2s [30–32]. PRMT5 mediated symmetric dimethylation of H4R3 promotes tri-methylation of histone, H3 lysine 27 (H3K27me3), a repressive mark that is one hallmark of compact chromatin [33]. In addition, knockdown of a related type II PRMT (PRMT7) is associated with a reduction in symmetric methylation of H4R3; moreover, reduced H4R3me2s levels corresponded to an increase in the activating, H3K4me3 marks [34]. The catalytic domain for the methyltransferase activity of PRMT5 was identified to be in Motif I (GAGRGP), and is essential for symmetrically methylating arginine 3 of histone H4 [31, 35]. PRMT5 has multiple versatile roles in cell growth and development and its association with specific binding partners influences substrate specificity and subcellular localization [29]. In the cytoplasm, pICln associates with PRMT5 and facilitates the methylation of Sm proteins, which increases their affinity for the SMN (survival of motor neuron) protein and facilitates proper assembly of the spliceosome for the formation of snRNPs [36, 37]. Another study of co-factors influencing substrate specificity of PRMT5 by Guderian et al. showed that PRMT5 bound to pICln or RioK1 recruits and symmetrically methylates nucleolin (a RNA binding protein) [38]. One of the cofactors important for regulating the histone methylation specificity of PRMT5 is MEP50 (also known as Wdr77) [39]. In addition to MEP50, another co-factor that regulates specific methylation of H4R3 by PRMT5 is a cooperator of PRMT5 (COPR5), which was identified in a yeast two-hybrid assay [40]. Recruitment of PRMT5 through COPR5 leads to a preferential symmetric dimethylation of histone H4R3 [40]. Importantly, symmetric dimethylation of H4R3 at the chromatin is associated with a compact chromatin causing transcriptional silencing [30, 41].
ORF59 is an early viral protein expressed within the first 24h of viral reactivation and acts as a processivity factor for the viral DNA polymerase during lytic DNA replication [42–44]. ORF59 homologs from other herpesviruses including human cytomegalovirus (hCMV) and Epstein-Barr virus (EBV) ppUL44 and BMRF1, respectively, are shown to be essential for lytic DNA replication [45]. The processivity factor dimerizes in the cytoplasm, binds to the viral DNA polymerase, and translocates it into the nucleus to assemble at the origin of lytic DNA replication (OriLyt) [46–49]. We recently reported that the phosphorylation of Ser378 and Ser379 is critical for ORF59’s activity, viral DNA synthesis, and virion production [50]. Although ORF59 was classically viewed as a processivity factor, studies have identified additional functions for ORF59; for example, ORF59 hinders the non-homologous end joining (NHEJ) repair of DNA double-stranded breaks by blocking the interaction between DNA-PKcs and the Ku complex during lytic replication to promote tumorigenesis [51]. Another report showed that ORF59 interacts with poly(ADP-ribose) polymerase 1 (PARP-1) and stimulates its proteosomal degradation to block the PARP-1 mediated cell cycle control and apoptosis [52].
In this study, we identified ORF59 binding proteins by immunoprecipitating the Flag epitope tagged ORF59 in BAC16 and classified bound proteins by mass spectroscopic analysis. The assay identified few viral proteins and a large number of cellular counterparts including RNA processing proteins and chromatin modifying enzymes. Here, we determined the association and significance of PRMT5 on viral chromatin and our data show that PRMT5 depletion in latent cells can trigger the transcription of lytic genes, suggesting PRMT5’s role in altering the chromatin landscape. Depletion of PRMT5 was linked to a reduced level of symmetric dimethylation of H4R3 (repressive mark) on viral chromatin, which correlated with lower H4R3me2s amounts detected during viral reactivation. Decrease in symmetric methylation during reactivation was linked to decreased levels of PRMT5 bound to the viral chromatin that facilitated the formation of euchromatin for active gene transcription. Our results confirmed that ORF59 competitively removed PRMT5 from its linker molecule, COPR5, which alters its specificity to symmetrically methylate H4R3. Furthermore, the loss of H4R3me2s marks lead to an increase in the tri-methylation of H3K4 (H3K4me3), an activating mark, and confirmed the role of symmetric methylation in regulating chromatin landscape. Taken together, we propose a mechanism by which ORF59 disrupts PRMT5’s mediated compact chromatin leading to the formation of open chromatin important for lytic replication.
ORF59 encoded viral processivity factor (PF-8) assists viral DNA polymerase with DNA processivity during lytic DNA replication [47–49]. ORF59, an early protein of viral reactivation has also been shown to be involved in additional functions and the mutational studies determined ORF59 to be important for DNA replication and virion production [50–52]. Therefore, we were interested in identifying proteins associating with ORF59, which we achieved by generating a BACmid harboring Flag-epitope tagged ORF59 and immunoprecipitating ORF59 with anti-Flag antibody from BAC16-ORF59Flag 293L cells induced for 48h for lytic reactivation (Fig 1). Flag epitope tag was introduced at the C-terminus of ORF59 by homologous recombination (Fig 1A, 1B and 1C). BAC16 WT, intermediate containing GalK-Kan-R and the final clone with Flag ORF59 was digested with BamHI and hybridized with GalK probe to confirm the loss of GalK-KanR cassette after the negative selection (Fig 1D). These BACs, BAC16WT and BAC16-ORF59-Flag, were transfected into 293L cells to obtain clones maintaining viral genome (Fig 1E). Expression of Flag tagged ORF59 in BAC16-ORFF59, but not in BAC16WT, was confirmed by a western blot and immunoprecipitation (Fig 1F). Since BAC16-WT did not have Flag tagged ORF59, it was used as a control for anti-Flag (ORF59) immunoprecipitation. Proteins immunoprecipitated with ORF59-Flag were identified by LC/MS analysis at Mitch Hitchcock Nevada Proteomics Center, University of Nevada, Reno. A list of ORF59-specific binding proteins is shown in Tables 1 and 2 for viral and cellular proteins, respectively. This confirmed the binding of previously reported proteins and many additional uncharacterized proteins including KSHV mRNA transcript accumulation protein, ORF57 (Table 1). Although there were several proteins of interest, we focused on determining the role of protein arginine methyltransferase 5 (PRMT5) during viral reactivation due to its role in regulation of chromatin architecture (Table 2). Furthermore, 293 L cells stably expressing GFP fused ORF59 also precipitated PRMT5 identified by protein sequencing (Fig. A in S1 Text), confirming specificity of their association independent of other viral factors. PRMT5 is an arginine methyl transferase that symmetrically dimethylates the arginine 3 of histone H4 (H4R3me2s) of core histones by binding to the chromatin through a linker, COPR5, which leads to the formation of a compact chromatin structure. Binding of ORF59 with PRMT5 led us to speculate that ORF59 may disrupt PRMT5 mediated symmetric methylation of H4R3 to alter the viral chromatin for transcription and replication.
In order to confirm the binding of PRMT5 with ORF59, we performed co-immunopreciptation assays on endogenous as well as over expressed proteins. First, the ORF59-HA and PRMT5-Flag expression vectors were transfected into 293T cells and an immunoprecipitation with anti-Flag antibody for PRMT5 co-precipitated the HA-tagged ORF59 (Fig 2A, lane 4). Lack of a detectable band of ORF59-HA with control vector (Flag), confirmed that these proteins associate specifically (Fig 2A, lane 3). Furthermore, the reverse CoIP in which PRMT5-Myc with ORF59-Flag or Flag vector were expressed and immunopreciptiated with anti-flag antibody, showed specific precipitation of PRMT5-Myc (Fig 2B, lane 4). Flag Vector was unable to precipitate PRMT5-Myc (Fig 2B, lane 3), which confirmed specificity of the assays.
Next, we used KSHV-infected cell lines, TRExBCBL1-RTA and iSLK.219 to affirm this finding in an endogenous system. Each cell line was induced for lytic replication by treatment with doxycycline and ORF59 was immunoprecipitated with anti-ORF59 antibody. Immune detection of PRMT5 in ORF59 precipitated lanes showed a distinct band in both, TRExBCBL-1 and iSLK.219 cell lines and confirmed the association of these two proteins during lytic reactivation (Fig 2C and 2D, lanes 3). The control antibody, IgG did not co-precipitate any PRMT5, which again confirmed specificity of the assay (Fig 2C and 2D, lanes 2).
After we confirmed the binding between these two proteins, we were interested in determining whether they localize to the same nuclear compartment during viral reactivation. To test this, we performed immune localization of these proteins in doxycycline induced TRExBCBL1-RTA cells. These proteins were detected using respective antibodies. Both ORF59 and PRMT5 are nuclear proteins that showed nuclear localization, as expected, and many of these foci showed localization of both proteins in the same nuclear compartment (Fig 2E, merge panel). Latent (uninduced) cells do not express ORF59, therefore undetected, but the subcellular localization of PRMT5 was seen to be similar as in the induced cells (Fig 2E). This confirmed that ORF59 associates with PRMT5 in reactivated cells.
To further investigate the association between ORF59 and PRMT5, a systematic series of truncation constructs of these proteins were generated (Fig 3A and 3B). The association of ORF59 with PRMT5 was first tested by using GST fused ORF59 truncations with in vitro translated 35S-methionine labeled full-length PRMT5. GST-tagged ORF59 segments or GST-control were used to precipitate the in vitro translated PRMT5 (Fig 3C). The first segment of ORF59 (1-132aa) precipitated PRMT5 most strongly compared to the other segments indicating the segment to be the primary site of PRMT5’s binding (Fig 3C, compare lane 3 with lanes 4 and 5). The specificity of their binding was confirmed by the lack of any bound PRMT5 with GST control (Fig 3C, lane 2). The binding assay was conducted with equal amounts of in vitro translated PRMT5 represented in the input lane (Fig 3, lane1), while GST proteins used for binding are shown with coomassie staining (Fig 3C, bands marked with asterisks).
Next, we determined the domain of PRMT5 primarily responsible for its interaction with ORF59. Full-length ORF59 fused to GST (GST-59) or control GST was used to precipitate the in vitro translated PRMT5 and its truncations. Equal proportions of the in vitro translated PRMT5 and respective truncations were used in the binding assays with GST or ORF59-GST. As expected, ORF59-GST but not control GST precipitated full-length PRMT5 (Fig 3D. lane 3, PRMT5 panel). Interestingly, ORF59-GST bound most strongly with the PRMT5 210-420aa region (Fig 3D. lane 3, PRMT5-M panel). Although a small proportion of PRMT5 1-210aa also bound with ORF59-GST it is very evident that ORF59 interacts most robustly with the 210-420aa regions of PRMT5, which notably possesses the catalytic domain for methyltransferase activity.
This association was further confirmed using an over-expression system in which HEK293T cells were transfected with the various truncation mutants of PRMT5 (Flag epitope tagged) and with HA-tagged ORF59. Immunoprecipitations with anti-Flag antibody for PRMT5 truncations showed efficient binding of PRMT5 segment 210-420aa with ORF59 detected in a western blot with HA antibody (Fig 3E, lane 7). The binding of PRMT5 truncations with ORF59 showed similar pattern as detected with the in vitro translated proteins i.e. PRMT5-M (210-420aa) of PRMT5 associated most strongly with ORF59 (Fig 3E). The N-terminal domain of PRMT5 showed also showed binding, although lower compared to the middle region, which is similar to the in vitro binding data (Fig 3E, compare lane 6 and 7). The C-terminal domain of PRMT5 did not associate with ORF59 in both in vitro binding and immunoprecipitation assays (Fig 3D and 3E). Since the middle region of PRMT5 (PRMT5-M) was the primary site of interaction with ORF59, and possesses the catalytic domain required for methyltransferase activity, we wanted to determine whether the minimal catalytic domain associated with ORF59 [31]. To do so, we constructed a PRMT5 truncation containing the catalytic motif fused to GFP and HA. This was used for binding with GFP tagged ORF59-Flag or GFP-Flag (Fig 3F). Immunoprecipitation with anti-Flag antibody showed specific precipitation of catalytic domain of PRMT5 with GFP-59-Flag but not with the GFP-Flag control (Fig 3F). This confirmed that ORF59 associated with PRMT5 through its N-terminal domain (1-132aa) to the middle, catalytic region of PRMT5, possibly to interfere with the methyltransferase activity.
A balance between repressive and activating epigenetic marks on the KSHV genome has been shown to regulate the viral gene expression and control the switch between latent-lytic cycles of the virus [19–21, 25]. The immediate early gene, RTA promoter region is one example of bivalent chromatin, displaying both repressive, H3K27me3 and activating, H3K4me3 marks but upon reactivation activating marks are enriched following a decrease of the repressive marks. This is a particularly ingenious mechanism as it allows the virus to respond to environmental stimuli rapidly. PRMT5, an arginine methyltransferase responsible for the H4R3me2s modification, promotes heterochromatinization. To understand the importance of PRMT5 and the symmetric methylation of histone H4R3 during the viral life cycle, we determined the symmetric methylation levels on H4R3 of viral chromatin during latency and lytic reactivation. We immunoprecipitated symmetrically dimethylated H4R3 containing chromatin from un-induced and doxycycline induced TRExBCBL1-RTA cells. Specificity of the antibody for the symmetrically methylated form of H4R3 chromatin was confirmed by immunoprecipitation and detection with symmetric antibody, which did not cross-react with asymmetric form (Fig. B in S1 Text). DNA extracted from the bound chromatin was sequenced (ChIP-Seq) and analyzed for enriched regions (ChIP-peaks) using the ChIP analysis tool of CLC Workbench [53]. The peak score representing the relative levels of H4R3me2s containing chromatin was detected throughout the genome with enrichment at specific regions on the uninduced latent genome (Fig 4B). Interestingly, the levels of symmetrically methylated H4R3 (H4R3me2s) chromatin on the lytically reactivated genome were significantly reduced throughout the genome with certain regions totally devoid of the H4R3me2s chromatin (Fig 4B). Considering the transcriptionally restrictive nature of the latent genome in contrast to highly active viral gene transcription during reactivation, the abundance of symmetrically methylated H4R3 in latent cells and subsequent reduction in lytic cells was to be expected (Fig 4B). In addition, we tested histone H4 occupancy on viral genome by performing histone H4 ChIP-Seq on the same latent and lytic TRExBCBL1-RTA cells and analyzing for any enriched ChIP-peaks (Fig 4C). Mapping the reads of histone H4 ChIP-seq to the viral genome showed a consistent occupancy and the ChIP peak calling software did not detect peaks with significant peak core due to the uniformed presence of histone H4 throughout the genome. Importantly, the occupancy of histone H4 on the viral genome remained similar after induction suggesting that reduction in H4R3me2s during induction was not due to overall reduction of histone H4 from the viral genome. To further ensure that the changes in H4R3me2s enrichment were not due simply to changes in genome copy numbers, we tested the levels in TREx-BCBL1-RTA cells induced for lytic replication for 12h and treated with replication inhibitor, 0.5mM PFA. At 12h time point, we still detected significant reduction in H4R3me2s enrichment at indicated region of the viral genome including OriLyt, RTA promoter, K8 promoter, and ORF21 promoter (Fig. C in S1 Text).
Since PRMT5 symmetrically methylates H4R3 to make the chromatin compact, we wanted to determine whether depleting PRMT5 would be sufficient to alter the chromatin into a transcriptionally active form. To this end, we performed PRMT5 knockdown by transducing shRNA in KSHV positive, BCBL-1 cells and transfection of siRNA on iSLK.219 cells. Both shRNA and siRNA for PRMT5 significantly reduced the levels of PRMT5 compared to the shControl and scrambled siRNA (si-Cntrl), respectively (Fig 4F and 4G). The transcriptionally active nature of the chromatin was determined by quantifying the levels of viral mRNA in PRMT5 depleted cells compared to the control cells. Interestingly, BCBL-1 cells with depleted PRMT5 showed significantly higher levels of almost all the viral transcripts compared to the control cells (Fig 4D). Furthermore, PRMT5 depleted iSLK.219 cells also showed an overall increase, although lesser fold than BCBL-1 cells, in the number of viral transcripts (Fig 4E). This suggested a restrictive role of PRMT5 in regulating KSHV gene expression.
Next, we wanted to determine whether the levels of symmetrically methylated H4R3 chromatin were reduced in those PRMT5 depleted cells. To achieve this, we performed H4R3me2s ChIP on both; BCBL-1 and iSLK.219, control and PRMT5 depleted cells and quantified the symmetric methylation of H4R3 at representative viral gene promoters and genomic regions (Fig 4B). The levels of H4R3me2s in PRMT5 depleted cells were calculated relative to the control cells, which are represented by light grey bars in while the darker grey bars indicate PRMT5 knock-down cells (Fig 4H and 4I). In both the cell lines, PRMT5-depletion correlated with a significant decrease in the amount of H4R3me2s at those representative regions (Fig 4H and 4I). These results suggest that the symmetric dimethylation of H4R3 on the viral genome is dependent upon the expression of PRMT5 and depleting PRMT5 results in a loss of the H4R3me2s modification. Thus, PRMT5 knockdown of KSHV-infected cells impairs H4R3me2s while simultaneously upregulating viral gene transcription.
The reduction of H4R3me2s marks on the viral chromatin during lytic reactivation suggests a dynamic state of arginine methylation that can be altered to favor viral transcription/replication during lytic reactivation. The reduction of H4R3me2s marks at various viral promoters shown in the PRMT5 knockdown cells implies that the presence of PRMT5 on the chromatin can be correlated with the levels of H4R3me2s marks. Taking into consideration PRMT5’s association with ORF59 and its depletion leading to transcriptional activation, we compared the levels of PRMT5’s association to the viral chromatin before, and following lytic induction. To test this, TRExBCBL1-RTA cells were harvested at uninduced, 12h induced or 24h induced time points and used for PRMT5 and ORF59 ChIP-Seq analysis. (Due to the insufficient quantity of ORF59 ChIP DNA isolated from uninduced samples, ORF59 ChIP-Seq was only performed on the 12h and 24h induced cells). PRMT5 showed enriched binding across the latent viral genome, however; upon reactivation at both, 12h and 24h the peaks were diminished suggesting a lower binding of PRMT5 after induction (Fig 5C). Not surprisingly, ORF59 was enriched at numerous loci on the viral genome at both, 12h and 24h post-induction (Fig 5D). Furthermore, in previous studies we tested ORF59’s binding to chromatin in the presence of replication inhibitor, PFA. When replication was inhibited, although perhaps slightly less than control-treated cells, ORF59 still showed significant enrichment at several viral loci (Fig. C in S1 Text).
This demonstrated an important link between ORF59 binding and chromatin structure modulation. The accumulation of ORF59 yet reduction in PRMT5 binding to the viral chromatin suggested a mechanism where ORF59 is responsible for displacing PRMT5 from the chromatin. Interestingly, PRMT5 binds to the chromatin through a ligand, cooperator of PRMT5 (COPR5) [40]. COPR5 functions as a linker molecule that causes PRMT5 to preferentially modify H4R3 residues [40]. To test if PRMT5 is displaced from the chromatin by the detachment of this linker molecule, we also performed COPR5 ChIP Seq on uninduced, and 24h induced TRExBCBL1-RTA cells. In contrast to the PRMT5 ChIP Seq results, COPR5 binding to the viral chromatin remained similar after induction (Fig 5B), confirming that displacement of PRMT5 from COPR5 binding alters symmetric methylation of H4R3. Despite the fact that ORF59, PRMT5, and COPR5 are all DNA-binding proteins, ChIP-seq performed with control-IgG on uninduced, 12h induced, and 24h induced cells did not show any particular peak on the viral genome confirming the specificity of these ligands in chromatin immunoprecipitation (Fig 5E).
To study the mechanism by which ORF59 could deplete PRMT5 binding from the viral genome in more detail, we tested the binding between ORF59 and the PRMT5-chromatin linker molecule, COPR5. To this end, we created constructs to mimic previously described regions of COPR5: 1-140aa, and 141-184aa. The first larger segment (1-140aa) corresponds to the linker-function of COPR5 and associates with histones whereas the smaller C-terminal truncation of COPR5 (141-184aa) is important for binding to PRMT5. Displacement of PRMT5 from the chromatin prompted us to test whether ORF59 was disrupting the association of PRMT5 with its linker, COPR5. Thus, we performed an in vitro binding assay using GST-fused ORF59 protein with 35S-methionine labeled, in vitro-translated COPR5. First, we determined the binding region of ORF59 by using ORF59 full length and its truncations used previously (Fig 6A). Equal amounts of in vitro-translated COPR5 were added to the binding reaction with each of the GST constructs and a representative amount of input is shown (Fig 6B, lane 1). While control-GST showed no association with COPR5 (Fig 6B, lane 2), ORF59 full length showed strong binding (Fig 6B, lane 3). Notably, ORF59-1, ORF59-2 showed some binding but segment 3 (ORF59-3) did not show any binding (Fig 6B, compare lanes 4, 5 with 6). Next, we determined the ORF59 binding domain in COPR5 by using in vitro translated COPR5 and its truncations (Fig 6C) with full-length ORF59-GST. Binding was compared with their respective inputs (Fig 6D, lanes 1–3). ORF59-GST interacted with full-length COPR5 and remarkably only the C-terminal region of COPR5 141-184aa (Fig 6D, lanes 7 and 9). The two proteins associated specifically as the control GST did not show bindings (Fig 6D lanes 4–6). The binding of ORF59 to the same domain of COPR5 required for recruiting PRMT5 (141-184aa), but not to the histone-linking region (1-140aa), suggested a mechanism in which ORF59 disrupts the binding of PRMT5 bound to COPR5. This was tested with competitive Co-IPs by expressing ORF59 in cells with COPR5 and PRMT5. In other words, the ability of PRMT5 to bind COPR5 (and vice-versa) was analyzed in the presence or absence of ORF59. In the first set, immunoprecipitation of PRMT5-Myc confirmed its association with COPR5 (Fig 6E, lane 5). However, the presence of ORF59-HA reduced the amounts of co-precipitating COPR5 with PRMT5 (Fig 6E, IB:Flag panel, compare lane 6 with lane 5). The levels of COPR5 in the lysates were comparable (Fig 6E, IB:Flag panel-input lanes, compare lanes 1–3).
We also attempted a reverse co-IP with COPR5 for assaying its binding to PRMT5 in the presence of ORF59 (Fig 6F). Detection of PRMT5 immunoprecipitating with COPR5 confirmed their binding (Fig 6F, lane 5, IB:Myc panel). Presence of ORF59 decreased COPR5’s binding with PRMT5 (Fig 6F, IB:myc, compare lanes 5 and 6). These co-precipitations demonstrate that expression of ORF59 competitively disrupts PRMT5’s association with COPR5 to assist in facilitating structural chromatin changes that favor lytic replication.
ORF59 is an essential protein for lytic viral replication; however, most of the studies were done in presence of a robust viral transcactivator protein, RTA, which is necessary and sufficient to facilitate lytic reactivation [54, 55]. Therefore, we wanted to investigate the role of ORF59 in modulating chromatin structure in cells lacking this overarching factor, iSLKTet-RTA-Bac16-RTASTOP [56].
iSLKTet-RTA-Bac16-RTASTOP cells were transiently transfected with ORF59 for the detection of chromatin modifications using ChIP assays. Expression of ORF59 and absence of RTA in iSLKTet-RTA-Bac16-RTASTOP cells was confirmed by immune detection (Fig 7A). We then determined the enrichment of ORF59 on the viral chromatin across a number of important loci including OriLyt region and the RTA promoter region (Fig 7B). The amounts of ORF59 bound to the viral promoters were calculated relative to the control-plasmid transfected iSLKTet-RTA-Bac16-RTASTOP cells. The data showed varying levels of ORF59 binding despite the absence of RTA on the representative regions of the viral genome (Fig 7A). Next, we used the same cells to determine the association of PRMT5 to the viral chromatin. This revealed a consistent pattern of lower PRMT5 bound at those representative sites (Fig 7C, darker bars represent PRMT5 binding in presence ORF59). Strikingly, the expression of ORF59 in those iSLKTet-RTA-Bac16-RTASTOP cells was enough to trigger significant reductions of H4R3me2s marks on the viral chromatin at various viral promoter regions and the OriLyt (Fig 7D). In fact, the reduction in symmetrically methylated H4R3 due to ORF59 expression was fairly comparable to those obtained from iSLKTet-RTA-Bac16RTASTOP cells induced for lytic replication with doxycycline for RTA expression (Fig 7F). These cells showed a robust expression of RTA and ORF59 (Fig 7E). These results suggest a model in which PRMT5 binds to the latent viral chromatin to symmetrically methylate the H4R3 but the expression of ORF59 represses H4R3me2s levels by altering PRMT5’s association with COPR5 in order to alter the chromatin landscape (Fig 7G).
The change in H4R3me2s marks and expression of viral transcripts due to PRMT5 depletion and ORF59 expression prompted us to evaluate whether reduction in the levels of H4R3me2s leads to an alteration of other epigenetic marks on the chromatin favoring lytic replication. To this end, we evaluated the levels of H3K4me3, a well-known activating mark on the viral chromatin, previously shown to be enriched during lytic reactivation [20, 22, 57]. Interestingly, studies have shown that the presence of H4R3me2s mark inhibits efficient tri-methylation of H3K4 residues to preserve a heterochromatic landscape [34, 58]. Our results confirmed a decrease of these prohibitive H4R3me2s marks at various viral promoters in the presence of ORF59, therefore it was necessary to test whether this led to an increase in the abundance of activating, H3K4me3 marks on those sites. To this end, chromatin immunoprecipitation with anti-H3K4me3 antibody was performed on cells with a series of different cellular/viral conditions. iSLKTet-RTA-Bac16-RTASTOP cells were induced for RTA expression by the addition of doxycycline for the precipitation of viral chromatin bound to H3K4me3. Relative levels of H3K4me3 bound chromatin after lytic induction (Fig 8B, dark bars) were calculated by normalizing with the levels in uninduced cells (Fig 8B, dark bars) (a relative fold change of 1 represents the H3K4me3 levels on viral genome before lytic induction (Fig 8B)). Consistent with previous findings [20, 22, 57, 59], lytic induction increased the abundance of activating, H3K4me3 marks on the viral chromatin, shown on the represented targets (Fig 8B). In addition, we used the same cell line, iSLKTet-RTA-Bac16-RTASTOP for detecting the role of ORF59 on H3K4 methylation levels in absence of RTA. The expression of ORF59 in these iSLKTet-RTA-Bac16-RTASTOP cells was confirmed by immune detection (Fig 8A). Chromatin bound to H3K4me3 histone showed a significant enrichment at various viral promoters in cells expressing ORF59 as compared to the vector-transfected cells (Fig 8C), which corroborated with reduction in the levels of H4R3me2s marks. These results confirmed that ORF59 by itself was capable of inducing changes in chromatin structure resembling those that occur during lytic reactivation. Next, we wanted to determine the levels of H3K4me3 bound chromatin in cells depleted with PRMT5, which showed a significant decrease in levels of H4R3me2s bound chromatin. Not surprisingly, the levels of H3K4me3 bound chromatin were increased in PRMT5 depleted iSLK.219 cells as compared to the control cells (Fig 8D). Similarly, BCBL-1 cells depleted for PRMT5 by shRNA showed significant increase in H3K4me3 bound chromatin compared to the control cells (Fig 8E). This was consistent with the previous observation that H4R3me2s has an inhibitory effect on H3K4me3 and removal H4R3me2s leads to an enrichment of active chromatin mark, H3K4me3 [34, 58, 60].
ORF59 has been previously characterized to be essential for efficient viral DNA replication, but given the role we describe here in regards to chromatin structure modulation, we wanted to investigate the effects of ORF59 protein on the transcription of viral genes. To this end, we utilized previously described KSHV Bacmid deleted with ORF59 (Bac36Δ59) in 293L cells [50]. We harvested RNA from wild type (293L-Bac36WT) and ORF59 deleted (293LBac36Δ59) cells transiently transfected with RTA for the induction of lytic cascade. The levels of RTA were comparable between these two sets, RTA transcripts were 505 and 510 folds in wt and ORF59 deleted cells, respectively (Fig 9). The levels of the viral genes transcripts in 293L-Bac36WT with RTA cells were calculated by comparing with the control vector transfected, 293L-Bac36WT cells. Similarly, the viral genes transcripts in 293LBac36Δ59 expressing RTA were calculated by taking vector transfected, 293LBac36Δ59 cells, as control. The relative fold change of viral genes with RTA from 293L-Bac36WT and 293LBac36Δ59 were plotted and showed that in cells lacking ORF59, the mRNA copies of many viral genes were reduced (Fig 9). We analyzed the effects of ORF59 on the expression of immediate early, early, and late genes. Immediate early (IE) genes, displayed in green, showed consistently higher levels of transcripts in the Bac36WT as compared to the cells deleted for ORF59 (Fig 9, Dark green-Bac36WT and light green-ORF59 deleted). A handful of early genes, shown in blue shade, also showed dependence on ORF59 for their expression and these include ORF59: ORF4, ORF18, ORF34, ORF38, and ORF47 (Fig 9, Dark blue- Bac36WT and light blue- Bac36Δ59). Interestingly, a greater number of late gene transcripts, represented by the red bars (Fig 9, Dark red-Bac36WT and light red-ORF59 deleted) were affected by the depletion of ORF59, suggesting that ORF59 plays an important role in controlling viral chromatin landscape for gene transcription.
ORF59 encoded processivity factor is one of the viral lytic proteins that helps the viral DNA polymerase with processivity activity and is essential for the productive replication of the viral genome [42, 44, 46, 48, 50]. We previously showed that phosphorylation of ORF59 with viral kinase is essential for the processivity function and virion production [50]. ORF59 was classically viewed as DNA replication protein but it has been shown to regulate additional processes involved in promoting lytic reactivation. These include, binding to cellular helicases, Ku70/Ku86 to impair non-homologous end joining of replicated DNA [51], and binding and degrading PARP-1 to alleviate the repressive effects of PARP-1 on viral lytic replication [52]. These additional functions prompted us to identify the proteins interacting with ORF59 during lytic replication to have a clear understanding of its role in the viral life cycle. Using a Flag epitope tagged version of ORF59 in BAC16 was advantageous for identifying the proteins specifically associating with ORF59, including a large number of viral and cellular proteins. Among them we observed a chromatin modifying protein, PRMT5, as a significant ORF59 binding protein because ORF59 is detected early during reactivation and is critical for processes involved in DNA replication [24, 49, 50, 61, 62].
PRMT5, an arginine methyltransferase, was characterized as a transcriptional repressor bound to the chromatin of a promoter region in conjunction with multiple repressive marks including hypoacetylated H3 and H4 and lysine methylated H3K9 residue [41]. Additional reports confirmed PRMT5 to be a chromatin binding protein specifically on a transcriptionally repressive region in cooperation with other repressive complexes or transcription factors including Blimp1, Snail, BRG1 and hBRM [63–66]. Our immunoprecipitation and localization studies revealed that ORF59 and PRMT5 interact in the nucleus of KSHV infected cells during lytic reactivation, which was consistent throughout multiple cell lines. After confirming the association between ORF59 and PRMT5, we postulated that ORF59, being a protein required for lytic DNA replication, which in turn occurs on open chromatin, might alter the repressive functions of PRMT5. To better understand the nature of the interaction of these two proteins, we determined the domains required for their association in in vitro binding assays. The N-terminal domain of ORF59, 1-132aa interacted most strongly with PRMT5. Interestingly, the N-terminal domain of ORF59 is the domain through which ORF59 homodimerizes and interacts with the viral DNA polymerase [44, 46, 47, 49]. On the other hand, the interaction domain of PRMT5 that binds to ORF59 mapped to its middle domain from residues 210-420aa. The middle segment of PRMT5 possesses the methyltransferase activity, which prompted us to generate a clone containing the catalytic domain of the middle segment for assaying its binding with ORF59. Intriguingly, ORF59 was able to interact with this small domain of PRMT5, indicating that ORF59 could be involved in modulating the methyltransferase activity of PRMT5.
A previous report showed that arginine methyltransferases could associate with viral proteins to methylate them and alter their function [67]. KSHV latent protein, LANA associates with an arginine methyltransferase type I, PRMT1, which methylates arginine residues in an asymmetric fashion [67–69]. This demonstrated that KSHV proteins are not only associated with cellular arginine methyltransferases, but are indeed post-translationally modified by them to alter specific functionality. This begged the obvious question of whether ORF59 is arginine-methylated by associating with PRMT5. To investigate this, we used a methylation prediction software, MASA (Methylation site based on the Accessible Surface Area) available at http://masa.mbc.nctu.edu.tw/predict.php [70], which found that only a single arginine, R384 could potentially be methylated (Fig. D in S1 Text) suggesting that its association with PRMT5 may be required for its function. However, we did not pursue the methylation of ORF59 in this report [71].
As mentioned earlier, PRMT5 symmetrically dimethylates the H4R3 residues of histone H4 to form a transcriptionally repressive chromatin, so we determined the levels of H4R3me2s marks on latent viral genome by performing a sequencing of the H4R3me2s bound DNA. This showed a significant enrichment throughout the latent genome while the levels of symmetrically methylated H4R3 on the viral chromatin were reduced upon reactivation. This coupled with the similar levels of H4 occupancy on the viral genome confirmed that histone arginine methylation is differentially regulated during different phases of the viral life cycle.
The dynamic nature of the chromatin landscape of the viral genome facilitates rapid change for controlling the KSHV lifecycle phases. The viral genome persists in a highly ordered chromatinized episome within the host cells and the associated chromatin is subject to modifications for the concealment or exposure of particular genes for transcription as needed during latency or the lytic reactivation cascade [20, 21, 25, 26]. The availability of genes for transcription is regulated by the presence or absence of a variety of multiple histone-tail modifications, including H4R3me2s [28]. To date, most of the reports on H4R3me2s methylation show its association with transcriptionally repressed gene [60, 72]. However, one study found PRMT5 to be over-expressed in transformed chronic lymphocytic leukemia (B-CLL) cells with elevated levels of global H4R3me2s marks [73]. Interestingly, this enrichment in H4R3me2s also resulted in the transcriptional repression and subsequent downregulation of a tumor suppressor family of genes, pRB [73]. Another study conducted using EBV positive cells as a model tested the effects of PRMT5 mediated gene repression [74]. They used a PRMT5 inhibitor to determine the importance of repressive epigenetic marks in context of EBV tumors and concluded that PRMT5 was critical for B cell transformation and malignancies because the overexpression of PRMT5 helped to silence the tumor suppressor genes [74]. Another report generated a sophisticated in silico modeling procedure to analyze the ChIP-Seq data from 20 different histone methylations concluded that in the context of simultaneous activating and repressive marks, H4R3me2s is one of the most abundant repressive marks associated with gene silencing [60]. Thus, after confirming the association between ORF59 and PRMT5, the reduction of H4R3me2s marks (that PRMT5 is responsible for bestowing) on the viral genome during lytic reactivation appeared in agreement with its role and we elucidated a mechanism by which ORF59 might be directly or indirectly facilitating these changes to the viral chromatin.
It was interesting to note that the binding of ORF59 to the viral genomic regions reduced the levels of chromatin bound PRMT5, which was primarily due to a change in specificity of PRMT5 with its linker COPR5 [40]. PRMT5’s binding specificity to the chromatin is determined in large by the co-factors it associates with and we found that ORF59 disrupted the association between PRMT5 and the linker molecule COPR5 that tethers PRMT5 preferentially to the chromatin for symmetrically methylating the H4R3 residue [40]. In essence, presence of ORF59, which expresses during lytic reactivation, disrupts the binding of PRMT5 to the chromatin by competitively causing its detachment from its linker molecule, COPR5. Notably, the binding of COPR5 to the viral chromatin did not appear to change significantly which in conjunction with the fact that ORF59 did not associate with the histone H4 binding region of COPR5 1-140aa, suggests that ORF59 does not disrupt that function but rather blocks PRMT5’s specific affinity for histone H4. Moreover, ORF59 appears to be directly involved in regulating these lytic reactivation-permissive changes to the chromatin as the KSHV genome lacking immediate early genes (RTA-stop cells) showed similar reduction in chromatin bound PRMT5 and H4R3me2s marks at the viral promoters when supplemented with ORF59.
The reduction of the levels of PRMT5 bound to the chromatin led to a reduction in the symmetrically methylated H4R3 and gene activation. Similarly, it was previously seen in cancer cells in which treatment with a drug, AS1411 (a quadruplex-forming oligonucleotide aptamer) led to the decreased association of PRMT5 with known gene promoters (cyclin E2, tumor suppressor genes)[29, 75]. As a result of this depletion, PRMT5 regulated genes regained their activity [75]. Thus, the presence or absence of PRMT5 on the chromatin can be linked to the levels of repressive marks and transcriptional activity, even on the KSHV genome.
Cross-talk between different histone modifications to epigenetically regulate gene expression has been extensively studied in recent years and although still imperfectly understood, it is known that the relative presence or absence of modifications at different residues of the histone tails influences precise transcription patterns [76]. Indeed, our experiments suggest that there is a cross talk between repressive H4R3me2s marks and activating H3K4me3 marks. When H4R3me2s was reduced at various viral promoters (due to PRMT5 depletion or ORF59 overexpression), an enrichment of H3K4me3 at the same promoters was detected (Fig 8). This corroborated with a previous report, which showed that H4R3me2s sterically inhibits H3K4 methyltransferase activity to promote gene silencing [34]. This study on an analogous Type II arginine methyltransferase (PRMT7) implicated the presence or absence of methyltransferase in the prevalence and levels of the corresponding histone marks [34]. Yao et al. determined that overexpression of type II arginine methyltransferase, PRMT7 in cancer cells inhibited the expression of a specific gene promoter by causing an elevation of H4R3me2s marks, shown in conjunction with reduced H3K4me3, H4ac, and H3ac [58]. Furthermore, knockdown of PRMT7 led to a restoration of gene expression by repressing the levels of H4R3me2s and increasing the H3K4me3 and acetylation of H4 [58]. This strongly confirms that regulating the activity of arginine methyltransferases can cause epigenetic changes to influence gene expression.
Other KSHV proteins are capable of modulating viral chromatin during infection, which includes LANA and RTA [19]. RTA has been previously shown to not only autoactivate its own promoter, but also transactivate other viral gene promoters including (but not limited to) vIL-6, PAN (polyadenylated RNA), ORF57, and ORF59 [77–80]. Due to the known robust effects of RTA, which is also present during viral reactivation, we wanted to determine the impact of ORF59 on KSHV chromatin structure and viral gene expression. Our data showed that ORF59 overexpression, independent of RTA, led to a reduction of PRMT5 binding, loss of H4R3me2s marks, and enrichment of H3K4me3 marks at various viral promoters signifying the role ORF59 in these alterations. The downstream effects of changes to the chromatin structure were evident by the modulated transcription of the lytic viral genes, meaning those structural chromatin changes orchestrated by ORF59 ultimately affect lytic gene transcription. Importantly, the transcription of the lytic viral genes in the absence of ORF59 was not as efficient in ORF59-depleted cells as compared to the wild-type cells. Therefore, this study clearly shows that the lytic protein ORF59 has multiple functions during viral reactivation to facilitate efficient gene transcription for viral replication.
293T and 293L (ATCC, Manassas, VA) cells were grown in high-glucose Dulbecco's modified Eagle's medium (DMEM) supplemented with 8% bovine growth serum (HyClone, Logan, UT), 2 mM l-glutamine, 25 U/ml penicillin, and 25 μg/ml streptomycin. Additionally, 293Ls harboring any of the following: BAC16WT, BAC16 ORF59-Flag, BAC16RTASTOP (generous gift from Dr. Jae Jung), BAC36WT, BAC36Δ50, or BAC36ΔORF59, were cultured in above DMEM supplemented with 50μg/mL hygromycin B. iSLK.219 (generous gift from Dr. Don Ganem), iSLKTet-RTABAC16WT, and iSLKTet-RTABAC16RTASTOP cells (generous gift from Dr. Jae Jung) were maintained in DMEM supplemented with 10% Tet-Free Fetal Bovine Serum with additional 600μg/mL hygromycin B, 400μg/mL G418, 1μg/mL puromycin. iSLK.219 cells with recombinant KSHV BACs were induced by doxycycline.
TRExBCBL1-RTA (generous gift from Dr. Jae Jung) and BCBL-1 (ATCC, Manassas, VA), and cells were grown in Roswell Park Memorial Institute medium (RPMI) supplemented with 8% fetal bovine serum (HyClone, Logan, UT), 2 mM l-glutamine, 25 U/ml penicillin, and 25 μg/ml streptomycin. BCBL-1 transduced with shRNA PRMT5 lentiviral vectors (GE Dharmacon, Lafayette, CO) were selected and maintained on 1μg/mL puromycin. BACmid-containing 293L cells were transduced with indicated lentiviral vectors and maintained on 1μg/mL puromycin. Constituently expressing ORF59-Flag, and dsRedORF59-Flag in 293L cells were generated using a lentivirus system and ORF59-Flag expressing BACmid. pLVX-ORF59-Flag or pLVXdsRed-ORF59-Flag lentiviral vectors were generated by introducing the gene into respective vectors, and transfecting into 293T cells along with the packaging vectors (CMV-dR8.2, pCMV-VSVG) (Addgene, Cambridge, MA) for producing virions. Respective vector was also transfected into 293T cells for producing vector control lentivirus. Collected viruses were added to the target cells for transduction followed by selection with puromycin (2 mg/ml) to obtain a pure population of cells. BAC16 ORF59-Flag was transfected into 293L cells with Metafectene Pro (Biontex Laboratories GmbH, San Diego, CA) as previously described [50] followed by selection with hygromycin to obtain cells maintaining the KSHV BACs. The selection in both cell lines was monitored with GFP or RFP signals encoded by the lentivirus and the BACmid. All cultures were incubated at 37°C in a humidified environment supplemented with 5% CO2.
The following plasmids were generated by PCR amplification and cloning: pLVXORF59-Flag, pLVXdsRedORF59-Flag, pA3M-PRMT5-Myc, pA3F-PRMT5-Flag, pxiORF59-HA, pA3M-COPR5-Myc, pA3F-COPR5-Flag, pLVX-RTA, pGex-ORF59-GST. The following plasmids were then sub-cloned: pA3F-PRMT51-210aa-Flag, pA3F-PRMT5210-420aa-Flag, pA3F-PRMT5420-637aa-Flag, pA3F-COPR51-140aa-Flag, pA3fCOPR5141-184aa-Flag, pGex-ORF591-132aa-GST, pGex-ORF59133-264aa-GST, pGex-ORF59265-396aa-GST. pGIPz shRNA PRMT5 and Control shRNA lentiviral vectors were obtained from commercial source (Thermo Scientific Inc.). Packaging lentiviral vectors were obtained from Addgene.
PRMT5, COPR5 and their truncations were generated by PCR amplification using specific primers listed in Table C in S1 Text and cloning into respective vectors. The integrity of clones was confirmed by sequencing at Nevada, Genomics Center, Reno.
To generate a recombinant KSHV BAC16 with ORF59-Flag tagged, the epitope tag was inserted at the C-terminus of ORF59 by homologous recombination using a two-step “seam-less” galK positive/counter selection scheme. First, the Galk-KanR cassette with homologous flanking sequence was PCR amplified using primers with 50bp-homologus sequences (before and after the target sequence) in the sense and anti-sense primers. Galk-KanR cassette was amplified by including 20-nt of homology to the Galk-KanR region. Primers used for Galk-KanR cassette insertion (bold case is homologous to ORF59 locus, lower case is homologous to the Galk-KanR plasmid):
Forward primer: 5’-GATCGTGGGAAGGTGCCCAAAACCACATTTAACCCCCTGATTGACTACAAAGACGATGACGACAAGTGAcctgttgacaattaatcatc-3’,
Reverse primer: 5’-CTGAAGAGCGACAGAGCGCGCTCACTGTCCAGGCGGCACATGGTGctcagcaaaagttcgattta-3’.
The PCR product containing the target sequence was subjected to DpnI digestion, followed by agarose gel purification to remove any residual template plasmid. PCR product was then electroporated into competent E. coli strain, SW102 containing BAC16. The Galk-KanR cassette containing mutants were selected on chloramphenicol/kanamycin agar plates and correct insertional mutants were confirmed by restriction digestion with BamHI and Southern blot analysis of fragment containing the GalK cassette. The GalK-KanR cassette was replaced by electroporating a double stranded oligo (5’-ACATTTAACCCCCTGATTGACTACAAAGACGATGACGACAAGTGACACCATGTGCCGCCTGGACAGTGAGCGCGCTCTGTCGCTCTTCAG-3’) with homology to the flanking site and plating the bacteria on 2-deoxy galactose (DOG) containing agar plates for counter selection. Correct colonies were screened and subjected to confirmation by southern blot analysis, PCR amplification of the junctions and sequence analysis.
The following antibodies were used: mouse anti-Flag (M2, Sigma-Aldrich, St. Louis, MO), rabbit anti-Flag (F7425, Sigma-Aldrich, St. Louis), mouse anti-RTA (mouse hybridoma), mouse anti-LANA (mouse hybridoma), mouse anti-Myc (mouse hybridoma), rabbit anti-HA (6908, Sigma-Aldrich, St. Louis, MO), mouse anti-HA12CA5 (sc-57592, Santa Cruz Biotechnology), mouse anti-GFP (G1546, Sigma-Aldrich, St. Louis), mouse anti-GAPDH (G8140, US Biological, Salem MA), and rabbit anti-Myc (SAB4300605, Sigma-Aldrich, St. Louis, MO), goat anti-PRMT5 (C-20, sc-22132, Santa Cruz Biotechnology), mouse anti-ORF59 (generous gift from Dr. Bala Chandran), rabbit anti-Control IgG (sc-2027, Santa Cruz Biotechnology), mouse anti-Control IgG (sc-2025, Santa Cruz Biotechnology), rabbit anti-H4R3me2s (61187, Active Motif, Carlsbad CA), rabbit anti-Histone H4 (#61299 Active Motif), rabbit-anti-PRMT5 antibody (#61001 Active Motif), rabbit anti-control IgG (ChIP grade—Cell Signaling Technology #2729, rabbit anti-COPR5 antibody (Novus Biologicals #NBP2-30884), and additional rabbit-anti ORF59 antibody custom synthesized for our lab by GenScript.
For overexpression experiments, 293T cells were plated to 70–80% confluency followed by transfecting them with expression vectors by combining PEI, transfection reagent and 150mM NaCl, mixing thoroughly and incubating it for 15 minutes at room temperature. The mixtures were then added onto 70–80% confluent 293T cells and incubated for 6 h at 37°C with 5% CO2 before changing the medium to remove PEI. Transfected cells were harvested after 48 h post-transfection for immunoprecipitation assays. Harvested cells were washed with ice-cold PBS and lysed in 0.5 ml ice-cold RIPA buffer (1% Nonidet P-40 [NP-40], 50 mM Tris [pH 7.5], 1 mM EDTA [pH 8.0], 150 mM NaCl), supplemented with protease inhibitors (1 mM phenylmethylsulfonyl fluoride, 1 μg/ml aprotinin, 1 μg/ml pepstatin, 1μg/mL sodium fluoride, and 1 μg/ml leupeptin). Cell debris were removed by centrifugation at 13,000×g (10 min and 4°C), and lysates were then precleared for 1h with rotation at 4°C with 30 μl of Protein A-Protein G-conjugated Sepharose beads. Approximately, 5% of the lysates were saved for input control and remaining was added with 1.0 μg of indicated antibodies to capture the protein by rotating overnight at 4°C. Immune complexes were captured with 30 μl of Protein A-Protein G-conjugated Sepharose beads with rotation for 2h at 4°C. The beads were pelleted and washed three times with RIPA buffer. Input lysates and the immunoprecipitated complexes were boiled for 5–7 min in Laemmli buffer, resolved on SDS-PAGE and transferred onto nitrocellulose membrane (Bio-Rad Laboratories). The membranes were incubated with appropriate antibodies followed by detection with infrared-dyes tagged secondary antibodies and imaged on an Odyssey imager (LICOR Inc., Lincoln, NE).
Chromatin immunoprecipitation was performed as described previously [16]. Briefly, approximately 20 million cells were cross-linked with 1% formaldehyde for 10 min at room temperature, followed by addition of 125 mM glycine to stop the cross-linking reaction. Cells were washed with cold PBS containing protease inhibitors (1 μg/ml leupeptin, 1 μg/ml aprotinin, 1μg/mL sodium fluoride, 1 μg/ml pepstatin, and 1 mM phenylmethylsulfonyl fluoride). Cells were resuspended in 1 ml cell lysis buffer [5 mM piperazine-N, N′-bis (2-ethanesulfonic acid) (PIPES)-KOH (pH 8.0)-85 mM KCl-0.5% NP-40] containing protease inhibitors, incubated on ice for 10 min followed by centrifugation at 2,500 rpm for 5 min at 4°C to collect the nuclei. Nuclei were resuspended in nuclear lysis buffer (50 mM Tris [pH 8.0]-10 mM EDTA-1% SDS containing protease inhibitors), followed by incubation on ice for 10 min. Chromatin was sonicated to an average length of 700 bp followed by removing the cell debris by centrifugation at 13,000 rpm for 10 min at 4°C. The supernatant containing sonicated chromatin was diluted with ChIP buffer (0.01% SDS-1.0% Triton X-100-1.2 mM EDTA-16.7 mM Tris [pH 8.1]-167 mM NaCl including protease inhibitor). Samples were precleared with a salmon sperm DNA-protein A-protein G Sepharose slurry for 1h at 4°C with constant rotation. The supernatant was collected after a brief centrifugation (2,000 rpm at 4°C). Ten percent of the supernatant was saved for input and the remaining 90% was used for capturing chromatin by rotating the complexes overnight at 4°C using indicated antibodies. The antibody bound chromatin was precipitated by protein A/G slurry. Beads were then washed sequentially three times with a low-salt buffer (0.1% SDS-1.0% Triton X-100-2 mM EDTA-20 mM Tris [pH 8.1]-150 mM NaCl), and twice in Tris-EDTA. Chromatin was eluted in an elution buffer (1% SDS-0.1 M NaHCO3) and reverse cross-linked by adding 0.3 M NaCl at 65°C overnight. Eluted DNA was precipitated, treated with proteinase K at 45°C for 2h and purified with Qiagen Min-Elute PCR purification columns. Purified DNA was used as a template for qPCR amplification of indicated regions of KSHV genome using primers listed in Table B in S1 Text.
We performed LowCell ChIP using the LowCell ChIP Kit (Diagenode Inc.). Briefly, 1 million cells were fixed as described above followed by a PBS wash and re-suspending them in the kit-provided chromatin-shearing buffer. Chromatin was sonicated to an average size of 500bp using the Bioruptor Pico (Diagendode Inc.) and chromatin with specific antibodies were precipitated as described above.
TRExBCBL1-RTA cells were used for ChIP-seq assay by precipitating the chromatin from both un-induced and doxycycline induced (12h or 24h) cells with indicated antibodies. DNA extracted from the chromatin bound to respective targets and their appropriate controls was used for preparing the DNA sequencing libraries using NEXTflexTM Illumina ChIP-Seq Library Preparation Kit (Bioo Scientific Inc. TX, USA). Mature libraries were analyzed for purity and quantification using Bioanalyzer 2100 (Agilent Technologies, Inc.) and kappa library quantitation kit (Kappa Biosystems, Inc.), respectively. Sequencing of these libraries was done on Illumina NextSeq500 (Illumina Inc.) and the sequences were analyzed using CLC Workbench 10.0.1 (Qiagen Inc.). Briefly, the sequences obtained from input, control IgG antibody and specific antibody samples were mapped to the KSHV genome (Accession number GSE98058, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE98058) using ‘map reads to the reference’ tool of the CLC Workbench 10.0.1. Mapped reads were analyzed for enriched peaks in ChIP samples, represented as peak score, with respect to input reads using the ‘ChIP-seq’ tool of the CLC Workbench 10.0.1 with the minimum peak-calling P-value set to 0.05 [53].
SmartPool siRNA for control or PRMT5 (Dharmacon, GE Life Sciences) was transfected into iSLK.219 cells using RNAiMax transfection procedure (Lipofectamine, Thermofisher Inc.). Lipofectamine RNAiMax reagent was diluted in Opti-MEM medium and combined with siRNA diluted in Opti-MEM medium and incubated 5 minutes at room temperature. siRNA-lipid complex was added to cells at a 30pmol final amount of siRNA per well of 6-well plate and incubated 96 h before harvesting.
pGIPz shPRMT5 and pGIPzshControl (Dharmacon, GE Life Sciences) vectors were transfected with packaging plasmids into 293Lenti-X T cells (Clontech Laboratories, Inc.) and induced with NaB for 12 hours. Supernatant was collected every 12 hours over the next 3 days and lentiviral particles were concentrated by centrifugation. Lentivirus was added to BCBL-1 cells treated with polybrene and the transduction efficiencies were monitored by the visualization of GFP. Transduced cells were selected with 1μg/mL puromycin and the knockdown efficiencies of both methods were tested with Western blotting and comparative qPCR. qPCR analysis was performed by calculating the relative fold change values against the control knock-down samples and then the fold change values from 3 replicates were averaged and displayed in the graphs Fig 4D and 4E with standard deviation represented by error bars.
293L with BAC36WT, BAC36Δ50, BAC36Δ59, BAC16WT, or BAC16RTASTOP were transiently transfected with pLVXdsRedORF59-Flag, pLVX-RTA, or pLVXdsRed, vector control using PEI transfection reagent and cells were harvested 36 hours post transfection. Additionally, above cell lines were lentivirally transduced with either pLVXdsRedORF59-Flag or vector control and the expression was confirmed by Western blot analysis. PRMT5 knock-down cell samples (detailed above) were also subjected to transcriptome analysis by real-time qPCR analysis. Briefly, approximately 2 million cells were harvested and RNA was isolated using GE RNA Spin Kit, samples were eluted in 40μ RNAse-free water. cDNA was generated using High-Capacity RNA-to-cDNA kit (Applied Biosystems Inc.) according to manufacturer’s instructions. qPCR was performed using cDNA as template for amplifying viral ORF targets listed in Table A in S1 Text. Target genes were quantitatively assessed by comparative CT values and normalized with untreated/control samples. Fold changes were calculated using ΔΔCt method and the error bars represent standard deviation of three experiment replicates.
GST proteins were purified as previously described[50]. Briefly, induced BL-21 bacterial culture with indicated GST-fusion plasmids were harvested and the purified proteins were collected on Glutathione beads. Aliqots were taken and resolved by SDS-PAGE and coomassie staining to estimate relative quantity.
TNT-T7 Quick Coupled Transcription/Translation System (Promega Inc.) was used to generate 35S-methionine labeled proteins according to manufacturer’s instructions. Briefly, reaction components were thawed on ice and combined with plasmid DNA template (5μg) and 2μl [35S]-methionine followed by incubation for 3 h at 30°C. Expression was confirmed by resolving small aliquots of translated protein on a SDS-PAGE and exposing to autoradiography screens.
In vitro translated [35S] methionine labeled proteins were first pre-cleared by rotating at 4°C for 30 minutes with Control-GST protein beads. Pre-cleared samples were then combined with equal amounts of respective GST-fusion proteins and the final volume of the binding mixture was brought up to 700 μl with in vitro binding buffer (1XPBS, 10% glycerol, 0.1%NP-40) supplemented with 1μM DTT and protease inhibitors. Samples were rotated 4°C overnight then washed 3 times with binding buffer before being resuspended in 1X PAGE Buffer. All samples, including inputs were then incubated at 95°C for 5 minutes before being resolved by SDS-PAGE and imaged by autoradiography.
The next generation sequence data of ChIP-seq are deposited to NCBI genbank with accession number Series GSE98058, that includes subseries GSE98057, GSE98087, GSE100044.
All the data were analyzed using Prism 6 software (Graphpad Inc.) for significance.
|
10.1371/journal.pgen.1004557 | Differential Management of the Replication Terminus Regions of the Two Vibrio cholerae Chromosomes during Cell Division | The replication terminus region (Ter) of the unique chromosome of most bacteria locates at mid-cell at the time of cell division. In several species, this localization participates in the necessary coordination between chromosome segregation and cell division, notably for the selection of the division site, the licensing of the division machinery assembly and the correct alignment of chromosome dimer resolution sites. The genome of Vibrio cholerae, the agent of the deadly human disease cholera, is divided into two chromosomes, chrI and chrII. Previous fluorescent microscopy observations suggested that although the Ter regions of chrI and chrII replicate at the same time, chrII sister termini separated before cell division whereas chrI sister termini were maintained together at mid-cell, which raised questions on the management of the two chromosomes during cell division. Here, we simultaneously visualized the location of the dimer resolution locus of each of the two chromosomes. Our results confirm the late and early separation of chrI and chrII Ter sisters, respectively. They further suggest that the MatP/matS macrodomain organization system specifically delays chrI Ter sister separation. However, TerI loci remain in the vicinity of the cell centre in the absence of MatP and a genetic assay specifically designed to monitor the relative frequency of sister chromatid contacts during constriction suggest that they keep colliding together until the very end of cell division. In contrast, we found that even though it is not able to impede the separation of chrII Ter sisters before septation, the MatP/matS macrodomain organization system restricts their movement within the cell and permits their frequent interaction during septum constriction.
| The genome of Vibrio cholerae is divided into two circular chromosomes, chrI and chrII. ChrII is derived from a horizontally acquired mega-plasmid, which raised questions on the necessary coordination of the processes that ensure its segregation with the cell division cycle. Here, we show that the MatP/matS macrodomain organization system impedes the separation of sister copies of the terminus region of chrI before the initiation of septum constriction. In its absence, however, chrI sister termini remain sufficiently close to mid-cell to be processed by the FtsK cell division translocase. In contrast, we show that MatP cannot impede the separation of chrII sister termini before constriction. However, it restricts their movements within the cell, which allows for their processing by FtsK at the time of cell division. These results suggest that multiple redundant factors, including MatP in the enterobacteriaceae and the Vibrios, ensure that sister copies of the terminus region of bacterial chromosomes remain sufficiently close to mid-cell to be processed by FtsK.
| Most bacteria harbour a single chromosome and, in the rare case in which the genetic material is divided on several chromosomes, the extra-numerous ones appear to have derived from horizontally acquired mega-plasmids that subsequently gained essential genes [1]. This is notably the case for Vibrio cholerae, the agent of the deadly human diarrheal disease cholera, whose genome is divided between a 2.961 Mbp ancestral chromosome, chrI, and a 1.072 Mbp plasmid-derived chromosome, chrII [2]. The preferential transcription of chrII genes during colon colonization compared to in vitro growth under aerobic conditions suggests that this genomic organization is important for rapid adaptation to different environments [3]. Likewise, other bacteria harbouring multipartite genomes can adopt several different life cycles [4], [5], [6], [7]: the rhyzobium, the burkholderia and the vibrio, can alternatively spread freely in the environment or interact as symbionts or pathogens with eukaryotic cells; the borrelia are obligate parasites that need to infect several different eukaryotic organisms in the course of their life cycle. Thus, multipartite genomes seem to offer a selective advantage for the adaptation to very different environmental conditions. However, the necessary coordination between replication, chromosome segregation and cell scission raises questions on the management of the different chromosomes of such bacteria.
Bacterial chromosomes harbour a single origin of bidirectional replication and are generally circular. Replication ends in a region opposite of the origin of replication, the terminus region, in which is usually found a specific recombination site dedicated to the resolution of chromosome dimers, dif [8]. Fluorescent microscopic observation of chromosome segregation in mono-chromosomal bacteria revealed that it is concurrent with replication and starts with the active positioning of sister copies of the origin region into opposite cell halves [9], [10], [11]. As replication progresses along the left and right chromosomal arms, newly replicated loci are progressively segregated towards their future daughter cell positions. However, the mean time during which sister loci remain together before separation is variable [12]. In particular, sister copies of the terminus region co-localize at mid-cell until the initiation of cell division in E. coli and P. aeruginosa [9], [10], [13], [14]. This mode of segregation can participate in the coordination between chromosome segregation and cell division. Indeed, nucleoid occlusion factors impede the assembly of the cell division machinery until a time when the only genomic DNA left at mid-cell consist of the sister copies of the terminus region in Escherichia coli and Bacillus subtilis [15], [16]; the long co-localization of sister termini at mid-cell is at least in part dictated by the MatP/matS macrodomain organisation system in E. coli [17], [18]; a DNA translocase, FtsK, which is recruited to mid-cell as part of the divisome and which pumps chromosomal DNA in the orientation dictated by repeated polar motifs that point towards dif, the KOPS, promotes the orderly segregation of the DNA within the terminus region of E. coli chromosome [13], [19], [20]. One of the functions of FtsK is to control the resolution of chromosome dimers, which result from homologous recombination events between circular sister chromatids, by the addition of a cross-over between sister dif sites at the time of constriction [21]. FtsK is also thought to participate in sister chromatid decatenation [22], [23] and to create a checkpoint to delay constriction until sister terminus regions have been fully segregated [19], [24], [25].
V. cholerae chrI and chrII are circular and harbour a single dif site in the region opposite of their origin of replication, dif1 and dif2, respectively (Figure 1A, [26]). Segregation of the two chromosomes is concurrent with replication and both chromosomes adopt a longitudinal organization within the cell [27]. However, chrII is replicated late in the C period of the cell cycle, when most of chrI has been replicated, and the initiation of its segregation is consequently delayed [28]. In addition, the origin region of chrI, OriI, locates to the old pole of newborn cells and one OriI sister migrates to the other pole after replication (Figure 1A, [27]). The origin region of chrII, OriII, locates to mid-cell in newborn cells and the two OriII sisters migrate towards the ¼ and ¾ positions after replication (Figure 1A, [27]). This is at least in part dictated by the presence of a partition machinery of a chromosomal type on chrI, parABS1, and a partition system that groups with plasmid and phage machineries on chrII, parABS2 (Figure 1A, [27], [29], [30]). The last chromosomal regions to be segregated are the terminus regions of chrI and chrII, TerI and TerII, respectively (Figure 1A, [27]). Both TerI and TerII locate at or close to the new pole in newborn cells (Figure 1A, [27]). Replication termination of the two V. cholerae chromosomes is synchronous [28] and unreplicated TerI and TerII are recruited to mid-cell at approximately the same time (Figure 1A, [27]). V. cholerae is closely related to E. coli in the phylogenetic tree of bacteria and its genome harbour the same dam co-occurring DNA maintenance machineries as E. coli [31]. This includes a unique E. coli MatP ortholog and the presence of cognate matS sites in both TerI and TerII. In addition, a common pair of tyrosine recombinases, XerC and XerD, serves to resolve dimers of each of the two V. cholerae chromosomes despite the sequence divergence of dif1 and dif2 [26]. Dimer resolution is controlled by a unique E. coli FtsK ortholog, whose translocation activity is oriented by KOPS motifs that point towards the dimer resolution site of each of the two chromosomes [26]. By analogy to E. coli, MatP is thought to maintain sister copies of TerI and TerII at mid-cell and FtsK to promote the orderly segregation of the DNA within TerI and TerII. Correspondingly, the separation of sister copies of a locus situated at 40 kbp from dif1 seemed coordinated with cell division (Figure 1A, [27], [32]). However, sister copies of a locus situated at 49 kbp from dif2 separated before cell division, which questioned the role of FtsK and MatP on TerII segregation (Figure 1A, [32]).
The aim of this work was to identify the contribution of MatP to the segregation dynamics of TerI and TerII. We show by replication profiling that dif1 and dif2 are located next to the replication terminus of chrI and chrII, respectively. Simultaneous visualization of the positions of dif1 and dif2 within the cell then allowed us to confirm the late and early separation of TerI and TerII, respectively. However, we show that TerII sisters keep colliding with each other at mid-cell during constriction by genetically probing the relative frequency of sister chromatid contacts occurring at mid-cell at the time of cell division along the two chromosomes and by time-lapse fluorescent microscopy. We further show that the frequency of these collisions depends on the MatP/matS macrodomain organization system, possibly because it restricts the movements of TerII within the cell. We also show that MatP promotes the late mid-cell co-localization of TerI sisters. However, TerI loci remain in the vicinity of the cell centre and sister chromatid contacts remain frequent in its absence.
Replication profiling of V. cholerae cells by deep sequencing indicated that termination most frequently occurred at a distance of ∼90 kbp and ∼70 kbp from the reference loci that had been used by Srivastava et al. for the simultaneous visualization of the positions of TerI and TerII (Figure S1A, [32]). It was therefore possible that the behaviour of these loci did not fully reflect TerI and TerII segregation dynamics. To confirm the segregation pattern of the terminus regions of chrI and chrII, we simultaneously visualized the intracellular location of dif1 and dif2 in cells that were exponentially growing in minimal media. We used the lacO/LacI-mCherry system to label the dif1 locus and the pMT1 parS/yGFP-ParB system to label the dif2 locus. Cells were classified according to their length in bins of 0.25 µm. They had a median length of 3.2 µm (Figure S2A). The smallest cells, i.e. the youngest cells, contained a single dif1 spot at one of the two cell poles (Figure 1B). This pole, which results from the previous division event, is hereafter referred to as the new pole. The preferential localization of dif1 towards the new pole was used to orientate the cells. A single dif2 spot was also observed in the youngest cells (Figure 1B). This spot was located in the younger cell half, at an intermediate position between the dif1 spot and the middle of the cell (Figure 1B). The polarity of the dif1 and dif2 spots decreased as a function of cell elongation and the median position of each spot reached mid-cell in cells of an intermediate length (Figure 1B). The majority of the longest cells, i.e. the closest to cell division, displayed a single dif1 spot, which was located at mid-cell and was flanked by two dif2 spots (Figure 1B). Indeed, <15% of the cells from the 4.25 µm–4.5 µm bin displayed two dif1 spots whereas >80% of them displayed two dif2 spots (Figure 1C). In addition, the proportion of cells containing two dif2 spots reached 100% in the cells that were longer than 4.5 µm whereas only 50% of these cells displayed two dif1 spots (Figure 1C, grey points). Marker frequency analysis indicated that the earlier timing of appearance of cells with two dif2 foci was not due to an earlier timing of replication of dif2 compared to dif1 (Figure S1A). The same pattern of segregation was observed when the dif1 and dif2 labelling systems were switched, excluding any possible artefact linked to the visualization strategy (Figure S3). Finally, dif2 sisters were found to segregate further away from each other and from mid-cell than dif1 sisters (Figure 1G). Taken together, these results suggest that in the vast majority of cases TerII sisters separated before cell division whereas TerI sister separation was delayed until the end of cell division.
We next investigated the influence of the MatP/matS macrodomain organization system on TerI and TerII segregation. V. cholerae cells in which MatP was disrupted were slightly longer than wild-type cells. In minimal medium, they had a median length of 3.77 µm (Figure S2B). Nevertheless, growth competition indicated that they lost less than 0.23% of fitness per generation (Figure S4). The smallest cells had a single dif1 and a single dif2 spot, which were both positioned closer to mid-cell than in wild-type cells (Figure 1D). This was accompanied by an increase in position variability (Figure 1D). As a consequence, mid-cell recruitment was no longer directly observable in cells of intermediate lengths (Figure 1D). In addition, the timing of separation of dif1 spots was now very similar to the timing of separation of dif2 spots (Figure 1E). Marker frequency analysis indicated that this was not due to a change in the relative replication timing of dif1 and dif2 (Figure S1B). Many cells of intermediate length now displayed two dif1 and two dif2 spots and most of the cells of the following bins had two dif1 and two dif2 spots (Figure 1E). This was directly reflected in the proportion of cells displaying a single dif1 spot and a single dif2 spot and the proportion of cells with two dif1 and two dif2 spots in the entire population (Figure 1F, number of spots). The separation of dif2 sisters remained slightly ahead of the separation of dif1 sisters (Figure 1E), which was reflected in the higher proportion of cells harbouring a single dif1 spot and two dif2 spots than cells harbouring a single dif2 spot and two dif1 spots (Figure 1F, 3 spots disposition). However, the disposition of spots became more random and many cells now displayed dif2 spots more centrally located than dif1 spots (Figure 1F, 3 spots disposition and 4 spots disposition). Finally, sister dif sites migrated to opposite cell halves after their separation (Figure 1D) and the distances between the sisters of both sites were similar (Figure 1G). Taken together, these results suggested that MatP contributed to the precise positioning of TerI before and after replication and that it delayed the separation of TerI sisters to the time of cell division. MatP also contributed to the precise positioning of TerII. However, it was unable to impede TerII sisters from separating before septum constriction.
As the densities of matS sites in TerI and TerII are very similar, we were intrigued by the apparent inability of MatP to block TerII sister separation. Lesterlin et al. designed an assay based on the interruption of the lacZ reporter gene by two copies of loxP to detect sister chromatid contacts (SCC) behind replication forks [33]. The assay was based on the proximity of the loxP sites: the cleavage points of the Cre recombinases on each strand of the tandem sites were separated by only 55 bp to prevent intra-molecular recombination. As a result, a functional lacZ ORF could only be reconstructed via intermolecular recombination events (Figure 2A). As dif-recombination is under the control of FtsK in V. cholerae [26], which was expected to restrict it to mid-cell and to the time of septum constriction [21], we reasoned that 55 bp dif-cassettes could be used to monitor the proximity of TerI and TerII sisters to the cell division machinery at the time of constriction (Figure 2B).
We engineered a strain in which XerC production was under the control of the arabinose promoter to permit the stable inheritance of dif-cassettes. To help repress any leaky XerC production, we inserted the E. coli lacZ promoter and the E. coli lacI repressor gene in anti-orientation at the end of the xerC ORF. We also replaced the ATG translation initiation codon by the less favourable TTG codon and removed the ribosomal binding site (Figure 2B).
The dif sites harboured by the first and second chromosomes of the El Tor N16961 strain, dif1 and dif2, possess divergent overlap regions (Figure 2C, [26], [34]). To compare the excision of 55 bp dif1- and dif2-cassettes (lac2dif1 and lac2dif2), we inserted them at the same genomic position, in place of the dif locus of chromosome II, and monitored the frequency of full blue colonies that were obtained three hours after the induction of XerC production (Figure 2C). Recombination worked well for both dif sites (Figure 2D). In both cases, blue colony formation strictly depended on XerC production and on the presence of a fully functional ftsK allele (Figure 2D).
Little or no recombination can occur between dif1 and dif2 thanks to their sequence divergence (Figure 2C). The use of lac2dif2 on chrI and lac2dif1 on chrII thus prevented any risk of Xer-mediated intrachromosomal rearrangements due to recombination between the dif sites of the cassette and the dimer resolution site of the chromosome during the course of the experiment (Figure 2E). Therefore, the dimer resolution site of the chromosome could be left, which avoided any artefact in the measured excision frequencies linked to the formation of chromosome dimers by recombination between sister copies of the cassettes (Figure S5). The dif sites of the cassettes used on each of the two V. cholerae chromosomes are identical to the dimer resolution site of the other chromosome. However, this site did not influence the proportion of blue colonies that were formed (Figure 2E and Figure S6).
Both intramolecular and intermolecular recombination events can generate single dif site products. In contrast, three dif site products can only be generated via intermolecular recombination. Such products are transient because they can be converted to single dif products by subsequent intramolecular recombination (Figure 2A). Nevertheless, we could detect their appearance with 55 bp cassettes, demonstrating that recombination occurred via SCC (Figure 3A). As a point of comparison, we engineered 1 kbp dif-cassettes, a distance sufficient for intramolecular recombination. With such cassettes, we did not observe any intermolecular recombination intermediates, suggesting that 1 kbp cassette excision mainly resulted from intramolecular recombination events on separate chromatids (Figure 3A).
FtsK-YFP localized to mid-cell in long cells (Figure 3B, white arrow) and at one of the two poles in short cells (Figure 3B, white arrow head). This was reminiscent of the pattern of localization of the cell division machinery of Caulobacter crescentus, which assembles at mid-cell but remains bound to the new pole after cell scission [35]. Time-lapse observations confirmed that such a scenario applied to V. cholerae FtsK, demonstrating that it assembled at mid-cell as part of the cell division machinery (Figure S7A). In addition, treating cells with cephalexin, which blocks septum constriction, led to a dramatic reduction in the level of dif-recombination without affecting the recruitment of FtsK to the cell division apparatus (Figure 3C). No loss of cell viability was observed during the course of the cephalexin treatment (Figure S7B). We conclude that dif-recombination occurs during or shortly after septum constriction in V. cholerae.
Finally, deletion of recA did not affect the proportion of excision events that could be detected using 55 bp- and 1 kbp-cassettes, indicating that activation of dif-recombination was independent from chromosome dimer formation in V. cholerae (Figure 3D). This result is strikingly different from what is observed using dif-cassettes in E. coli [36], [37]. The reasons for this difference are the subject of another study (Gally, Midonet, Demarre and Barre, unpublished results).
Taken together, these results demonstrate that the proportion of blue colonies formed following lac2dif1 and lac2dif2 recombination events can be used as a relative measure of the respective frequency of contacts between monomeric sister chromatids that occur at mid-cell at the time of septum constriction in V. cholerae.
Cells in which lac2dif2 were inserted in the immediate vicinity of dif1 yielded a high level (∼60%) of blue colonies, demonstrating dif1 SCC during constriction (Figure 4A), in agreement with the co-localization of dif1 sisters (Figure 1). However, interchromatid recombination dropped rapidly when lac2dif2 was not in the immediate vicinity of the dif1 locus (Figure 4A). The frequency of blue colony formation did not diminish in cells in which recA was deleted, confirming that 55 bp cassette recombination on chrI was not restricted to chromosome dimers (Figure S8A).
Strikingly, we obtained a very high proportion of blue colonies (∼90%) when lac2dif1 was inserted at dif2 (Figure 4B) despite the apparent early separation of dif2 sisters (Figure 1). In addition, blue colony formation remained high (∼45%) within a 160 kbp region surrounding dif2, from a position at 9 kb on the left of the dif locus to 152 kb on the right of it (Figure 4B). The same results were obtained after recA deletion, confirming that TerII SCCs were unlikely due to chromosome dimers (Figure S8A). Taken together those results suggested that dif2 sisters contacted each other at mid-cell at the time of cell division as frequently as dif1 sisters, despite their apparent early separation.
On chrII, the extent of the region displaying a high frequency of SCC at the time of septum constriction corresponded to the putative MatP domain (Figure 4B). The only notable exception was next to a matS site that is isolated from the rest of the matS region by the V. cholerae superintegron (Figure 4B). Correspondingly, we observed more than a 4-fold reduction in blue colony formation within TerII upon matP disruption (Figure 4C). Indeed, dif2 was the only locus where cassette excision remained above the background level (Figure 4C). Cassette excision remained independent from chromosome dimer formation (Figure S8B). In contrast, the disruption of matP only had a very modest, albeit significant, effect on SCCs within TerI (Figure 4C). The remaining SCCs were still independent from homologous recombination (Figure S8B). Correspondingly, SCCs occurred in a much smaller region than the putative MatP domain on chrI (Figure 4B). Taken together, these results suggested that MatP was the main contributor to TerII SCC occurring at mid-cell at the time of cell division.
The high frequency of SCCs detected at dif2 with our genetic assay suggested that dif2 sisters frequently collided at mid-cell during septum constriction despite their early separation. To directly demonstrate that such collisions occurred, we followed the segregation dynamics of dif2 sisters by time-lapse fluorescence microscopy. We expected collisions to be transient because two dif2 spots were observed in almost all of the wild type cells longer than 4.5 µm (Figure 1 and Figure S9A). Therefore, we reasoned that short time intervals had to be used between each image acquisition. However, a balance had to be achieved between the detection of the supposedly transient dif2 collisions and the fraction of the cell cycle during which dif2 spots could be tracked in any given cell due to photobleaching. With 30 s time intervals, dif2 foci could be observed for 100 min.
A total of 74 wild-type cells were followed, out of which 44 showed a complete cell division event. In 42 of these cells, i.e. in ∼95% of the observed cell division events, dif2 sisters separated before septum invagination, in agreement with our snapshot analysis. However, dif2 sister collisions were frequent (Figure 5A and Movie S1). As a result, dif2 sisters were found to co-localize at mid-cell at some stage of the cell constriction process in 70% of the cells, which fits with the high frequency of dif2 SCCs observed with the genetic assay (Figure 5A and Movie S1). On average, 3.2 collisions were observed after the initial separation of the dif2 sisters and before cell fission. In the majority of cases, re-joining of the dif2 sisters was transient, i.e. co-localization was only observed during 2 consecutive frames. In some instances, however, dif2 sisters remained co-localized for several minutes.
We also followed 131 matP− cells, out of which 30 displayed a complete analysable cell division event. In all of these cells, dif2 sisters separated before septum invagination (Figure 5A and Movie S2). The positions of the two dif2 sisters were no longer restricted to the ¼–¾ cell region and, in several cases, one of the two dif2 spots located near the old pole at the time of division (Figure 5A and Movie S2). Indeed, only 0.6 collisions were observed on average in each cell after the initial separation of the dif2 sisters and before cell fission. These events lasted for a single frame in the vast majority of cases. Finally, co-localization of the dif2 sisters during septum constriction was only observed once, which fits with the loss of dif2 SCCs monitored with the genetic assay.
Taken together, these results suggested that MatP allowed FtsK to process dif2 sisters during cell division by restricting the range of their movements to the ¼–¾ cell region and that other factors played a similar role for dif1 sisters in its absence.
In the present study, we investigated the positioning of the replication terminus regions of the two V. cholerae chromosomes using a combination of two techniques. On one hand, we directly visualized the positioning of the chromosome dimer resolution locus of each of the two chromosomes by snapshot and time-lapse microscopy (Figure 1 and Figure 5). On the other hand, we monitored the proximity of these loci using Xer recombination between sister dif sites as a genetic reporter (Figure 4). The requirement for a direct contact between the Xer recombinases and the FtsK cell division protein for recombination to occur ensured that sister sites were recombined at mid-cell (Figure 2 and 3). The requirement for constriction initiation ensured that they were recombined at the time of septum constriction (Figure 3). The high frequency of intermolecular recombination events at the chromosome I dif locus (Figure 4) could be due to the late separation of the two sister chromatids at this particular location (Figure 1 and Figure 5). Intermolecular recombination events at the chromosome II dif locus could also happen before the sisters segregated, i.e. after dif duplication but before replication completion and/or when separated sisters were still trapped together by catenation links. However, this can only account for a limited of number of recombination events since snapshot analysis and time-lapse microscopy suggested that sisters of the chromosome II dif locus separated before septum constriction in ∼95% of the dividing cells (Figure 1 and Figure 5). Thus, the high frequency of recombination events between chromosome II dif sisters is probably mainly linked to collisions events that occurred their initial separation (Figure 4).
Possibly the most striking observation of our study was that TerII sisters kept colliding against each other at mid-cell after their initial separation in the cell cycle, up to and after the initiation of the constriction process (Figure 4 and 5). During the three hours of our genetic assays, cells underwent ∼8 divisions, as judged by the number of colony forming units at the beginning and at the end of the experiments. Therefore, the ∼90% frequency of blue colony formation that we observed with a recombination inserted at dif2 corresponded to a rate of 25% of β-galactosidase+ cell formation per generation. As only one out of the two possible intermolecular recombination events could yield β-galactosidase+ cells (Figure 2A), this result suggested that >50% of SCC occurred between TerII sisters during each cell division event (Figure 4). Moreover, we observed the same frequency of blue colony formation with the lacZdif2 probe when it was inserted at the dif2 locus on chrII (Figure 2D, lac2dif2) and when it was inserted at the dif1 locus on chrI (Figure 4B, dif1 locus), suggesting that SCCs at cell division were as frequent within TerII as within TerI. Accordingly, frequent collisions of dif2 sisters were observed at the time of cell division when following the growth of individual cells by fluorescence microscopy with 30 s time intervals (Figure 5). Interchromatid recombination events during constriction were observed in a specific 160 kb region of chrII, which corresponded to the putative MatP domain of the chromosome (Figure 4). The relative frequency of interchromatid recombination curve consisted of a plateau with a central peak at the dif2 locus (Figure 4). Our results suggested the plateau was due to the action of the MatP/matS system (Figure 4).
Our snapshot analysis of the positioning of dif1 in wild type and matP− cells indicated that MatP was a major contributor to the organization and management of TerI at the time of cell division, as observed in E. coli (Figure 1). However, the relative frequency of interchromatid recombination curve on chrI simply consisted of a sharp peak centred on dif1 with no plateau in the MatP region (Figure 4). In addition, the relative frequency of SCCs was not dramatically affected in matP− cells (Figure 4). This is in sharp contrast to what we could have expected based on the role of MatP in the formation of a FtsK loading region in E. coli [13]. Taken together, these observations suggest that other factors than MatP contribute to the management of dif1 sisters at the time of cell division, which partially masked its action in our genetic assay. We are currently investigating the relative contribution of likely candidates for TerI mid-cell localization using the power of our SCC assay. We think that these factors might be common to other bacteria in which sister copies of the terminus regions remain at mid-cell for a long period during cell division, such as P. aeruginosa and C. crescentus. However, they could not, or might not yet, be adapted to the management of the recently acquired chrII of V. cholerae. As a result, the MatP/matS system was left as the sole contributor for TerII SCCs during cell division, which helped reveal its action.
The disruption of matP had a profound impact on the subcellular localization of dif1 and dif2 (Figure 1). In particular, MatP seemed to impede the separation of dif1 sisters until cell division (Figure 1). MatP is able to create bridges between two matS sites [38]. However, we do not think that sister chromatids are tethered together by such bridges since MatP did not impede the separation of dif2 sisters (Figure 1). Careful analysis of the location of dif1 and dif2 spots in wild type and matP− cells rather suggested that MatP helped create a molecular leash that confined Ter regions in the ¼–¾ portion of the cell: even though the median positions of dif2 sisters in the cell population indicated their separation before cell division, they did not migrate very far apart from each other and away from mid-cell (Figure 1). In particular, ∼90% of dif2 spots were located at a distance of less than a quarter of the cell length in cells longer than 4.5 µm (Figure S9A). Results from our genetic assay suggested that the movements of such sister loci around the median position probably allowed for their frequent collision at mid-cell at the time of cell division. Even though their medians were equivalent, the distributions of the distances between dif2 sisters in wild type and matP− cells were markedly different (Figure 1G). Indeed, in matP− cells longer than 4.5 µm, only ∼57% of the dif2 spots remained in the ¼–¾ portion of the cell (Figure S9B). This might be sufficient to explain a large drop in sister collisions. In contrast, ∼83% of dif1 spots remained in the ¼–¾ portion of the cell, which might explain the low impact of the matP disruption on the frequency of SCC (Figure S9C and S9D). Further work will be required to investigate the molecular nature of the MatP leash. An attractive possibility would be that MatP restrains the movement of catenation loops between the two circular chromatid sisters by binding together the matS sites of each sister chromatid.
Our results suggest that multiple redundant factors, including MatP in the enterobacteriaceae and the Vibrios, ensure that sister copies of the terminus region of bacterial chromosomes remain sufficiently close to mid-cell to be processed by FtsK. In this regard, it is remarkable to observe that, even though initiation of chrII replication responds to the same global cell cycle regulatory networks than chrI initiation [39], it occurs at a later time point in the cell cycle [28], which results in synchronous chrI and chrII replication termination (Figure S1, [28]). This is likely to participate in delaying TerII sister separation until the time of cell division. We observed that matP− cells were longer than wild type cells in agreement with the notion that coordination of cell division and chromosome segregation is a key feature of the bacterial cell cycle (Figure S1). What is the functional role of this coordination? The late segregation of the terminus region might facilitate the action of FtsK on unresolved catenation links or chromosome dimers. Under laboratory conditions, we did not observe any significant chromosome dimer resolution defect (Figure S4). However, these results have to be interpreted with caution since the disorganization induced by the absence of MatP should only slightly delay the time required for FtsK to bring together sister dif sites.
Genetic engineering methods are described in Text S1. Bacterial strains and plasmids used in this study are listed in Tables S1 and S2, respectively. All V. cholerae strains were derivatives of the El Tor N16961 strain.
A lacO array was inserted adjacent to dif1 and a PMT1 parS was inserted adjacent to dif2. LacIE.coli-mCherry and yGFP-ParBpMT1 were produced via the leaky expression of a synthetic operon under the E. coli lacZ promoter that was inserted at the V. cholerae lacZ locus. A C-terminal fusion between FtsK and a yellow fluorescent protein, FtsK-YFP, was inserted in place of the endogenous V. cholerae ftsK allele to visualize its localisation. Protocols for Microscopy are detailed in Text S1. The snapshot images were analysed using the Matlab-based sofware MicrobeTracker [40], [41]. Details for the analysis are described in [27].
For bright field (BF) and fluorescence microscopy 2 µl of an exponentially growing culture sample were placed on a microscope slide coated with a thin agarose layer (1%) made using the growth medium. The slide was incubated at 30°C during the images acquisition. The images were acquired with an Evolve 512 EMCCD camera attached to an Axio Observe spinning disk from Zeiss and recorded every 30 seconds with step size of 0.4 µm in the Z-axis (3 images were acquired for each channel). The BF image 3 is subtracted from the BF image 1 to obtain the phase image.
Blue colony formation assay: 0.2 mM IPTG were used to repress xerC transcription. 0.1% arabinose was used to produce XerC. Freshly grown cultures were diluted in 5 mL of LB supplemented with arabinose to reach 0.02 of optical density at 600 nm. They were incubated for 180 mn at 37°C with shaking. Serial dilutions of the cells were plated on LB agar plates supplemented with X-gal and IPTG before and after the induction of recombination.
Southern blot assay: Cephalexin was added at the final concentration of 10 µg/ml at the same time as the arabinose. Cells were collected at the beginning of the incubation and after 40, 80 and 120 mn for genomic DNA extraction. Recombination products were analysed an EcoRV/HphI digest and 1 kbp fragment corresponding to the lacZ promoter as a probe. Signals were detected using a Typhoon instrument and quantified using the IQT 7.0 software (GE Healthcare).
|
10.1371/journal.pgen.1004152 | Distinct Requirements for Cranial Ectoderm and Mesenchyme-Derived Wnts in Specification and Differentiation of Osteoblast and Dermal Progenitors | The cranial bones and dermis differentiate from mesenchyme beneath the surface ectoderm. Fate selection in cranial mesenchyme requires the canonical Wnt effector molecule β-catenin, but the relative contribution of Wnt ligand sources in this process remains unknown. Here we show Wnt ligands are expressed in cranial surface ectoderm and underlying supraorbital mesenchyme during dermal and osteoblast fate selection. Using conditional genetics, we eliminate secretion of all Wnt ligands from cranial surface ectoderm or undifferentiated mesenchyme, to uncover distinct roles for ectoderm- and mesenchyme-derived Wnts. Ectoderm Wnt ligands induce osteoblast and dermal fibroblast progenitor specification while initiating expression of a subset of mesenchymal Wnts. Mesenchyme Wnt ligands are subsequently essential during differentiation of dermal and osteoblast progenitors. Finally, ectoderm-derived Wnt ligands provide an inductive cue to the cranial mesenchyme for the fate selection of dermal fibroblast and osteoblast lineages. Thus two sources of Wnt ligands perform distinct functions during osteoblast and dermal fibroblast formation.
| Craniofacial abnormalities are relatively common congenital birth defects, and the Wnt signaling pathway and its effectors have key roles in craniofacial development. Wntless/Gpr177 is required for the efficient secretion of all Wnt ligands and maps to a region that contains SNPs strongly associated with reduced bone mass, and heterozygous deletion is associated with facial dysmorphology. Here we test the role of specific sources of secreted Wnt proteins during early stages of craniofacial development and obtained dramatic craniofacial anomalies. We found that the overlying cranial surface ectoderm Wnts generate an instructive cue of Wnt signaling for skull bone and skin cell fate selection and transcription of additional Wnts in the underlying mesenchyme. Once initiated, mesenchymal Wnts may maintain Wnt signal transduction and function in an autocrine manner during differentiation of skull bones and skin. These results highlight how Wnt ligands from two specific tissue sources are integrated for normal craniofacial patterning and can contribute to complex craniofacial abnormalities.
| The bones of the skull vault develop in close contact with the embryonic skin to enclose the brain. In the mouse embryo, both bone-forming osteoblasts and skin-forming dermal fibroblasts are derived from cranial neural crest and paraxial mesoderm [1]. At E11.5, cranial dermal fibroblast progenitors undergo specification beneath the surface ectoderm while osteoblast progenitors are specified in a deeper layer of cranial mesenchyme above the eye [2]–[4]. Subsequently, osteoblast progenitors proliferate and migrate apically beneath the dermal progenitors [1], [4]. Both cell types secrete collagen as extracellular matrix, but skull bones provide physical protection for the brain, while the overlying dermis lends integrity to the skin and houses the epidermal appendages [5].
Both paracrine and autocrine intercellular signals function in early bone and skin development. In craniofacial bone formation the mesenchyme sets the timing of ossification [6], [7], while the surface ectoderm functions in a permissive manner [8]. Likewise, during skin formation ectodermal signals are essential for formation of the trunk hair-follicle forming dermis [9], [10], but the cranial dermal mesenchyme determines epidermal appendage identity such as hair or feather [11]. Further delineation of specific ectoderm-mesenchyme signaling during early development of the bone and dermis is required to overcome challenges in the engineering of replacement connective tissues.
Mesenchymal canonical Wnt/β-catenin signal transduction is essential in the specification and morphogenesis of both craniofacial dermis and bone [2], [3], [12]–[15], and dysregulation in components of such signaling pathways is associated with diseases of bone and skin [1], [2], [16]–[18]. Wntless (Wls) functions specifically in trafficking of Wnt ligands and is required for the efficient secretion of Wnt ligands. [2]–[4], [19]–[28]. Genetic deletion of Wls in mice is likely to dramatically reduce the levels of active Wnt ligands and can recapitulate phenotypes obtained by genetic ablation of Wnt ligands in mice [1], [4], [29]. Wnt ligand binding to target cell surface receptors (Fzd and LRP5/6) results in nuclear translocation of β-catenin, which binds to TCF/LEF transcription factors and activates expression of downstream targets. Certain Wnt ligands also activate the non-canonical Wnt/Planar Cell Polarity (PCP) pathway, which influences cellular movements [5], [30], [31]. β-catenin is essential in osteoblast differentiation and inhibition of chondrogenesis [6], [7], [12]–[14]; however, deletion of individual Wnt ligands resulted only in mild effects on bone differentiation [8], [32], [33]. β-catenin is also a central regulator of early dermal specification [3], [9], [10], [34], [35], and roles for Wnt ligands so far have only been directly shown later during hair follicle initiation [9], [11], [36]. In bone and skin development, redundant functions of multiple Wnts may compensate for deletion of individual ligands. Conventional knockouts of individual ligands removed Wnt expression from all cells in the embryo, and have confounded the identification of tissue sources of Wnt ligands in bone and skin development. Thus, the relative contributions from different sources of Wnt ligands for fate selection in cranial mesenchyme remain unknown.
Previous limitations were the lack of genetic tools to spatiotemporally manipulate early surface ectoderm and mesenchyme, and an inability to circumvent the intrinsic redundancy of Wnt ligands. We took a conditional approach to ablate the efficient secretion of Wnt ligands from either surface ectoderm or cranial mesenchyme prior to fate selection of the cranial bone and dermal lineages. Our findings provide key insights into how local developmental signals are utilized during morphogenesis to generate the cranial bone and dermal lineages.
We found that the genes for most Wnt ligands were expressed in the cranial mesenchyme (Figure 1A) and surface ectoderm (Figure 1B) during the specification of two separate lineages such as cranial osteoblast and dermal fibroblasts in E12.5 mouse embryos (Figure S1, S7, Table 1). To identify the cells with the potential to secrete Wnt ligands, we examined the spatiotemporal expression of Wls, the Wnt ligand trafficking regulator. We detected Wls protein expression from E11.5-E12.5 in the cranial surface ectoderm and in the underlying mesenchyme (Figure 1C, G). Both the Runx2-expressing cranial bone progenitor domain and the Dermo1/Twist2-expressing dermal progenitor domain expressed Wls [3], [37] (Figure 1C, D, E, G). Wnt signaling activation was also visualized in the cranial ectoderm, bone and dermal progenitors by expression of target gene, Lef1 and nuclear localized β-catenin (Figure 1D, F, H, I). During specification of cranial bone and dermis, ectodermal and mesenchymal tissues secreted Wnt ligands, and the dermal and bone progenitors actively transduced Wnt signaling via β-catenin (Figure 1J).
To dissect the requirements of ectodermal and mesenchymal Wnt signals, we generated mutant mice with conditional deletion of Wls [38] in the early surface ectoderm using Crect [39] and in the whole cranial mesenchyme using Dermo1Cre [40]. Crect efficiently recombined the Rosa26 LacZ Reporter (RR) in the cranial ectoderm by E11.5 (Figure S4K), but left Wls protein expression intact in the mesenchyme (Figure 2A, E, B, F) [41]. Dermo1Cre recombination showed β-galactosidase activity and Wls deletion restricted to the cranial mesenchyme and meningeal progenitors at E12.5, and Wls protein was still expressed in the ectoderm in mutants (Figure 2C, D, G, H).
First, we compared the extent to which Wls deletion from ectoderm or mesenchyme affected formation of the craniofacial skeleton. E18.5 Crect; RR; Wls fl/fl mutant embryos, which experienced perinatal lethality, demonstrated a hypoplastic face with no recognizable upper or lower jaw most likely due to decrease in cell survival of branchial arch mesenchyme (Figure S5). In the remaining tissue, facial mesenchyme patterning was grossly comparable to controls for most of the markers examined (Figure S5). Notably, the mutants showed no sign of mineralization in the skull vault (Figure 2I–L). The later deletion of Wls from the ectoderm using the Keratin14Cre line resulted in comparable skull bone ossification as controls (Figure S2). Dermo1Cre; RR; Wls fl/fl mutant embryos exhibited lethality after E15.5, which precluded assessment of skeletogenesis by whole-mount. We generated En1Cre/+; RR; Wls fl/fl mutants, using a Cre that recombines in early cranial mesenchyme but lacks activity in meningeal progenitors (Figure S3 E′, F′) [3]. En1Cre/+; RR; Wls fl/fl mutants survived until birth, and demonstrated reduced bone differentiation and mineralization (Figure S3) as well as intact dermis in the supraorbital region with hair follicles (Figure S3). The more severe arrest in Crect; RR; Wls fl/fl mutants (Figure 2) suggested ectoderm Wls appears to play an earlier role than mesenchymal Wls in cranial development.
We next examined the effects of ectoderm or mesenchyme Wls deletion on cranial bone and dermal development by histology. We found Von Kossa staining for bone mineral was absent in Crect; RR; Wls fl/fl mutants (Figure 3A, B). The thin domain of mesenchyme above the eye in mutants appeared undifferentiated and showed no condensing dermal cells or early stage hair follicles. Additionally, the baso-apical expansion of both dermis and bone was evident by E15.5 in controls, but not in the thin cranial mesenchyme of mutants (Figure 3A–B red arrowhead). Although ossification was absent, we observed the presence of thin nodules of ectopic, alcian blue-stained cartilage (Figure 3E–F). Therefore the result of Wls deletion in the ectoderm was an absence of skull ossification and hair-inducing dermis, a failure of baso-apical expansion of mesenchyme, and the presence of ectopic chondrocyte differentiation. By comparison, Dermo1Cre; RR; Wls fl/fl mutants showed a reduction in mineralized bone (Figure 3C–D) without ectopic cartilage formation (Figure 3 G–H). The mutant mesenchyme nonetheless condensed and formed sufficient hair-follicle generating dermis in the supraorbital region to support the supraorbital vibrissae hair follicle and fewer primary guard hair follicles (Figure 3 C, D, C′, D′, black arrowheads). Compared to the control apical region of the head, the mutant lacked sufficient condensed dermal layer to support normal number and differentiation of hair follicles (Fig. 3 C″, D″). Reduced mineralization without ectopic chondrogenesis as well as hair-follicle formation were also present in En1Cre/+; Wls fl/fl mutants (Figure S3). Our data suggest that Wls deletion using the Dermo1Cre resulted in diminished bone mineralization with thinner dermis and fewer hair follicles.
Deletion of Wls from the ectoderm resulted in complete absence of skull vault mineralization with failure of dermis formation, pointing to early defects in formation of the two lineages. Therefore we tested if cranial mesenchyme undergoes proper patterning, fate selection, and differentiation in the absence of Wls. Msx2 and Dlx5 that are early markers of skeletogenic patterning in cranial mesenchyme were expressed in Crect; Wls fl/fl mutants (Figures 4A, H, S4). The number of Msx2+ progenitor cells was not significantly different in controls and mutants (191±9.4 in controls and 206±24 in mutants, P-value = 0.23). However, few Runx2+ osteoblast progenitors formed in Crect; RR; Wls fl/fl mutant embryos, and expression shifted directly beneath the surface ectoderm (Figure 4B, I). During subsequent differentiation, condensing osteoblast progenitors express alkaline phosphatase (AP; Figure 4C, S4), but ectoderm Wnt-secretion deficient embryos lacked AP activity entirely (Figure 4J, S4). Markers of early osteoblast progenitors from other signaling pathways, Bmp4 and PTHrP (Figure 4D–E, K–L) were also absent in mutants, suggesting an arrest in osteoblast progenitor differentiation. The block was persistent as committed osteoblast progenitors expressing Osx were present in controls but not mutants (Figure 4F, M). Cell survival was not affected in the cranial mesenchyme prior to changes in marker expression (Figure S4A–D). We did not find significant difference in cell proliferation of the underlying mesenchyme (47%±4 in controls and 51%±2; P-value = 0.12). Whereas chondrocytes expressed Sox9 only at the skull base in controls, in mutants, ectopic Sox9-expressing chondrocyte progenitors and cartilage formed within the frontal bone domain (Figure 4G, N, Q, U). In spite of the effect of ectoderm-Wls deletion on mesenchyme, surface ectoderm expression of the differentiation marker, Keratin 14 (K14) was unaffected (Figure S4E,F). Next, we examined formation of dermal fibroblast progenitors in Crect; RR; Wls fl/fl mutant embryos. Cranial dermal fibroblast progenitors expressed the markers, Twist2 [3], [37] and Insulin Growth factor 2 (IGF2) by E12.5 in supraorbital mesenchyme (Figure 4O, P), but mutant embryos lacked Twist2 and IGF2 expression (Figure 4S, T). Twist2 expression became more progressively restricted to upper dermal fibroblasts during differentiation in controls, but was completely absent from cranial supraorbital mesenchyme of mutants (Figure 4 R, V). The altered cell fate marker expression at E12.5 (Figure 4, S4 I, J) immediately after deletion of ectoderm Wls (Figure S4K) was indicative of primary defects in mesenchymal cell fate selection. Together, our data suggest ectoderm Wnts form a non-cell autonomous inductive signal to the underlying mesenchyme for specification of osteoblast and dermal fibroblast progenitors, and for repression of chondrogenesis.
Next, we determined if mesenchyme Wls deletion resulted in a later defect in differentiation of cranial bone and dermal fibroblast progenitors. In En1Cre; RR; Wls fl/fl mutants, Runx2 expression in osteoblast progenitors was intact without ectopic Sox9 expression, but showed diminished expression of the skeletal differentiation marker, Osx and ossification (Figure S3). Wnt responsiveness by Axin2 expression was comparable in control and mutant cranial mesenchyme at E14.5 (Figure S3). In Dermo1Cre; RR; Wls fl/fl mutants, Runx2 expression was also unaffected during fate selection stages (Figure 5A, G, B, H). However, during later osteoblast progenitor differentiation (E15.5), Osx was diminished in mutants at E15.5 (Figure 5C, I). In dermal progenitors undergoing specification, Twist2 expression was unaffected (Figure 5D,J), and surface ectoderm differentiation marker, K14, was appropriately expressed (Figure S6C, D). Additionally at later stages in the mutant, we observed thinner dermis, which was sufficient to support initiation of fewer guard hair follicles (data not shown) and supraorbital vibrissae hair follicle formation (Figs. 3C, D; 5E, K). Furthermore, no ectopic expression of Sox9 occurred in mesenchyme Wls-deficient mutants (Figs. 5F, L). Deletion of mesenchyme-Wls did not lead to decrease in cell survival as monitored by expression of activated-Caspase3 (Figure S6A–B). Prior to E15.5, cell proliferation of osteoblast, dermal, and surface ectoderm progenitors was not significantly different from controls (Figure S6). Based on Dermo1Cre- and En1Cre- deletion of Wls, mesenchyme-derived Wnt ligands are not required for differentiation of dermal progenitors but are indispensable for later differentiation of osteoblast progenitors. Next, we tested the spatiotemporal requirement for mesenchyme Wls in Wnt signaling transduction. Nuclear β-catenin and Axin2 expression were comparable in the mesenchyme of mutants during fate selection stages at E12.5 (Figure 5M, N, Q, R). As differentiation occurs, expression of Axin2 and Lef1 was selectively diminished in the osteoblast progenitor domain of mesenchyme-Wls mutants compared to the controls (Figure 5O, P, S, T). Thus, mesenchyme Wnt ligands appeared to be important in mesenchyme Wnt signal transduction during osteoblast differentiation and ossification as opposed to earlier lineage specification events.
Next, we examined the source of Wnts for the onset of Wnt responsiveness in the mesenchyme. During dermal and osteoblast progenitor cell fate selection, Wnt ligands, inhibitors, and target genes are expressed in spatially segregated patterns. Wnt10a and Wnt7b were expressed in surface ectoderm (Figure 6A–B), Wnt11 was expressed in sub-ectodermal mesenchyme (Figure 6C), and Wnt16 mRNA was expressed in medial mesenchyme (Figure 6D). Notably, the soluble Wnt inhibitor, Dickkopf2 (Dkk2) mRNA was localized to the deepest mesenchyme overlapping with cranial bone progenitors (Figure 6E). Wnt ligands can induce nuclear translocation of β-catenin in a dose-dependent manner leading to the expression of early target genes [42], [43]. At E11.5, expression of nuclear β-catenin was present in both dermal and osteoblast progenitors, and the highest intensity of nuclear localization was found in the surface ectoderm and dermal mesenchyme (Figure 1F). Wnt target genes Lef1, Axin2, and TCF4 were patterned in partially complementary domains. Expression of Tcf4 protein was visible in the skeletogenic mesenchyme (Figure 6F). Tcf4 expression expanded into the mesenchyme under the ectoderm in ectoderm Wls-deficient mutants (Figure 6I–J) and was diminished in mesenchyme Wls-deficient mutants compared to controls (Figure 6K–L). Lef1 and Axin2 were expressed at the highest intensity in the dermal progenitors beneath the ectoderm (Figure 6 G, H). At E12.5, Lef1 expression was completely abolished in the mesenchyme of ectoderm-Wls mutants, but was comparable to controls in the absence of mesenchyme-Wls (Figure 6M–P). The onset of Wnt signaling response in the mesenchyme as measured by Lef1, Axin2, and nuclear β-catenin expression (Figure 6O–T) required ectoderm Wls. By contrast, no single tissue source of Wnt ligands was required to maintain TCF4 expression.
Finally, we tested whether cranial surface ectoderm Wnt ligands regulate the onset of Wnt ligand mRNA expression in the underlying mesenchyme (Figure 7). The non-canonical ligands Wnt5a and Wnt11 were expressed in cranial mesenchyme, with the highest expression corresponding to dermal progenitors. Wnt4, which signals in canonical or non-canonical pathways [44], was expressed strongly in dermal progenitors, as well as in osteoblast progenitors and in the skull base (Figure 7A–C). Wnt3a and 16, which signal in the canonical pathway via β-catenin and have roles in intramembranous bone formation, were expressed medially in the cranial mesenchyme containing cranial bone progenitors (Figure 7D, E) [12]–[14], [45]. Expression of Wnt5a Wnt11, Wnt3a, Wnt16 mRNAs was absent from the mesenchyme of Crect; RR; Wls fl/fl mutants whereas some Wnt4 expression was maintained (Fig. 7F–J). En1Cre deletion of β-catenin in the cranial mesenchyme [12] also resulted in an absence of Wnt5a and Wnt11 expression, except in a small portion of supraorbital lineage-labeled mesenchyme, suggesting a phenocopy of Crect;Wls mutants (Figure 7K, L, M). In contrast, Wnt5a, Wnt11, and Wnt4 expression were present in the Dermo1Cre; RR; Wlsfl/fl mutants (Figure 7N–S). However, the Wnt-expressing domains were smaller and only located close to the surface ectoderm, but nonetheless were lineage-labeled (Figure 7E–G, L–N; not shown). Thus, consistent with a role as initiating factors, ectoderm Wnt ligands and mesenchyme β-catenin were required for expression of certain Wnt ligands in the cranial mesenchyme during lineage selection. Mesenchymal Wnt ligands may in turn be required later for osteoblast differentiation (Figure 7T).
Here we obtained data suggesting that ectodermal and mesenchymal Wnts function distinctly in early dermal and osteoblast progenitor specification and differentiation. Wnt ligands are expressed in the cranial surface ectoderm and mesenchyme, and ectoderm Wnts are required to generate an inductive cue for the specification of multiple lineages in the cranial mesenchyme. The dermal progenitors and osteoblast progenitors closest to the ectoderm experience the highest concentrations of nuclear β-catenin, in response to Wnt ligands from overlying ectoderm. Subsequent differentiation of osteoblast and dermal fibroblast progenitors requires Wls from the mesenchyme. Thus our study demonstrates that two different sources of Wnt signals coordinate to form two separate lineages, bone and dermis.
We present evidence to demonstrate that ectoderm Wnts generate an inductive cue of Wnt signaling in the mesenchyme to specify cranial bone and dermal lineages. The mechanism remains elusive; however, there are at least three possible models. First, the spatial segregation of Wnt pathway transcription cofactors such as Lef1 and TCF4, partially by lineage, provides a mechanism to generate different lineage programs. Second, a threshold-dependent model may also exist to generate multiple lineages from the same signal. At E11.5–12.5, dermal progenitors are closest to the ectoderm Wnt source and exhibit the highest Wnt signaling reporter activity and markers induced by constitutive activation of β-catenin in mesenchyme (Figure 1) [3], [9], [46]. High levels of Wnt pathway activity preclude osteoblast marker expression in the mesenchyme [12]. Consistently, osteoblast progenitors are present farther away from the ectoderm in an overlapping domain to at least one Wnt inhibitor, Dkk2 [47] (Figure 6E). Finally, the osteoblast response to ectodermal Wnts may be indirect; osteoblast progenitors may require a separate signal relayed from dermal progenitors. Future genetic experiments with new reagents will be required to distinguish between these models and test direct or indirect requirements of Wnt sources in osteoblast and dermis formation.
During fate selection of cranial dermal and osteoblast progenitors, upstream ectodermal Wnt ligands initiate expression of a subset of mesenchymal Wnt ligands via β-catenin. Ectoderm Wnts also act upstream of mesenchyme Wnts in mouse limb development [48]. Here, ectoderm Wnts act in a temporally earlier role than mesenchyme Wnts, and other studies support a direct relationship. In at least one instance, mesenchyme Wnt ligands are direct targets of canonical Wnt signaling [49]. Alternatively, ectoderm and mesenchyme Wnts may signal in parallel pathways to the mesenchyme. The signal that acts upstream to initiate Wnt ligand expression in the cranial ectoderm remains unknown.
We report here that osteoblast differentiation requires distinct Wnt signals from surface ectoderm and mesenchyme. β-catenin deletion in the ectoderm did not inhibit skull bone mineralization [39], so autocrine effects of Wls deletion on the ectoderm were unlikely to contribute to the skull phenotype. However, removal of surface ectoderm Wls resulted in ectopic chondrogenesis (Figure 3), which phenocopied mesenchymal β-catenin deletion [12]. In contrast, mesenchymal Wls deletion did not result in ectopic cartilage formation, suggesting repression of chondrogenesis in cranial mesenchyme requires an early, ectoderm Wnt signal. Our results thus implicate β-catenin here as a Wnt pathway factor that acts in the nucleus to repress chondrogenesis and functions downstream of ectoderm ligands. Ectoderm Wnt ligands thus provide an inductive cue acting on osteoblast progenitors while the cells are closest to the ectoderm. Indeed, later deletion of Wls from the ectoderm using the K14Cre line did not give rise to a skull bone ossification phenotype (Figure S2). During osteoblast progenitor differentiation, Wls deletion with Dermo1Cre resulted in a similar but more severe differentiation arrest than the more restricted En1Cre. Consistently, using a different Wls mutant allele, deletion of mesenchymal Wnts led to absence of osteoblast differentiation expression and reduced cell proliferation [50]. We show that the mesenchyme Wnts maintain the differentiation process but require an inductive ectoderm Wnt signal.
We demonstrate that dermal progenitors require ectodermal Wls for specification and mesenchymal Wls for normal differentiation (Figs. 4–5). Cranial dermal progenitors located beneath the ectoderm require β-catenin for specification [3], but the tissue contribution of Wnt sources remained previously undetermined. Here, a mesenchymal Wls source is indispensable in the dermal lineage for normal differentiation, thickness, and hair follicle patterning. Previous reports in murine trunk skin development suggested that ectoderm Wnts alone are essential in hair follicle induction [9], [10]. Differential requirements may exist for mesoderm-derived trunk dermal progenitors and cranial neural crest-derived dermal progenitors. Future studies will be needed to uncover the requirements for a mesenchymal Wnt signal in dermal fibroblast differentiation in different parts of the embryo.
Conditional Wls deletion resulted in a failure of cranial dermal and osteoblast progenitors to undergo baso-apical extension (Figure 3), a process that occurs independently of β-catenin [12]. Since Wls deletion blocked secretion of canonical and non-canonical Wnt ligands, extension defects in the mesenchyme are consistent with known roles for non-canonical Wnt ligands in orienting cell movements [51]. Homozygous null mutants of core planar cell polarity (PCP) components lacked proper skull tissue development and neural tube closure [52]. However, mutants for individual non-canonical Wnt ligands lack a cranial PCP phenotype. In the cranial mesenchyme, non-canonical Wnt5a or Wnt11 ligands were expressed in overlapping expression domains, suggesting the ligands function redundantly [53] (Figure 7). Therefore, the role of PCP signaling remains to be rigorously tested in conditional mutant mice. The non-canonical and canonical Wnt signaling pathways interact extensively. In our study, canonical β-catenin transduction, in response to ectodermal Wnts, initiates non-canonical Wnt ligand expression (Figure 7), consistent with reports from other systems [30], [49], [51]. Our results reinforce the role of non-canonical Wnt ligands in the pathogenesis of craniofacial anomalies [54], [55]. The ability of exogenous non-canonical Wnts to compensate for Wls deletion in the baso-apical extension of dermal and osteoblast progenitors remains to be tested.
Our results from tissue-specific deletion of Wls have implications in diseases with dysregulation of dermal fibroblasts or osteoblasts, and in understanding the pathogenesis of craniofacial birth defects. Removal of Wls from the ectoderm by E12.5 of mouse development reveals a default state for formation of cartilage in the cranial skeleton and dermis if all Wnt secretion were absent from the ectoderm. This forms an important baseline state that can be used to interpret less severe genetic conditions resulting from loss or mutation of individual Wnt ligands. In this respect, we hypothesize that mutations in the Wnt secretory pathway may underlie diseases of osteoblasts, and dermal fibroblasts, warranting continued investigation into the role of Wnt production in bone and skin formation and homeostasis [15], [17], [18], [45], [56]–[58]. Understanding the signals surrounding osteoblast and dermal fibroblast formation is crucial to meet the demands of engineering appropriate connective tissues.
Conditional functional studies were conducted using Crect, Keratin 14Cre; Dermo1Cre, En1Cre, β-catenin deleted, conditional β-catenin floxed mice [39], [40], [59]–[62]. Mice and embryos were genotyped as described previously. The conditional loss-of-function floxed allele for Wls (Wlsfl/fl) was described previously [38]. RR/RR mice harboring a LacZ transgene downstream of a floxed stop transcription signal in the ubiquitous Rosa26 locus were obtained for lineage tracing [41]. For timed matings the vaginal plug day was assigned as E0.5. At desired time points, embryos were harvested and processed for frozen sections as previously described [34]. For each experiment, at least three to five different mutants with littermate controls from 2–3 litters were analyzed. At least three to five litters were used for all analyses. Case Western Reserve Institutional Animal Care and Use Committee approved all animal procedures.
Embryos were fixed in 4%PFA, cryopreserved, and sectioned at 8–12 µm. In situ hybridization, β-galactosidase with eosin counter-staining, and immunohistochemistry were performed essentially as described [34], [35]. Alcian blue staining of sections was performed as described. For Von Kossa staining of frozen sections, slides were fixed with 4% PFA, incubated in the dark with 2% silver nitrate, rinsed, exposed to light, and counterstained with eosin. In situ probes for Twist2 (Eric Olson, Dallas, TX), Pthrp, Wnt4 (V. Lefebvre, Cleveland, OH), Wnt5a (Andrew McMahon, Boston, MA), Wnt11 (Steve Potter, Cincinnati, OH), Axin2 (Brian Bai), BMP4, Wnt7b, Dlx5 (Gail Martin, San Francisco, CA), Wnt16 (Yingzi Yang, Bethesda, MD) and Osx (Matthew Warmann, Boston, MA) were gifts. For Wnt10a, cDNA was amplified from E12.5 RNA using primer F: GCTATTTAGGTGACACTATAGGCGCTCTGGGTAAACTGAAG, primer R: TTGTAATACGACTCACTATAGGGAGAGCCAACCACCTCTCTCA, and in vitro transcription of antisense mRNA with T7 polymerase. For Dkk2, PCR primers DKK2-F(5′-GACATGAAGGAGACCCATGCCTACG-3′ and DKK2-T7R 5′-TGTAATACGACTCACTATAGGGCATAGATGAGGCACATAACGGAAG-3′ were used.
Primary antibodies for Runx2, Sox9, Twist2, Lef1, Osx, Msx2, Ki67, IGF2, Wls, and β-catenin (goat anti-Runx2; 1∶20, R&D Biosystems; rabbit anti-Sox9; 1∶100; Millipore; mouse anti-Twist2, 1∶500, Santa Cruz; rabbit anti-Lef1, 1∶100, Abcam; rabbit anti-Osx, 1∶400, Abcam; mouse Msx1/2, 1∶50, DSHB; activated Caspase3, 1∶250, Abcam; rabbit Ki67; 1∶500 Abcam; rabbit IGF2 1∶400, Cell Signaling); rabbit anti-Wls, 1∶2000, gift from Richard Lang; mouse β-catenin 1∶100 BD Biosciences) were used for indirect immunofluorescence and immunohistochemistry. All control/mutant pairs were photographed at the same magnification. Number of Msx2+ cells was counted from a fixed field in 10 different sections from 4 embryos.
Proliferation index was assessed by percent of cells with Ki67 expression in the Runx2 expression domain, in the dermal mesenchyme in the Twist2 domain, and surface ectoderm in the Keratin14 expressing cells. Similar numbers of cells in each domain were analyzed between four controls and mutants. Statistical significance for all quantifications was calculated using two-tailed Student t-test.
Embryos were sacrificed, skinned and eviscerated, fixed in 95% ethanol, then stained for 24 hours each in 0.03% Alcian blue and 0.005% Alizarin red. Stained embryos were subsequently cleared in graded series of potassium hydroxide and glycerol until photography, after which they were stored in 0.02% Sodium Azide in glycerol. Whole mount Alkaline phosphatase staining was performed as previously described [63] with the addition of a 70% ethanol overnight incubation step after fixation in 4% PFA.
Cranial mesenchyme and surface ectoderm were micro-dissected from E12.5 embryos and flash frozen in liquid nitrogen. Total RNA was isolated using the Qiagen RNEasy micro kit, and cDNA was reverse transcribed using the ABI kit. RT-PCR for most of the Wnt ligands was amplified for 35 cycles of 94°C for 15 seconds, 66°C for 30 seconds, and 72°C for 60 seconds and the products were resolved on a 3% agarose gel. For Wnt1, 5b, 8a, 8b, 10b the annealing temperature was 55°C for 30 seconds. Primer sequences for RT-PCR are listed in Table 1.
|
10.1371/journal.pntd.0000815 | The OmpL37 Surface-Exposed Protein Is Expressed by Pathogenic Leptospira during Infection and Binds Skin and Vascular Elastin | Pathogenic Leptospira spp. shed in the urine of reservoir hosts into freshwater can be transmitted to a susceptible host through skin abrasions or mucous membranes causing leptospirosis. The infection process involves the ability of leptospires to adhere to cell surface and extracellular matrix components, a crucial step for dissemination and colonization of host tissues. Therefore, the elucidation of novel mediators of host-pathogen interaction is important in the discovery of virulence factors involved in the pathogenesis of leptospirosis. In this study, we assess the functional roles of transmembrane outer membrane proteins OmpL36 (LIC13166), OmpL37 (LIC12263), and OmpL47 (LIC13050), which we recently identified on the leptospiral surface. We determine the capacity of these proteins to bind to host tissue components by enzyme-linked immunosorbent assay. OmpL37 binds elastin preferentially, exhibiting dose-dependent, saturating binding to human skin (Kd, 104±19 nM) and aortic elastin (Kd, 152±27 nM). It also binds fibrinogen (Kd, 244±15 nM), fibrinogen fragment D (Kd, 132±30 nM), plasma fibronectin (Kd, 359±68 nM), and murine laminin (Kd, 410±81 nM). The binding to human skin elastin by both recombinant OmpL37 and live Leptospira interrogans is specifically enhanced by rabbit antiserum for OmpL37, suggesting the involvement of OmpL37 in leptospiral binding to elastin and also the possibility that host-generated antibodies may promote rather than inhibit the adherence of leptospires to elastin-rich tissues. Further, we demonstrate that OmpL37 is recognized by acute and convalescent leptospirosis patient sera and also by Leptospira-infected hamster sera. Finally, OmpL37 protein is detected in pathogenic Leptospira serovars and not in saprophytic Leptospira. Thus, OmpL37 is a novel elastin-binding protein of pathogenic Leptospira that may be promoting attachment of Leptospira to host tissues.
| Leptospirosis is a potentially fatal disease in humans and livestock caused by Leptospira bacteria. Effective antibiotic treatment depends on timely, accurate diagnosis. However, current diagnostic and vaccine options are limited by their specificity for the lipid-sugar coat of leptospires, which varies among 200 serum-reactive groups. We aim to understand how leptospires infect a host, and in turn, to develop broadly effective diagnostic and immunization products. We recently described OmpL37, a new protein on the surface of leptospires. Here, we show it is made by pathogenic strains, suggesting it can be a target for detecting and protecting against a wide range of Leptospira. Moreover, leptospirosis patients and hamsters infected with leptospires make antibodies against OmpL37. Purified OmpL37 binds host proteins, including human elastin, fibrinogen, fibronectin, and mouse laminin. Although other leptospiral proteins bind multiple host proteins, OmpL37 has novel preferential affinity for skin and aorta elastin, suggesting a role in a common route of transmission through abraded skin and exposed blood vessels. Indeed, OmpL37 binding and leptospiral attachment to elastin are both enhanced by OmpL37 antiserum, further implicating a possible role for OmpL37 during infection. Thus, OmpL37 may mediate host attachment and has potential clinical application with a broad range of Leptospira.
| Leptospirosis is a zoonosis caused by pathogenic Leptospira spp. transmitted from reservoir hosts (typically rodents) to humans via water contaminated by infected animals and has a significant impact on public health throughout the developing world [1]–[4]. Leptospirosis also has significant adverse effects on the agricultural industry by causing abortions, infertility, and death in livestock [5], [6]. After being shed in the urine of a reservoir host animal, leptospires can persist in freshwater or soil until contact with abraded skin or mucous membranes of a new host occurs. The resulting infection is potentially fatal, and is frequently characterized by jaundice, renal failure, and/or pulmonary hemorrhage [1], [4], [7]. Large outbreaks of leptospirosis occur in tropical and subtropical regions after heavy rainfall and the dispersal of leptospires in contaminated water [3], [8]. Current vaccines against leptospirosis target the lipopolysaccharide (LPS) coat of the leptospires, which is highly variable; this variation is thought to be the major antigenic determinant defining the differences between approximately 230 serovars that contribute to serovar-specific immunity [6], [9]. In contrast, vaccines directed towards well-conserved leptospiral outer membrane proteins (OMPs) [10], [11] would have an advantage in inducing cross-protective immunity [12].
The leptospiral lifecycle involves interactions with host tissues at multiple stages of infection, including: (i) entering the host, (ii) evading its immune response, and (iii) adhering to tissues [13]–[15]. Identification and characterization of novel proteins that mediate the stage-specific interactions with the host are essential for the understanding of leptospiral pathogenesis, and in the development of diagnostic and protective antigens for leptospirosis. Pathogenic leptospires have been shown to bind to a variety of host ligands, including fibronectin, fibrinogen, collagen, laminin, elastin, and proteoglycans, indicating that cell surface and extracellular matrix (ECM)-binding OMPs, or adhesins, are likely to be expressed by the spirochetes [16]–[21]. It is possible that leptospires express distinct adhesins at different stages of infection, including initial attachment, dissemination, and colonization. Many leptospiral proteins, including LigA/B, Lsa21, Lsa27, Lsa63, Lsa24 (LfhA/LenA), LenB to F, LipL32, Lp95, TlyC, and LipL53, have been shown to have affinity for host ligands in vitro [18], [19], [21]–[32]. However, it is unclear to what extent these putative adhesins mediate interactions of leptospires with cell surface and ECM proteins. Only Lsa24, LigA/B, and Lsa63 have been tested for their capacity to inhibit leptospiral adherence to ECM proteins [18], [21], [24], [32]. In each case, only partial inhibition was observed, suggesting that additional fibronectin-, laminin-, collagen-, and elastin-binding proteins likely exist in Leptospira [18], [21], [24], [32].
In this study, we investigated whether the surface-exposed proteins in Leptospira, OmpL36 (LIC13166), OmpL37 (LIC12263), and OmpL47 (LIC13050) that we recently described [33] can bind to any host ligands. We now report that OmpL37 is the first leptospiral protein found to have pronounced specificity for human skin elastin. OmpL37 exhibits strong, saturating binding to skin elastin with one of the highest affinities of all leptospiral ligand-binding proteins described. In addition, OmpL37 binds efficiently to human aortic elastin, fibrinogen, fibrinogen fragment D, and to a lesser extent laminin and plasma fibronectin. OmpL47 also binds to laminin, plasma fibronectin, fibrinogen, fibrinogen fragment D, along with collagen type III, and aortic elastin, but showing much lower activities than OmpL37. OmpL36 shows no binding to any of the host tissue components investigated.
Elastin is a connective tissue component of ECM responsible for the elasticity and resilience of skin, lung, blood vessels, uterus, placenta, and other tissues [34]–[36]. These elastin-rich tissues are highly relevant to leptospirosis as infection includes entry through skin abrasions or mucous membranes, dissemination through the circulation, and attachment to vascular, renal, pulmonary, uterine, and other tissues. Until our discovery of the elastin-binding properties of OmpL37, only LigB was known to have the capacity to bind elastin and tropoelastin [21]. We also show that OmpL37 antiserum can enhance the binding to skin elastin by both live Leptospira and recombinant OmpL37, suggesting that leptospiral binding to elastin is at least partially mediated by OmpL37. Expression of OmpL37 during infection is confirmed with the recognition of OmpL37 by sera from Leptospira-infected hamsters and also acute and convalescent leptospirosis patients. While the gene for an OmpL37 homologue is present in saprophytic leptospires, OmpL37 is detectable only in pathogenic Leptospira serovars. Taken together, our data suggest that OmpL37 is an elastin-binding protein of Leptospira with potential roles in leptospirosis, including the attachment to elastin-rich tissues, such as the dermis, vasculature, and lungs.
This study was conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the Institutional Review Board of the Research and Development Committee, VA Greater Los Angeles Healthcare System, Research Service (PCC # 2008-121778).
All animals were handled in strict accordance with good animal practice as defined by the relevant national and/or local animal welfare bodies, and all animal work was approved by the VA Greater Los Angeles Healthcare System, Research Service (PCC # 2009-010088).
Leptospira interrogans serovar Copenhageni strain Fiocruz L1-130 was isolated from a patient during a leptospirosis outbreak in Salvador, Brazil [8]. L. interrogans serovar Pomona strain PO-01, L. kirscheneri serovars Mazdok strain 5621 and Grippotyphosa strain RM52, L. borgpetersenii serovars Tarrasovi strain Perepelicin and Javanica strain Veldrat Bataviae 46, Leptonema illini strain 3055, Leptospira weilii serovar Celledoni strain Celledoni, L. wolbachii serovar Biflexa strain codice, L. inadai serovar Lyme strain 10, and L. biflexa serovar Patoc strain Patoc 1 were obtained from the National Leptospirosis Reference Center (National Animal Disease Center, Agricultural Research Service, U.S. Department of Agriculture, Ames, Iowa). Leptospires were cultivated in Ellinghausen-McCullough-Johnson-Harris (EMJH) medium supplemented with 1% rabbit serum (Rockland Immunochemicals, Gilbertsville, PA) and 100 µg/ml 5-fluorouracil at 30°C [37].
The polyclonal rabbit sera specific for OmpL37, OmpL47, and OmpL54 have been described previously [33]. Immunoglobulins G (IgG) from OmpL37 and OmpL47 antisera were purified by Melon Gel IgG spin purification kit (Thermo Scientific, Rockford, IL) according to manufacturer's instructions. LipL32 monoclonal antibody 1D9 [38], [39] was a kind gift from Dr. José Antonio Guimarães Aleixo (Universidade Federal De Pelotas, Pelotas, Brazil). Pooled sera from infected Syrian hamsters (Harlan Laboratories) were obtained ten days following intradermal challenge of one month-old animals with L. interrogans L1-130. As a negative control, a serum sample from a hamster injected intradermally with EMJH was collected after 12 days. Patient sera from leptospirosis outbreaks in 1996 and 1997 in Salvador, Brazil, were kindly provided by Dr. Albert I. Ko (Oswaldo Cruz Foundation, Salvador, Bahia, Brazil). Acute and convalescent samples from the same patients were prepared by pooling sera from 13 individuals with laboratory-confirmed leptospirosis. Normal human serum was obtained from Thermo Scientific and Millipore (Billerica, MA).
The gene encoding potential outer membrane lipoprotein, LIC10091 (LipL40) [40], was amplified from Fiocruz L1-130 DNA using forward primer, 5′- TTCGCATATGAAAACGCCTCCTCCTAAAG -3′, and reverse primer, 5′- TAAAATCTCGAGTTTCAAAACTTCTACGGGC- 3′, by PCR conditions as described for ompL36, ompL37, and ompL47 [33]. PCR product was digested with NdeI and XhoI (New England BioLabs, Ipswich, MA), cloned into NdeI- and XhoI- digested expression vector, pET-20b(+) (Novagen, San Diego, CA), and purified as previously described for OmpL36, OmpL37, and OmpL47 [33].
Protein samples were boiled for 5 min in Novex NuPAGE sample buffer (Invitrogen, Carlsbad, CA) in the presence of 2.5% β-mercapthoethanol and separated in Bis-Tris 4–12% polyacrylamide gradient NuPAGE gels (Invitrogen).
For immunoblotting, proteins were transferred to a polyvinylidene difluoride (PVDF) Immobilon-P membrane (Millipore) and probed with rabbit polyclonal antisera, Syrian hamster sera, or leptospirosis patient sera. Bound antibodies were detected using horseradish peroxidase (HRP)-conjugated anti-rabbit IgG (GE Lifesciences, Buckinghamshire, England), anti-Syrian hamster IgG (Jackson Immuno Research, West Groove, PA), or anti-human IgG (Sigma-Aldrich, St. Louis, MO), respectively. Immunoblots were visualized by enhanced chemiluminescence reagents according to the manufacturer's instructions (Thermo Scientific).
Host ligands included human plasma fibronectin (Sigma-Aldrich), human plasma fibronectin 30-kDa proteolytic fragment (heparin-binding domain, Sigma-Aldrich), human plasma fibronectin 45-kDa proteolytic fragment (gelatin-binding domain, Sigma-Aldrich), human fibroblast fibronectin (Calbiochem, La Jolla, CA), human plasma fibrinogen (HYPHEN BioMed, France), human plasma fibrinogen fragment D (HYPHEN BioMed), murine laminin (Sigma-Aldrich), bovine skin collagen type I (Sigma-Aldrich), human placenta collagen type III (Sigma-Aldrich), human placenta collagen type IV (Sigma-Aldrich), soluble human skin elastin (Elastin Products Company, Owensville, MO), soluble human aorta elastin (Sigma-Aldrich), bovine kidney heparan sulfate (Sigma-Aldrich), shark cartilage chondroitin sulfate (Sigma-Aldrich), fetal calf serum fetuin (Sigma-Aldrich), and bovine serum albumin (BSA, Sigma-Aldrich). Ultra-high binding Immulon 4HBX microtiter plates (Thermo Scientific) were coated with 1 µg of host ligand in 0.1 ml of phosphate buffered saline (PBS), pH 7.2, and incubated overnight at 4°C. Fresh (stored for less than 4 months at 4°C) recombinant OmpL36, OmpL37, OmpL47 [33], and LIC10091 (used as a negative control) binding to individual ligands was assessed by ELISA. Briefly, non-specific binding sites were blocked with Protein-Free Blocking buffer (PFBb; Thermo Scientific) for 1 h at room temperature and 1 µg of recombinant protein in 0.1 ml of PFBb was added per well and incubated for 1 h at 37°C. For assays of ligand binding as a function of leptospiral protein concentration, serial dilutions of recombinant OmpL37 and OmpL36 (negative control) ranging from 0 to 2 µM in 0.1 ml of PFBb were added to wells and incubated for 1 h at 37°C. Wells were washed three times with PBS, pH 7.2, and bound protein was detected by probing with anti-His Tag monoclonal antibody (5 Prime, Gaithersburg, MD), developing with HRP-conjugated anti-mouse IgG (Novagen) and a tetramethyl benzidine substrate (Thermo Scientific), and recording by spectrophotometry at 450 nm. For saturating binding the apparent dissociation constant (Kd) was estimated as the concentration of OmpL37 resulting in half-maximal binding.
For assays assessing effects of antibodies on binding of recombinant OmpL37 to skin elastin, recombinant OmpL37 was pre-incubated either with serum against OmpL37 or OmpL47 (negative control) at 1∶2500 dilution or purified IgG at 1∶40 to 1∶640 serial dilutions in PFBb at room temperature for 30 min. In addition, OmpL37 was pre-incubated with convalescent leptospirosis patient sera or Leptospira-infected hamster sera at 1∶640 to 1∶10,240 serial dilutions in PFBb at room temperature for 1 h. Then mixtures containing 0.5 µg of OmpL37 in 0.1 ml of PFBb were added to microtiter wells, incubated for 1 h at 37°C, and bound protein detected as described above.
Oneµg of recombinant OmpL37 or OmpL36 (negative control) in 0.1 ml PBS, pH 7.2, was used to coat Immulon 4HBX microtiter wells overnight at 4°C. Non-specific binding sites were blocked with PFBb and 1 µg of human plasma fibronectin, human plasma fibrinogen, human plasma fibrinogen fragment D, murine laminin, bovine skin collagen type I, human collagen type IV, or human skin elastin in 0.1 ml of PFBb was added and incubated for 1 h at room temperature. After three washes with PBS, pH 7.2, OmpL37-bound host ligands were detected by probing with anti-human fibronectin rabbit antibody (Sigma-Aldrich), anti-human fibrinogen rabbit IgG (HYPHEN BioMed), anti-laminin rabbit IgG (Sigma-Aldrich), anti-collagen type I monoclonal antibody (Clone COL-1, Sigma-Aldrich), anti-collagen type IV monoclonal antibody (Clone COL-94, Sigma-Aldrich), or anti-elastin rabbit IgG (Santa Cruz Biotechnology, Inc., Santa Cruz, CA), then adding HRP-conjugated anti-mouse IgG (Novagen) or HRP-conjugated anti-rabbit IgG (GE Lifesciences) and developing as described above.
Microtiter plates were coated with human skin elastin and non-specific binding sites were blocked as described above. L. interrogans cultures were harvested by centrifugation at 2000× g for 15 min at room temperature and resuspended in PBS-5 mM MgCl2 to a final concentration of 1×109 cells/ml. To assess the effect of the OmpL37 antibodies on leptospiral binding to elastin, serial dilutions (1∶40 to 1∶1280) of anti-OmpL37 or anti-OmpL54 (used as a negative control) were mixed with the leptospires or PBS-5 mM MgCl2 and 1×108 cells in 0.1 ml of PBS-5 mM MgCl2 were added to the microtiter wells. For experiments assessing the inhibition of leptospiral binding by recombinant OmpL37, prior to the addition of leptospires, 0.5 µM of recombinant protein in 0.1 ml of PFBb was either added directly to elastin-coated microtiter wells or pre-incubated with anti-OmpL37 or anti-OmpL47 (negative control) at a 1∶500 dilution for 30 min at room temperature and added to the microtiter wells. Plates were incubated for 1 h at 37°C, and washed three times with PBS. After addition of leptospires, plates were incubated at 30°C for 90 min, unbound leptospires were removed by four washes with PBS-5 mM MgCl2, and adherent cells were fixed with methanol at −20°C for 10 min. Elastin-bound leptospires were detected by probing with LipL32 monoclonal antibody 1D9 and developing as described above.
To measure the effect of OmpL37 antibodies on the binding of immobilized leptospires to freely soluble skin elastin, 1×108 of L. interrogans in PBS-5 mM MgCl2 were allowed to adhere to Immulon 4HBX microtiter wells for 90 min at 30°C, washed twice with PBS-5 mM MgCl2, and non-specific binding sites were blocked with PFBb for 30 min at room temperature. Rabbit sera for OmpL37 and OmpL47 (negative control) diluted 1∶40 and 1∶2560 in PBS-5 mM MgCl2 were added to the wells and incubated for 90 min at room temperature, followed by three washes with PBS-5 mM MgCl2. One µg of human skin elastin in 0.1 ml PBS-5 mM MgCl2 was added and incubated overnight at 4°C. Wells were washed three times with PBS, pH 7.2, and cell-bound complexes were fixed with methanol at −20°C for 10 min, followed by additional blocking with PFBb. Leptospira-bound elastin was detected by probing with anti-elastin mouse serum (Novus Biologicals, Littleton, CO) and developing as described above.
To investigate the capacities of our recently described surface-exposed transmembrane OmpL proteins [33] to interact with host tissue ligands, soluble recombinant OmpL36, OmpL37, and OmpL47 were assessed for binding to immobilized host components. BSA and the highly glycosylated serum protein, fetuin, were used as controls for non-specific binding. Soluble recombinant LIC10091 was used as a non-binding protein control. OmpL37 exhibited significant binding to ECM components, such as human skin (P<0.001 compared to BSA) and aorta elastin (P<0.001), laminin (P<0.001), fibrinogen (P<0.01), fibrinogen fragment D (P<0.001), and plasma fibronectin (P<0.05) (Fig. 1). OmpL37 binding to the 30-kDa and 45-kDa fragments of plasma fibronectin, fibroblast fibronectin, collagen types I, III, and IV, heparan sulfate, and chondroitin sulfate was not statistically significant (Fig. 1). OmpL47 also showed significant binding to laminin (P<0.001), fibrinogen (P<0.001), fibrinogen fragment D (P<0.001), plasma fibronectin (P<0.001), collagen III (P<0.001), and aorta elastin (P<0.05). However, the OmpL47 activities were much lower than those observed for OmpL37 and were not investigated further. None of the other recombinant proteins exhibited significant binding to any of the host tissue components investigated (Fig. 1).
In order to compare binding affinities, the interaction with immobilized skin and aorta elastin, fibrinogen, fibrinogen fragment D, plasma fibronectin, plasma fibronectin 30-kDa and 45-kDa fragments, and laminin was measured as a function of OmpL37 concentration from 0 to 2 µM (Fig. 2 and Table 1). OmpL37 exhibited very strong, saturating binding to skin elastin (Kd , 104±19 nM), aorta elastin (Kd , 152±27 nM), fibrinogen (Kd , 244±15 nM), and fibrinogen fragment D (Kd , 132±30 nM) as estimated by ELISA from three independent experiments. OmpL37 binding to plasma fibronectin (Kd , 359±68 nM), its 30-kDa (Kd , 408±94 nM) and 45-kDa fragment (Kd , 460±70 nM), and laminin (Kd , 410±81 nM) were noticeably lower (Fig. 2 and Table 1). OmpL36 was used as a negative control based on our observation that OmpL36 does not bind to host ligands (Fig. 1). The comparison of apparent Kds for the saturation binding of host proteins by recombinant OmpL37 is summarized in Table 1.
In addition, we investigated whether OmpL37 immobilized on microtiter wells bound freely soluble host proteins (Fig. 3). Immobilized OmpL37 exhibited significant binding to free laminin (P<0.001 compared to collagen IV), human skin elastin (P<0.001), and plasma fibronectin (P<0.05) (Fig. 3). It is evident that immobilized OmpL37 binds to virtually the same host proteins as freely soluble OmpL37, with the exception of fibrinogen and fibrinogen fragment D (Fig. 1 and 3). Although immobilized OmpL37 can bind to freely soluble ligands, the interaction appears to be weaker when compared to that of free OmpL37 binding to immobilized ligands (Fig. 1 and 3). This could be due to an impairment of the active conformation of OmpL37 caused by its immobilization to the microtiter wells, to the differences between antibodies utilized for detection or to the intrinsic property of less efficient binding of soluble versus immobilized host proteins, which has been described for pathogenic bacteria and their surface proteins [41]–[43].
We examined whether OmpL37 mediates the attachment of leptospires to human skin elastin by testing for OmpL37 antiserum inhibition of live Leptospira binding to immobilized skin elastin in an ELISA. Surprisingly, the OmpL37 antiserum enhanced adhesion in a dose-dependent manner, whereas antiserum for another surface-exposed OMP, OmpL54, had no effect (Fig. 4A). Leptospiral binding to skin elastin was enhanced by the OmpL37 antiserum at dilutions of 1∶40 (1.6-fold compared to no antibody, P<0.001), 1∶80 (1.5-fold, P<0.001), and 1∶160 (1.3-fold, P<0.05); the effect was not evident at higher dilutions (Fig. 4A). The increase in adhesion was not due to leptospiral agglutination as determined with a systematic examination for the presence of leptospiral aggregates or decrease in leptospiral numbers in multiple fields by dark-field microscopy (data not shown). We also investigated the effect of OmpL37 antiserum on the binding to free skin elastin by Leptospira immobilized on microtiter wells and found that a 1∶40 dilution of antiserum increased binding 1.4-fold (data not shown). Further, we assessed the effect of the OmpL37 antibodies on the binding of recombinant OmpL37 to skin elastin (Fig. 4B). The OmpL37 antiserum significantly enhanced recombinant OmpL37 binding to skin elastin at a dilution of 1∶2500 (1.4-fold, P<0.05), a >10-fold lower antiserum concentration compared to that required for an effect on the adhesion of leptospires (Fig. 4A and 4B). IgG purified from the OmpL37 antiserum slightly enhanced (1.2-fold) OmpL37 binding to skin elastin at dilutions of 1∶40 to 1∶160 (Fig. 4B); however the effect was not statistically significant (P>0.05), indicating the possibility that along with OmpL37-specific antibody, another serum component removed in the purification of IgG could be required for binding enhancement. The antiserum and purified IgG for another surface-exposed OMP, OmpL47, was used as a negative control and did not have any effect on OmpL37 binding (Fig. 4B). The OmpL37 antiserum did not exhibit any statistically significant effect on recombinant OmpL37 binding to fibronectin, fibrinogen, and laminin (data not shown). Further, we investigated whether convalescent leptospirosis patient sera enhance leptospiral binding to skin elastin. Dark-field microscopy revealed leptospiral agglutination by patient sera (data not shown), which prevented the assessment of leptospiral adhesion. We also investigated the effects of convalescent leptospirosis patient sera and Leptospira-infected hamster sera on recombinant OmpL37 binding to skin elastin but no statistically significant enhancement was observed (Fig. S1 and S2). We also tested for the inhibition of leptospiral binding to skin elastin by pre-treating the wells with a saturating concentration (0.5 µM) of recombinant OmpL37 alone or pre-incubated with the OmpL37 antiserum and found no effect (Fig. S3), indicating that additional leptospiral surface proteins could mediate elastin binding or that the avidities of recombinant and native cellular OmpL37 are significantly different.
To investigate whether OmpL37 can elicit an immune response from an infected host, we examined the reactivity of sera from L. interrogans L1-130-infected hamsters and leptospirosis patients to recombinant OmpL37 by immunoblot (Fig. 5). The results show that sera pooled from two L1-130-infected hamsters recognized OmpL37 and OmpL36 but not OmpL47 (Fig. 5B). Control hamster serum obtained after sham infection with sterile EMJH medium did not react to any of the proteins tested (Fig. 5B). Similar results were obtained when sera from acute and convalescent leptospirosis patients were tested (Fig. 5C). Pooled sera from 13 individuals diagnosed with either acute leptospirosis or recovering from the disease recognized OmpL37 and OmpL36, while no significant reactivity was found for OmpL47 and pooled normal human serum did not react to any of these proteins (Fig. 5C).
The expression of OmpL37 by pathogenic and saprophytic Leptospira serovars was investigated by immunoblot using OmpL37 antiserum (Fig. 6A). Pathogenic Leptospira included L. interrogans serovars Copenhageni strain Fiocruz L1-130 and Pomona strain PO-01, L. kirschneri serovars Mazdok strain 5621 and Grippotyphosa strain RM 52, and L. borgpetersenii serovars Tarrasovi strain Perepelicin and Javanica strain Veldrat Bataviae 46. Saprophytic Leptospira and Leptonema included Leptonema illini strain 3055, Leptospira weilii serovar Celledoni strain Celledoni, L. wolbachii serovar Biflexa strain codice, L. inadai serovar Lyme strain 10, and L. biflexa serovar Patoc strain Patoc 1. Loading of equal amounts of proteins from whole cell lysates was confirmed by Coomassie Brilliant G-250 staining (6B). As expected, the immunoblot revealed the highest reactivity of OmpL37 antibodies against L. interrogans L1-130 (Fig. 6A). OmpL37 was detected in all other pathogenic Leptospira strains investigated with various levels of band intensity reflecting variations in either expression or antiserum reactivity (Fig. 6A). Interestingly, the OmpL37 homolog in L. interrogans serovar Pomona (LIP_1392, 99% identity) is expressed in considerably lower amounts (Fig. 6A). OmpL37 was not detected in any saprophytic strains investigated, with the exception of Leptonema illini, where a very weak band could be detected (Fig. 6A). In L. biflexa a homolog of OmpL37 has been annotated (LBF_0995), but the amino acid sequence reveals only 47% identity (Fig. S4), and is either not recognized by the OmpL37 antibodies or is not expressed in L. biflexa under the in vitro conditions used (Fig. 6A). The genomes of the other Leptospira serovars we tested have not been sequenced precluding identification of OmpL37 homologues in those organisms. Taken together, our results show that OmpL37 was detectable only in pathogenic Leptospira spp.
Bacterial adhesins are surface-exposed OMPs, often playing roles as determinants of pathogenicity that allow bacteria to colonize host tissues by attaching to host molecules, such as ECM proteins [14], [15]. L. interrogans is an extracellular pathogen that enters the susceptible host through skin abrasions and mucous membranes, disseminates through the bloodstream, and colonizes kidneys, lungs, and other organs. The capacity of L. interrogans to adhere to tissues of various organs requires surface-exposed proteins with high affinity for host cell-surface and ECM components. The abundance of elastin in the inner layers of skin (reticular region of the dermis), blood vessels, and lungs implies that Leptospira have the ability to recognize and attach to elastin. Whereas various leptospiral proteins have been shown to have the capacity to bind multiple ECM components, such as laminin, fibronectin, fibrinogen, and collagens [18], [19], [22]–[30], [32], only the LigB repeated domains have been shown to exhibit elastin-binding activity [21]. Thus, it is of interest to identify any additional elastin-binding proteins. We previously identified four novel surface-exposed transmembrane OMPs, OmpL36, OmpL37, OmpL47, and OmpL54 [33] and have been investigating whether these proteins bind to host proteins. The recombinant OmpL36, OmpL37, and OmpL47 proteins are soluble and suitable for functional analysis by ELISA.
The binding activity of OmpL37 for the host components tested varied widely, with the most significant binding to human skin elastin, followed by human aorta elastin, laminin, fibrinogen fragment D, fibrinogen, and plasma fibronectin (Fig. 1). The weakest binding was observed for chondroitin sulfate, heparan sulfate, and collagen type IV (Fig. 1). Although we used BSA as a widely accepted negative control for host-ligand binding, it is important to note that albumin occurs in blood and the OmpL37 binding observed with BSA and fetuin may be important in host-pathogen interactions that require further investigation. This is the first report of a leptospiral protein showing pronounced specificity for human skin elastin. The strong dose-dependent, saturating binding to skin elastin gave an estimated apparent Kd of 104±19 nM (Fig. 2 and Table 1), which is one of the highest affinities for all leptospiral ligand-binding proteins investigated to date, comparable to that of LigB U1 binding to fibronectin and fibrinogen with Kd of 72.6±11.7 nM and 87.1±2.4 nM, respectively [18], LenB binding to fibronectin and laminin with Kd of 106±8 nM and 118±39 nM, respectively [30], and LigB Cen binding to lung elastin with Kd of 101±11 nM [21]. It is apparent that LigB exhibits strongest affinity for fibronectin and fibrinogen when compared to its affinity for lung elastin and other host proteins tested [18], [21]. It is noteworthy that OmpL37 showed much stronger affinity for human skin and aorta elastin (Fig. 2 and Table 1) compared to the binding of lung elastin by multiple parts of LigB, with the exception of LigB Cen [21].
Dose-dependent, saturating binding by OmpL37 was also observed for fibrinogen and its fragment D, with considerably weaker activity for plasma fibronectin and its 30-kDa and 45-kDa fragments (Fig. 2 and Table 1). Note that whereas OmpL37 seemed to interact very efficiently with laminin (Fig. 1), the reaction kinetics revealed barely saturating binding with Kd greater than 410±81 nM (Fig. 2 and Table 1). Our results show that OmpL37 interacts with multiple host proteins, a property that is common with previously characterized ligand-binding proteins of Leptospira [18], [19], [22]–[30], [32].
L. interrogans was recently shown to bind to elastin in vitro and LigB has been attributed a partial role in this binding, and it was proposed that there are additional elastin-binding proteins [21]. This observation and our results showing strong elastin binding by recombinant OmpL37 prompted us to investigate the contribution of OmpL37 to leptospiral adherence to elastin. The addition of saturating amounts of recombinant OmpL37 did not affect L. interrogans binding to immobilized human skin elastin, raising the possibility that the avidity of recombinant OmpL37 for elastin is different from that of the native OmpL37 on the cell surface, affecting the ability of the recombinant protein to compete with native OmpL37. However, we observed significant increases in leptospiral elastin binding in the presence of rabbit antiserum for OmpL37, with the effect being concentration dependent (Fig. 4A). In contrast, antiserum for the surface-exposed OmpL54 did not affect elastin binding by the spirochetes (Fig. 4A), suggesting that antiserum specific for OmpL37 can promote the interaction of OmpL37 on the cell surface with elastin. Convalescent leptospirosis patient sera were also tested for effects on leptospiral binding to elastin, but the strong agglutinating effect prevented assessment of leptospiral adhesion.
Further, the OmpL37 antiserum but neither leptospirosis patient sera nor Leptospira-infected hamster sera enhanced the binding of recombinant OmpL37 to immobilized skin elastin (Fig. 4B and Fig. S1–S2). The discrepancy might be due to the different nature of the antibodies for OmpL37, with antibodies in rabbit serum capable of recognizing additional epitopes not recognized by antibodies generated in the infected hosts. As with the enhancement of leptospiral adhesion, the effect on recombinant OmpL37 was also specific for the OmpL37 antiserum, since rabbit antiserum for OmpL47 did not produce an effect (Fig. 4B). Interestingly, purified IgG specific for OmpL37 did not exhibit statistically significant enhancement compared to that of antiserum for OmpL37 (Fig. 4B). This discrepancy did not appear to be due to inactivation of IgG molecules as we confirmed their recognition of OmpL37 by immunoblot (data not shown). However, we cannot exclude the possibility that some level of degradation or partial loss of IgG activity could have occurred. The possibility that the enhancement effect is due to another Ig type is not likely as the IgG purification process also yields up to 20% of IgA and IgM molecules. Since antisera for other surface-exposed leptospiral proteins did not exhibit enhancement (Fig. 4A and B), we hypothesize that another serum component in addition to OmpL37 antibody might be required to enhance OmpL37 binding to elastin. Further, the OmpL37 antiserum did not enhance recombinant OmpL37 binding to fibronectin, fibrinogen, and laminin (data not shown), suggesting that this effect is specific to elastin. Finally, the OmpL37 antiserum enhanced immobilized Leptospira binding to freely soluble skin elastin (data not shown).
The enhancement of adhesin binding to host ligands by antibodies has been observed previously for fibronectin-binding protein FnbA of Streptococcus dysgalactiae [44], fibronectin-binding protein FnBPA of Staphylococcus aureus [45], and Lewis X (Lex) expressed by both human gastric mucosa and Helicobacter pylori [46]. Antibodies recognizing the fibronectin-binding site, Au, of FnbA bound to Au only in the presence of fibronectin [44]. Moreover, the Au antibodies enhanced ligand binding by recombinant proteins or synthetic peptides containing the Au sequence, suggesting that the antibodies can stabilize the ligand-adhesin complex, resulting in enhanced fibronectin binding. This would provide an advantage to the pathogen, where antibodies could enhance adherence to host tissue rather than inhibit this critical step in tissue colonization [44]. Interestingly, although the fibronectin-binding repeats of FnBPA are recognized weakly by antibodies in the absence of fibronectin, epitopes induced in FnBPA by its interaction with fibronectin greatly enhance its recognition by both monoclonal antibodies and sera from patients with staphylococcal infections [45]. In H. pylori it has been shown that Lex antibodies can specifically increase bacterial adhesion [46]. This potentially positive effect on colonization could be due to an increase in bacterial aggregation or to the antibodies mediating a bivalent H. pylori Lex – antibody– human gastric mucosa Lex interaction that forms a bridge between bacteria and host cells [46]. Our interesting but somewhat surprising finding that OmpL37 antiserum enhances leptospiral binding suggests that OmpL37 contributes to the adhesion of leptospires to skin elastin. Since the OmpL37 antibodies do not promote leptospiral agglutination, we hypothesize that they in conjunction with another serum component alter the conformation of OmpL37 to promote more efficient binding to elastin. Alternatively, the antibodies may stabilize the elastin-OmpL37 complex as described for FnbA [44], resulting in enhanced elastin binding and leptospiral adhesion. The exact mechanism and functional role of this enhancement needs further investigation, which is currently under way in our laboratory.
To support our hypothesis that antibodies against OmpL37 may aid Leptospira during the infection process, we wanted to verify that there is a host immune response towards OmpL37. We found that an antibody response to OmpL37 occurs in both the hamster model and leptospirosis patients (acute and convalescent) (Fig. 5). Further, we investigated whether OmpL37 is expressed in pathogenic leptospiral isolates and found that OmpL37 could be detected only in pathogenic Leptospira serovars (Fig. 6), albeit a homologue is present in the L. biflexa genome (Fig. S4). Given that the amino acid sequences of OmpL37 in serovars Copenhageni and Pomona are 99% identical, the lower immunoblot reactivity indicates that OmpL37 is expressed at much lower levels in serovar Pomona (Fig. 6).
It has been reported that OmpL36 (LIC13166) and OmpL47 (LIC13050) are potential virulence factors of L. interrogans that are recognized by leptospirosis patient sera [47]. However, we observed the recognition of OmpL36 but not OmpL47 by the sera from infected hosts (Fig. 5). This could be due to the different techniques (ELISA versus immunoblot) employed in the studies. In addition, proteomic analysis of total protein extracts separated by 2D-gel electrophoresis identified OmpL36 and OmpL47 along with other well described leptospiral OMPs in virulent L. interrogans serovar Pomona cultured from infected hamsters [48]. The failure of these proteomic studies to detect OmpL37 may be explained by another mass spectrometry study focusing on cellular protein concentrations in L. interrogans serovar Copenhageni, which has revealed that OmpL36 is the most abundant transmembrane OMP (13th highest of all cell proteins), OmpL47 is very abundant (29th highest), and OmpL37 is less abundant (204th highest) [49]. The lower abundance of OmpL37 in serovar Copenhageni compared to OmpL36 and OmpL47 and our data showing lower expression of OmpL37 by serovar Pomona compared to serovar Copenhageni suggest that there could be insufficient OmpL37 for detection by proteomic analysis of total proteins separated by 2D-gel electrophoresis. Although OmpL47 had very little ligand-binding activity and OmpL36 did not show significant ligand-binding activity (Fig. 1), the involvement of these proteins in virulence awaits further investigation.
Based on the results presented here, we hypothesize that OmpL37 might be involved in the pathogenesis of leptospirosis. During the initial stage of infection, the strong binding affinity of OmpL37 for human skin elastin would facilitate the attachment of leptospires to an inner elastin-rich layer of the skin exposed by abrasion. Later during dissemination, the ability of OmpL37 to efficiently bind human aorta elastin suggests that OmpL37 could promote leptospiral attachment to elastin-rich vascular structures, including the walls of the pulmonary, cardiac, and other blood vessels, possibly explaining the propensity of leptospirosis to result in hemorrhagic complications [1], [50], [51]. The enhancement of leptospiral binding to elastin by OmpL37 antiserum supports the hypothesis that the host immune response to OmpL37 could promote rather than inhibit the adherence of infecting spirochetes to elastin-rich tissues, possibly also aiding in evading immunological clearance. Future studies are planned to map the elastin-binding sites in OmpL37 to further understand the antiserum enhancement effect. We also plan to identify the serum component that is apparently required in addition to OmpL37 antibodies in the enhancement of elastin-binding. Finally, we plan to study the contribution of OmpL37 during the initial and subsequent stages of leptospirosis.
|
10.1371/journal.pntd.0001323 | Luciferase-Expressing Leishmania infantum Allows the Monitoring of Amastigote Population Size, In Vivo, Ex Vivo and In Vitro | Here we engineered transgenic Leishmania infantum that express luciferase, the objectives being to more easily monitor in real time their establishment either in BALB/c mice—the liver and spleen being mainly studied—or in vitro. Whatever stationary phase L. infantum promastigotes population—wild type or engineered to express luciferase—the parasite burden was similar in the liver and the spleen at day 30 post the intravenous inoculation of BALB/c mice. Imaging of L. infantum hosting BALB/C mice provided sensitivity in the range of 20,000 to 40,000 amastigotes/mg tissue, two tissues—liver and spleen—being monitored. Once sampled and processed ex vivo for their luciferin-dependent bioluminescence the threshold sensitivity was shown to range from 1,000 to 6,000 amastigotes/mg tissue. This model further proved to be valuable for in vivo measurement of the efficiency of drugs such as miltefosine and may, therefore, additionally be used to evaluate vaccine-induced protection.
| Leishmania infantum/L. chagasi parasites are inoculated in the skin of mammals by sand flies. Though most often these L. infantum-mammal interactions are asymptomatic, they can proceed, in some individuals, to a systemic disease known as visceral leishmaniasis. If left untreated this disease is fatal. The lack of protective or curative vaccines and the limited number of parasite-targeting drugs were incentive to set up experimental conditions that could allow easy monitoring of the fluctuation of the population size of parasites in living laboratory animals. Thus, in the present report, we depict two distinct readout assays that rely on a population of L. infantum we genetically engineered for stably expressing the firefly luciferase gene. These transgenic parasites were either inoculated to BALB/c mice or added to a culture of monocytic cells. Post intravenous inoculation, BALB/c mice were imaged over time, with special attention being given to the liver and the spleen. The sensitivity of this technique ranged from 20,000 to 40,000 parasites/mg of tissue and from 1,000 to 6,000 parasites/mg tissue, for in vivo and ex vivo measurements, respectively. Though preliminary, the data, relying on monocytic cells, are promising for further in vitro screening of small compound libraries.
| Leishmania are obligate intracellular dimorphic protozoan parasites that cause a broad spectrum of clinical diseases in mammalian hosts. Visceral leishmaniasis, due to L. infantum, is endemic in the mediterranean basin (Mediterranean Visceral Leishmaniasis, MVL) and is a fatal disease, if untreated. To date, no efficient vaccine exists against human MVL and therapeutic options for managing MVL are limited with significant toxicity in some cases.
Exploring novel molecules for use as leishmanicidal drugs or vaccines necessitates experimental models such as in vitro culture of mouse or human-derived macrophages or laboratory susceptible mice. The standard method for monitoring infection in the mouse model is based on the estimation of parasite loads in target organs such as liver, spleen, or lymph nodes by microscope examination of touch imprints or smears. Alternatively, limiting culture dilution or qPCR amplification of parasite DNA is performed [1]. These techniques are however cumbersome and require large groups of mice to be euthanized to follow the efficiency of leishmanicidal drugs or time-course efficacy of vaccines against MVL. To overcome these drawbacks, a few real time monitoring methods, using reporter genes encoding GFP or firefly luciferase, have been developed for in vitro drug screening [2]–[7] or in vivo individual follow-up of Leishmania infection in the mouse model [8]–[13]. Most in vivo studies have focused on dermotropic species. Only a limited number of experiments have been conducted concerning visceral species (L. donovani and L. chagasi), which target deep organs [10], [13]. In particular, no complete study using transgenic L. infantum has until now been reported.
In this paper, we have engineered transgenic L. infantum stably expressing luciferase. We then aimed to assess their usefulness for monitoring - in vivo and ex vivo - L. infantum biomass overtime in liver and spleen of BALB/c mice inoculated with a high parasite dose. Particular attention has been paid to the suitability - in vitro and in vivo - of recombinant L. infantum-expressing luciferase for screening the efficiency of leishmanicidal drugs such as miltefosine.
Six to eight week-old female BALB/c mice were purchased from Charles River (France). Mice were maintained and handled according to the regulations of the European Union, the French Ministry of Agriculture and to FELASA (the Federation of Laboratory Animal Science Associations) recommendations. Experiments were approved by the ethics committee of the Nice School of Medicine, France (Protocol number: 2010-45).
L. infantum MON-1 (MHOM/FR/94/LPN101), was isolated from a patient with MVL contracted in the Nice area (South of France). L. infantum promastigotes were routinely grown at 26°C in M199 medium supplemented with adenosine 0.1 mM, biotin 1 µg/ml, bovine hemin 5 µg/ml, streptomycin 100 µg/ml, penicillin 100 U/ml, 2 µg/ml biopterin, L-glutamine 2 mM, folic acid 10 µg/ml and 10% fetal calf serum (culture medium) [14].
L. infantum clones encoding firefly luciferase were generated as previously described [8]. Briefly, the 1.66 kb coding region of firefly luciferase was cloned into the Leishmania expression vector pF4x1.HYG, with marker gene for selection with hygromycin B. Following linearization by SwaI restriction digest, an insert containing 18S rRNA flanked with the luciferase gene and hygromycin, was prepared for integration into the 18S rRNA locus of the nuclear DNA of L. infantum. L. infantum promastigotes (0.5×108), in exponential growth phase, were suspended in 0.5 ml of cytomix buffer (25 mM HEPES pH 7.5, 0.15 mM CaCl2,120 mM KCl,10 mM KH2PO4, 2 mM EDTA, 5 mM MgCl2). Transfections were performed by electroporation (Gene Pulser X cell System, Biorad) using 3 µg of DNA inserts with the following conditions: 25 µF, 1600 v, 9 ms in 4 mm cuvette. Following electroporation, transfected parasites were cultured in complete culture medium and plated on semi-solid medium containing 100 µg/ml of hygromycin B. L. infantum colonies were collected, expanded in culture medium with hygromycin, aliquoted and frozen in liquid nitrogen in 90% fetal calf serum with 10% DMSO until use. Before inoculation experiments, a single promastigote clone termed LUC-parasite was passaged twice in M199 medium containing hygromycin. Occasionally, to maintain virulence, the LUC-parasite was injected by intra-peritoneal (IP) route to BALB/c mice and two months later, spleen or liver was collected and cultured as a source of promastigotes.
Variable inocula of the stationary phase LUC-parasites or wild type (WT) promastigotes (ranging from 0.12×108 to 3×108 parasites) in stationary phase of growth, were injected by intravenous (IV) route to groups of 4 of 6–8 week-old female BALB/c mice. One month later, the animals were euthanized by CO2 and spleens and livers were collected. Spleen and liver aliquots were homogenized at 100 mg/ml in PBS containing 1% Nonidet P40 and a protease inhibitor cocktail (Roche), and parasite loads were quantified by an ELISA-based method as previously described [15], [16]. For follow-up studies, mice were infected as above, and at different time-points imaging was performed. Occasionally, mice were infected by the intraperitoneal route (IP) with 500×106 stationary phase LUC-parasites and regularly imaged.
Mice infected with LUC-parasites were periodically imaged using the Photon Imager (Biospace Lab, France) system as follows: mice were administered with luciferin (Caliper life science, France) at 300 mg/kg by IP route, and within 10 min, the animals were anesthetized in 5% isoflurane/1L O2.min−1 atmosphere. The animals were then placed in the imaging chamber of the Photon Imager and anesthetisis was maintained using 2.0% isoflurane/ 0.2 L O2 per mouse min-1 atmosphere. Acquisition of emitted photons, with a charge-coupled device camera, was monitored over a 20 min period in previously defined regions of interest (ROI) that delimited the surface of analysis. To standardize imaging, and to allow comparison between mice, the images presented in the figures correspond to an acquisition of 1 min duration, taken once luminescence plateaued. In some experiments, bioluminescence imaging (BLI) localization of transgenic luciferase expressing L. infantum amastigotes to mouse liver, spleen or other sites was confirmed by reimaging mouse after dissection.
In vitro luminescence activity was measured using the Luciferase Assay System E1501 (Promega). To quantify ex vivo parasite burdens by luminescence, Nonidet P40 extracts of spleen or liver, prepared as above, were serially diluted with the reporter lysis buffer. Luminescence was quantified from 10 µl aliquot (1 mg tissue) using a luminometer (Centro LB 960 Berthold Technologies, Germany). To compare luciferase activity of promastigote and amastigote forms, the exponentially growing LUC-promastigote clone, and spleen or liver aliquots from infected mice, were extracted with NP 40 buffer and parasites number was quantified by ELISA. Detergent extracts were serially diluted and luminescence activity was analysed as above.
The THP-1 cell line (ATCC TIB-202) was routinely grown in RPMI 1640 medium containing 10% fetal calf serum and differentiated overnight at 106/ml density with 10 ng/ml of phorbol myristate acetate (PMA) in 96-well microtiter plate. Wells were then washed with RPMI and fresh complete medium was added. PMA-differentiated THP-1 cells were infected with stationary phase LUC-parasites, at parasite to THP-1 cell ratio of 10∶1, for 3 h at 37°C. Free parasites were aspirated off and the plate was incubated for 48 h in culture medium for amastigote transformation. Miltefosine (hexadecylphosphocholine, Sigma) was added at concentrations ranging from 10−8 to 10−4 M and within 48 h the plate was washed and the wells were extracted using 50 µl of reporter lysis buffer. Luminescence activity was quantified from 20 µl aliquot as described above. Cell viability was evaluated using Trypan blue exclusion.
In order to assess the infectivity of the L. infantum luciferase parasites (LUC-parasite), as well as the usefulness of bioluminescence for the monitoring of parasite proliferation in target organs, BALB/c mice were inoculated by IV route with various inocula of the stationary phase LUC-parasites or WT parasites. One month following inoculation, when generally both liver and spleen are infected, mice were imaged and sacrificed. Detergent extracts from liver or spleen were prepared for parasite quantification by ELISA or bioluminescence analysis ex vivo. Figure 1A shows that mice inoculated with increasing numbers of the LUC-parasite or WT parasites exhibited increasing and similar spleen and liver parasite burdens. This indicates that the selected LUC-parasite exhibits an infectivity identical to that of WT parasites (Fig. 1A). Identical results were obtained with another LUC-parasite (data not shown). In order to evaluate the threshold sensitivity of BLI, LUC-parasite infected mice were imaged and the luminescence (expressed as photons/s/cm2) present in a ROI corresponding to liver or spleen was recorded. Despite variable infection levels in mice infected with the same parasite dose, imaging of bioluminescence signals confirmed the parasite dose-dependency of the infection measured by ELISA (Fig. 1A). Luminescence measured in the ROI and express as photon/s/cm2, was directly proportional to parasite density, and the generated regression curves were used for estimating the threshold sensitivity of BLI for both organs (Fig. 1B). The threshold sensitivity of BLI was calculated as luminescence three times stronger than background luminescence of naïve mice, which exhibit the same background values as non-bioluminescent parasite infected mice. In our hands, threshold sensitivity was around 20,000 and 40,000 parasites per mg spleen and liver, respectively (above the grey zone of Fig. 1C).
We next assessed the possibility of measuring parasite loads by ex vivo analysis of organ luminescence. L. infantum parasites are present in the host as intracellular amastigotes, and are the form targeted by drugs and vaccines. As these amastigote forms generally display a lower metabolic activity than exponentially growing promastigotes we compared in vitro luciferase activity of both parasite forms. Figure 1D shows the dose-response curves drawn from luciferase parasites under promastigote or spleen and liver amastigote forms. As expected, luciferase activity of amastigotes was notably lower than that of exponentially growing promastigotes with threshold sensitivities (calculated as twice the background value) at around 10, 1,000 and 6,000 parasites for metacyclic promastigotes, spleen and liver amastigotes, respectively (Fig. 1D). The specificity of ex vivo analysis was confirmed by the lack of luminescence activity of equal numbers of spleen and liver WT amastigotes. Analysis of luminescence ex vivo thus facilitates a rapid and simple mean for quantitating parasite burdens in mouse organs provided a relatively high infection level is present. Finally, parasite density of samples taken from 36 livers or spleens were analysed both by luminescence analysis and ELISA. Strong correlation between the two methods verified the reliability of the bioluminescence assay (Fig. 1E).
BALB/c mice inoculated with L. infantum WT parasites generally display a liver infection episode, which peaks 1–2 months post inoculation and is partially self-resolved because of granuloma formation [17]. Partial liver clearance is followed by progressive and chronic spleen infection with a concomitant destruction of spleen architecture [17]. We used this model of infection to assess the usefulness of bioluminescence to follow this process. For this, BALB/c mice were intravenously infected with 3×108 and 1×108 stationary phase promastigotes of the LUC-parasite, and at different time-points, the spread of infection was recorded following IP injection of the luciferin substrate. Time-course infection followed by BLI (Fig. 2A) revealed the classical transitory hepatic episode (day 14) followed by parasite spleen colonization (day 40) (Fig. 2A). Of note, in the two heavily infected mice (3×108), inguinal and mediastinal lymph nodes were persistently bioluminescent (Fig. 2A). Conversely, bone marrow parasite localisation could not be evidenced.
Next, in order to validate this technique for drug screening purposes, either in vivo or in vitro, we conducted experiments using miltefosine as a leishmanicidal drug. For in vivo testing, infected mice were treated with miltefosine for 5 days (1 mg per mouse per day) and imaged 2 days later. Figure 2A shows that parasite decrease can be readily detected by BLI imaging in vivo, as luminescence was undetectable in spleen or lymph nodes following 5 days of treatment (Fig. 2A). This reduction was seen regardless of the infection level before treatment and was confirmed following dissection of mice and measurement of spleen burdens (Fig. 2A). Therefore, parasite clearing by miltefosine in spleen was accompanied by the disappearance of luminescence.
Similarly, luminescence was used in vitro to evaluate the efficacy of miltefosine on intracellular amastigotes. PMA-differentiated THP-1 cells were infected with metacyclic LUC-parasites and within 48 h, a time sufficient for amastigote transformation (data not shown), miltefosine was added. 48 h later, THP-1 cells were lysed and luciferase activity was quantified. Figure 2B shows the drug-concentration dependency of amastigote killing by miltefosine (Fig. 2B). The IC 50 (1.5 µM) was identical to that previously reported [18]. Using this method, drug-concentration dependency of amastigotes killing could be measured (Fig. 2B). This protocol may therefore provide a high throughput method for screening drugs.
Preliminary experiments showed that BALB/c mice infected via the IP route with 5×108 LUC-parasites displayed hepatic and spleen infection episodes similar to that observed following IV challenge (data not shown). However, infection levels in target organs were generally lower than that observed following infection by the IV route. Nevertheless, by day 40 post inoculation, peritoneum bioluminescence, clearly distinct from spleen bioluminescence (Fig. 3A, 3B), was detectable in nearly all infected mice and was still observable 3 months post inoculation (data not shown). Dissection of mice showed that peritoneal bioluminescent parasites localized to a mass of adipose tissue located above the intestine (Fig. 3C, 3D). Interestingly, BLI and quantitation of amastigotes in adipose tissue indicated that compared to spleen, 20 times more photons per parasite could be recovered from peritoneum amastigotes. Therefore, BLI can be useful to demonstrate unexpected sites of parasite proliferation.
The BALB/c mouse model has been widely used for drug screening and vaccine trials, since therapeutic effects of drugs or vaccine-induced protection are evaluated keeping the contextual influences of the living animal. However, estimation of parasite loads at different time-points in target organs necessitates euthanizing and dissecting animal groups, thus limiting longitudinal studies devoted to the therapy of MVL. In addition, parasite loads are extrapolated from parasite counts obtained on a limited sample size, which is not always representative of the burden occurring in the organ. Finally, spread of parasites to an unexpected site of infection may be missed because the infected tissue is not harvested or analysed.
Recently, the possibility of labelling invasive microorganisms with reporter genes, such as firefly luciferase has provided, the ability to trace the infection dissemination at the tissue/organ level by BLI in animal models. BLI potentially presents many advantages over conventional methods of infection monitoring. BLI technology allows detection of only live, metabolically active cells and because of its non-destructive and non-invasive nature it can be performed repeatedly. BLI thus permits each animal to be used as its own control over time, overcoming the problem of animal-animal variations. Finally, as already reported, by using this technique new microorganism localization can be evidenced [19].
We report here the generation of recombinant luciferase L. infantum parasites and their use in tracing parasite dissemination in vitro and in vivo. Our results show that luciferase transfected parasites are suitable for both in vitro and in vivo studies. While light emission of the amastigote form was reduced as compared to that of extracellular promastigotes it was sufficient to provide an accurate and rapid estimation of parasite loads ex vivo with a sensitivity of around 1 to 6,000 amastigotes/mg tissue, which is sufficient for most applications. Importantly, luciferase parasite clones proved to be also useful tools to measure drug efficacy of miltefosine in vitro and in vivo at the amastigote stage. Luminescent L. infantum parasites may thus represent a high throughput method for leishmanicidal drug testing on different target organs in the context of the whole body.
BLI of BALB/c mice infected with luciferase clones showed that in our conditions, the threshold sensitivity of parasite burdens was around 20×103 to 40×103 parasites/mg for spleen and liver. This indicates that BLI of deep organs requires relatively high parasite loads. In our hands, these infection levels could be easily reached by IV route using high parasite inoculum not enriched in mammal-invasive metacyclic promastigotes. Following IP parasite delivery we noticed a new site of parasite development located in the peritoneum. The demonstration of this unexpected localization, which has not been observed in mice infected by IP with L. donovani [8], suggests that this localisation may be species specific. The high photon emission observed in adipose tissue (numbers of photon/parasite) as compared to spleen emphasizes the fact that, due to different local environments, luminescence emitted by different organs cannot be compared but for a given organ it reflects differences in parasite loads. Absorption and scattering of light by overlaying tissues may account for this difference though it is likely that the delivery of luciferin directly into the peritoneum facilitated better photon emission from the peritoneum. Collectively we demonstrate that BLI represents a versatile tool for drug screening in vitro or in vivo as well as for assessing the potential protectivity of vaccine preparations. More sensitive cameras and/or luciferase with improved catalytic activity or a higher vector expression [20], will undoubtedly improve the performance and ability of BLI of L. infantum.
|
10.1371/journal.ppat.1004347 | Transgenic Analysis of the Leishmania MAP Kinase MPK10 Reveals an Auto-inhibitory Mechanism Crucial for Stage-Regulated Activity and Parasite Viability | Protozoan pathogens of the genus Leishmania have evolved unique signaling mechanisms that can sense changes in the host environment and trigger adaptive stage differentiation essential for host cell infection. The signaling mechanisms underlying parasite development remain largely elusive even though Leishmania mitogen-activated protein kinases (MAPKs) have been linked previously to environmentally induced differentiation and virulence. Here, we unravel highly unusual regulatory mechanisms for Leishmania MAP kinase 10 (MPK10). Using a transgenic approach, we demonstrate that MPK10 is stage-specifically regulated, as its kinase activity increases during the promastigote to amastigote conversion. However, unlike canonical MAPKs that are activated by dual phosphorylation of the regulatory TxY motif in the activation loop, MPK10 activation is independent from the phosphorylation of the tyrosine residue, which is largely constitutive. Removal of the last 46 amino acids resulted in significantly enhanced MPK10 activity both for the recombinant and transgenic protein, revealing that MPK10 is regulated by an auto-inhibitory mechanism. Over-expression of this hyperactive mutant in transgenic parasites led to a dominant negative effect causing massive cell death during amastigote differentiation, demonstrating the essential nature of MPK10 auto-inhibition for parasite viability. Moreover, phosphoproteomics analyses identified a novel regulatory phospho-serine residue in the C-terminal auto-inhibitory domain at position 395 that could be implicated in kinase regulation. Finally, we uncovered a feedback loop that limits MPK10 activity through dephosphorylation of the tyrosine residue of the TxY motif. Together our data reveal novel aspects of protein kinase regulation in Leishmania, and propose MPK10 as a potential signal sensor of the mammalian host environment, whose intrinsic pre-activated conformation is regulated by auto-inhibition.
| Leishmaniasis is an important human disease caused by Leishmania parasites. A crucial aspect of Leishmania infectivity is its capacity to sense different environments and adapt for survival inside insect vector and vertebrate host by stage differentiation. This process is triggered by environmental changes encountered in these organisms, including temperature and pH shifts, which usually are sensed and transduced by signaling cascades including protein kinases and their substrates. In this study, we analyzed the regulation of the Leishmania mitogen-activated protein kinase MPK10 using protein purified from transgenic parasites and combining site-directed mutagenesis and activity tests. We demonstrate that this kinase is activated during parasite differentiation and regulated by an atypical mechanism involving auto-inhibition, which is essential for parasite viability.
| Leishmaniasis is an infectious disease characterized by a variety of pathologies, affecting more than 12 million people worldwide and ranging from self-healing cutaneous lesions to fatal visceral infection [1]. This disease is caused by pathogenic protozoa of the genus Leishmania, which show two major life cycle stages depending on the host. The extracellular promastigote stage develops inside the midgut of sandflies and is transmitted during blood feeding to a vertebrate host where they are ingested by phagocytic cells, notably macrophages. Inside the host cell phagolysosome, promastigotes develop into proliferating intracellular amastigotes. These developmental transitions are triggered by environmental changes, mainly pH (7.4 to 5.5) and temperature (26°C to 37°C), encountered in insect and vertebrate hosts, respectively, and can be mimicked in vitro [2]–[4]. Interfering with amastigote stage development and proliferation by altering the parasite's ability to sense its environment could be a very efficient way to eliminate intracellular Leishmania and thus signaling proteins involved in extra- or intracellular signal transduction are interesting drug target candidates.
In eukaryotes, environmental signals are generally sensed and transduced by signaling cascades involving receptors and downstream-regulated protein kinases. The MAPK signaling pathway is a good example of such a phosphorylation cascade [5] as it is composed of mitogen-activated protein kinase kinase kinases (M3Ks), which activate mitogen-activated protein kinase kinases (M2Ks), which in turn activate mitogen-activated protein kinases (MAPKs) by dual phosphorylation on the highly conserved TxY motif present within the MAPK activation loop [6], [7]. MAPKs regulate various important cellular functions, such as cell cycle progression and differentiation, through phosphorylation of a large number of substrates, including transcription factors and MAPK-activated protein kinases, thus modifying gene expression and post-translational regulation, respectively [8]–[14]. While the core cascade M3K-M2K-MAPK is conserved in Leishmania, the absence of classical transcription factors and the largely constitutive gene expression suggest that the response to environmental signals occurs mainly post-translationally through the regulation of the level of protein phosphorylation, rather than through modulation of gene expression [15].
As in higher eukaryotes, the Leishmania MAPK pathway comprises two distinct kinase families, the STE family, which includes five putative M2K and M2K-like and 20 putative M3K members, and the CMGC family, including 17 putative MAPK and MAPK-like members [16]. The comparison between the Leishmania major and the human kinomes revealed an evolutionary expansion relative to genome size of these two kinase families in the parasite. The STE and CMGC families represent 19% and 25% in the Leishmania kinome, compared to 9% and 12% in the human kinome, respectively [17]. These expansions indicate that the MAPK pathway could be of particular importance to parasite development and survival, a possibility that is supported by recent investigations showing that Leishmania MAP kinases are required for flagellar development, intracellular survival and viability [16],[18]–[23]. Thus, the study of Leishmania MAPKs could hold the key to the understanding of the mechanisms that allow the adaptation of Leishmania to environmental changes required for extra- and intracellular parasite survival during host infection.
The activity of the three Leishmania MAPKs MPK4, MPK7 and MPK10 have been shown to be induced in a stage-specific manner in axenic amastigotes, which occurred concomitant to an increase of their phosphorylation as expected for this class of kinases [24]–[26]. The importance of MPK4 and MPK7 for cell survival and infectivity, respectively, has been established [21], [26], [27], but neither the mode of regulation nor the functions of MPK10 have been studied. MPK10 is conserved in all Leishmania species (Figure S1A) with a percentage of identity above 90%. This percentage is even higher if we compare the sequence of L. donovani MPK10 with that of L. major, L. mexicana and L. infantum (99%, 98% and 100%, respectively, Figure S1B). The close relationship between this kinase and human ERK2 or p38 kinase suggests a potential role in cell differentiation or in response to stress [28], [29]. Leishmania MPK10 is a highly conserved member of the eukaryotic MAPK family. However, MPK10 does not resemble a classical eukaryotic MAPK with respect to two structural features. First, the alignment of the protein sequence with its mammalian orthologs revealed the presence of a long carboxy-terminal extension [16], [24]. This is a common feature for Leishmania MAPKs (observed in 15 out of 17 putative Leishmania MAPKs), whereas only five mammalian MAPKs have such an extension. They are usually regulatory domains implicated in the control of the kinase activity, localization or auto-inhibition [30]–[34]. Second, a structural analysis of MPK10 published recently by Horjales et al. provided evidence that recombinant MPK10 adopts an activated conformation, despite the absence of TxY phosphorylation [29]. Following phosphorylation of this motif by upstream M2Ks, mammalian MAPKs are activated by the switch of a conserved DFG motif from an inactive (DFG-out) to an active (DFG-in) conformation. This change leads to the alignment of two structural motifs comprising non-consecutive hydrophobic residues that are referred to as the regulatory and catalytic spines [35], [36]. In contrast to mammalian MAPKs, the Leishmania DFG motif of MPK10 is replaced by DFN, causing the R and C spines to be aligned thus stabilizing an apparent active conformation in the absence of TxY phosphorylation [29]. These findings suggest that MPK10 has a strikingly different mode of regulation compared to mammalian MAPKs, which could be particularly adapted to adjust quickly to environmental changes thus acting as a signaling switch.
Here, by utilizing a transgenic approach we unraveled an unusual mechanism of MPK10 regulation that tightly controls kinase activity. Stage-specific increase in MPK10 activity during the pro- to axenic amastigote conversion did not follow the largely constitutive phosphorylation status of the regulatory tyrosine residue in the kinase activation loop, suggesting additional and non-classical mechanisms of MPK10 regulation. Combining limited tryptic digestion with mutagenesis analysis and phosphoproteomics investigation, we uncovered an essential role of the C-terminal domain of MPK10 in limiting stage-specific kinase activity, and identified a novel regulatory phospho-serine residue that is required for axenic amastigote viability. Together our data propose MPK10 as a potential signal sensor of the mammalian host environment whose intrinsic pre-activated conformation is regulated by auto-inhibition. These results shed important new light on Leishmania-specific signaling mechanisms, and significantly advance our understanding on parasite-specific protein kinase biology.
To investigate the biochemical properties of MPK10, we generated non-mutated recombinant MPK10 (NM) and the corresponding MPK10-K51A enzymatically dead mutant, both tagged with GST-Strep. We used the bacterially expressed and purified proteins in an in vitro kinase assay at 30°C or 37°C, monitoring the transfer of radiolabeled phosphate from γ-33P-ATP to MPK10 (auto-phosphorylation) or to the canonical MAPK substrate myelin basic protein (MBP) [37]–[39]. The kinase reactions were subjected to SDS-PAGE, proteins were visualized by Coomassie staining for normalization (Figure 1A, upper panel), and phosphotransferase activity was revealed by auto-radiography (Figure 1A, lower panel). We observed a weak auto-phosphorylation signal at 30°C that was slightly enhanced when the kinase reaction was performed at 37°C (Figure 1A, lower panel). We did not observe any significant activity towards MBP at both temperatures. Likewise, performing kinase assays at 37°C and pH 7.5 with other substrates, including casein, histone H1 or Ets1, revealed only a very weak substrate-specific activity, with casein giving rise to the strongest signal, which still was faint compared to the levels of auto-phosphorylation (Figure 1B, lower panel). By contrast, recombinant human MEK1 (a M2K), used as a positive control, revealed strong substrate-specific activity towards MBP. No substrate phosphorylation or auto-phosphorylation could be detected with the kinase dead MPK10-K51A control, suggesting that the signals observed were specific for MPK10 and not due to a co-purified bacterial kinase. Based on these results, casein appears to be the most suitable substrate for recombinant MPK10.
We next varied the pH of the kinase assay in an attempt to improve MPK10 activity. Kinase assays performed with casein and MPK10 or MPK10-K51A at pH 5.5, 6.5, 7.5 and 8.5 revealed different pH optima for auto- and substrate-specific phosphorylation at pH 6.5 and 7.5, respectively (Figure S2, lower panel). In conclusion, recombinant MPK10 shows some minor auto-phophorylation activity, but largely fails to phosphorylate the canonical MAPK substrate MBP [37]–[39], which may either depend on activation through an upstream M2K, parasite-specific kinase substrate interactions, or auto-inhibition.
We performed a limited tryptic digestion of purified recombinant GST-Strep-tagged MPK10 to investigate the presence of potential auto-inhibitory accessory domains by delineating the structured, protease-resistant kinase core. We treated recombinant MPK10 with 0.25 µg of trypsin for 150 minutes, analyzing samples after 2.5, 5, 15, 30, 60 and 150 minutes by SDS-PAGE and Coomassie staining. MPK10 was sensitive to partial digestion as we observed the appearance of several bands within the first 15 minutes corresponding to different tryptic MPK10 products (Figure 2A). After 30 min of treatment, only one band remained, revealing the core of the kinase. N-terminal sequencing and SELDI-TOF analysis of four of the digestion products (as marked by arrowheads in Figure 2A and represented by the cartoon shown in Figure 2B) revealed that both ends of the tagged protein were cleaved by trypsin at lysines 12, 24, and 30, arginine 392. We also identified a digestion product that resulted from cleavage at aspartate 387 and thus lacked the last 46 amino acids of MPK10 (Figure 2C). This product is likely due to a miscleavage or cleavage by a contaminating bacterial protease.
To investigate the activity of the MPK10 kinase core alone and to test whether deletion of the C-terminal region (46 aa) increases its activity, we generated and purified recombinant non-mutated His6-MPK10 (NM) or mutated His6-MPK10 deleted for the last 46 amino acids (ΔC) and monitored their activity towards canonical substrates by performing in vitro kinase assays in the presence of γ-33P-ATP. We changed from the GST-Strep to His6 tag to show that the lack of MPK10 substrate phosphorylation activity is independent from this modification. The auto-radiogram shown in Figure 2D (left panel) confirms weak auto- and substrate phosphorylation activity of His6-MPK10 NM similar to GST-STREP-MPK10 NM ruling out tag-specific interference with kinase activity. In contrast to His6-MPK10 NM, His6-MPK10-ΔC presented a stronger auto-phosphorylation activity and enhanced phosphorylation of Ets1 and casein. These differences reflect a true increase in phosphotransferase activity as judged by Coomassie staining, which showed equal loading of both recombinant proteins and substrates (Figure 2D, right panel). The signals were specific for His6-MPK10-ΔC and not caused by a co-purified kinase, as we did not detect any signal with the His6-MPK10-ΔC_K51A control. Yet again, the auto-phosphorylation signals of NM or truncated MPK10 were significantly stronger than those for substrate phosphorylation. Moreover, no 33P incorporation could be detected for MBP. Thus, deleting the C-terminal tail substantially increases the ability of MPK10 to phosphorylate itself, suggesting a potential role of this domain in negative regulation. The C-terminal domain of L. donovani MPK10 is conserved in Leishmania species (84 to 100%, Figure S3B) but not in Trypanosoma species (38 to 53%, Figure S3B), except for the conserved motif (DHMxRTxSxME), of unknown function (underlined, Figure S3A). This motif is only conserved in trypanosomatid MPK10 as extending the analysis of the C-terminal extensions to the 14 Leishmania and 2 human MAPKs (ERK5 and ERK8) as well as T. brucei TbECK1 [32] by pattern recognition analysis (Pratt version 2.1) and multiple sequence alignment (ClustalW) did not reveal this motif or any other conserved motifs or patterns (data not shown). As our study reveals important limitations of the bacterially expressed protein with regard to kinase activity, we analyzed in the following in situ activated MPK10 isolated from transgenic L. donovani parasites.
To gain insight into MPK10 regulation and activity in a physiologically relevant context, we generated transgenic parasites expressing a GFP-MPK10 fusion protein from the episomal vector pXG-GFP2+ (kindly provided by S. Beverley). We first performed a time course experiment to investigate MPK10 activity in promastigotes and during axenic amastigote differentiation. GFP-MPK10 was purified using a monoclonal anti-GFP antibody from L. donovani promastigotes harvested from logarithmic (log) or stationary (stat) phase cultures, and from cultures at different time points between 12 h to 120 h after induction of axenic amastigote differentiation by pH and temperature shift. Immuno-purified proteins were incubated for 30 min at 37°C in the presence of radiolabeled ATP and MBP, and the kinase reaction was subjected to SDS-PAGE. The phosphotransferase activity was determined by auto-radiography (Figure 3A, a and b), and MPK10 and MBP were visualized by Coomassie staining for normalization (Figure 3A, c and d). After exposure, the bands corresponding to the signals of phosphorylated MPK10 or MBP were recovered from the dried gel and 32P incorporation was measured using a scintillation counter. The results were expressed relative to GFP-MPK10 NM from promastigotes of logarithmic culture set to 100% (the relative counts are represented by the numbers in Figure 3A a and b). As opposed to recombinant MPK10 we observed that MBP phosphorylation was higher than MPK10 auto-phosphorylation when using protein purified from parasite extracts. This finding suggests that the purification of GFP-MPK10 from transgenic parasites allows the assessment of biologically relevant kinase activity. MPK10 purified from parasites undergoing axenic amastigote differentiation showed a higher level of MBP phosphorylation compared to promastigotes, especially during the first 48 h following temperature and pH shift (202% versus 100% respectively, Figure 3A, b). Afterwards, the level of activity decreases to reach its lowest point at 96 h (31%), and increases again at 120 h (144%). These data suggest that MPK10 activity is stage-specifically regulated. The signals observed are specific to MPK10 and not due to a co-purified kinase as no signals were observed with the K51A mutant (data not shown).
The mean values with standard deviation of three independent experiments are represented by the histogram plot shown in Figure 3B. As judged by statistical analysis (see Figure S4), significant differences in MPK10 kinase activity are observed (i) between promastigotes and parasites at 48 h of axenic amastigote differentiation confirming its stage-specific regulation, (ii) between parasites at 48 h and 72 h or 96 h, demonstrating a transient reduction in MPK10 activity during later stages of axenic differentiation, and (iii) between parasites at 96 h and 144 h providing evidence for a second peak in MPK10 activity in axenic amastigotes. Conversely, no significant difference in MPK10 activity was observed between promastigotes from logarithmic or stationary phase culture.
Phosphorylation of both the threonine and the tyrosine residues of the TxY motif have been shown to be required for eukaryotic MAPK activation and dual phosphorylation is correlated with kinase enzymatic activity. As an alternative read out for MPK10 activity we therefore assessed tyrosine phosphorylation using an anti-phosphotyrosine antibody. First, we tested whether phosphorylated Tyr192 was the only residue recognized by the antibody in MPK10. We compared the signals obtained with GFP-MPK10 NM, GFP-MPK10-K51A and the corresponding TxY motif mutants T190A, Y192F and T190A_Y192F purified from respective transgenic promastigotes harvested in logarithmic growth phase (low level of MPK10 activity, Figure 3B), or cells during axenic amastigotes differentiation harvested at 48 h after induction of differentiation (high level of MPK10 activity, Figure 3Ab and B). The western blot presented in Figure 3C shows that GFP-MPK10 NM, GFP-MPK10-K51A and GFP-MPK10-T190A are recognized by the anti-phosphotyrosine antibody as documented by the detection of a strong signal at 75 kDa (Figure 3C). The absence of this signal in GFP-MPK10-Y192F and GFP-MPK10-T190A_Y192F demonstrates that pY192 is the only residue recognized by this antibody in GFP-MPK10. Moreover, Y192 phosphorylation does not require T190 phosphorylation as Y192 is phosphorylated in the GFP-MPK10-T190A single mutant. Remarkably, we did not observe any difference in the level of tyrosine phosphorylation between MPK10 purified from log promastigotes or from 48 h axenic amastigotes despite their difference in activity, suggesting that the phosphorylation state of Y192 is dissociated from the regulation of kinase activity. In contrast, no tyrosine phosphorylation was observed for recombinant MPK10 NM, which further supports the observed inactive state of the bacterially purified kinase (Supplementary Figure S5).
We next studied dissociation of MPK10 activity and Y192 phosphorylation state in a more detailed time course experiment by western blot analysis of pY192 levels in promastigotes (log and stat) and during axenic amastigote differentiation (Figure 3D). Again, no difference was observed in the phosphorylation state of Y192 between MPK10 purified from log promastigotes, stat promastigotes or axenic amastigotes during the first 48 h of differentiation, despite their significant differences in activity. After 72 h we observed a decrease in tyrosine phosphorylation, which this time correlated with decreased MPK10 activity.
These findings were supported by an independent experiment performed with Leishmania mexicana using a targeted quantitative phosphoproteomic analysis termed Selected Reaction Monitoring (SRM, [40]), from which results concerning MPK10 were extracted and presented in Figure 3E. In this analysis the phosphorylation states of the two regulatory phosphorylation sites T190 and Y192 were quantified in L. mexicana late log phase promastigotes and axenic amastigotes at 72 h of differentiation. Peptides containing T190-H-pY192 (unphosphorylated T190 and phosphorylated Y192) were found at similar levels in promastigotes and axenic amastigotes, whereas peptides containing pT190-H-pY192 (pT190 and pY192 dual phosphorylation) showed significantly increased abundance in the axenic amastigote fraction. Peptides containing a single pT190 phosphorylation were not detected. These results show that phosphorylation of T190 and Y192 occurs independently: while pY192 is identified both in promastigotes and in axenic amastigotes, pT190 is mostly identified in amastigotes and thus this phosphorylation event may be the rate limiting step for MPK10 activation. Altogether, these findings strongly suggest that the phosphorylation of the Y192 residue is largely constitutive and occurs independently from MPK10 activation, which is in contrast to MAPK regulation in most other eukaryotes. This characteristic of MPK10 is likely a conserved feature among Leishmania species as we found similar results with independent experiments performed with L. donovani and L. mexicana.
In most eukaryotes, both residues of the TxY motif need to be phosphorylated for MAPK activation and as a consequence their mutation abrogates kinase activity [41]–[43]. To analyze the requirement of T190 and Y192 phosphorylation for MPK10 activity, we used our transgenic cell lines to investigate the impact of the overexpression of the three different MPK10 TxY-motif mutants on parasite growth and survival as well as to measure MPK10 kinase activity, using GFP-MPK10 and GFP-MPK10-K51A as positive and negative controls, respectively. We first followed the growth and percentage of cell death in promastigotes by flow cytometry analysis and observed no differences between the untransfected L. donovani control (UC) and parasites over-expressing GFP-MPK10 NM or mutant forms (Figure S6). We next measured the kinase activity of GFP-MPK10 NM and GFP-MPK10 mutant proteins purified from promastigotes. The results presented in Figure 4A show that GFP-MPK10 NM undergoes weak auto-phosphorylation (exposure 3 h), but catalyzes robust MBP phosphorylation (exposure 1 h). It is interesting to note that although the auto-phosphorylation signals are much weaker than the MBP phosphorylation signals, they are similarly regulated. No MBP phosphorylation could be detected with the GFP-MPK10-K51A, -T190A, -Y192F or -T190A_Y192F mutants. These data demonstrate that, similarly to other eukaryotic MAPKs, T190 and Y192 are essential for MPK10 activity, as their mutation considerably reduced the activity of the kinase.
We next investigated the impact of these MPK10 mutants on amastigote growth and survival. As presented in Figure 4B (upper panels), untransfected control (UC) parasites showed a decrease in cell growth during the first 24 hours of differentiation, but started to grow thereafter to reach a plateau at around 96 h. This profile was the consequence of a high level of cell death at 24 h (26±13%), which decreased to reach a percentage of cell death of 15±2% at 96 h. This phenomenon has been previously documented and is likely due to the adaptation of parasites to elevated temperature and acidic pH [44]. Parasites overexpressing GFP-MPK10 NM (Figure 4B, upper panels) resumed growth after 48 h of differentiation to finally reach the same cell concentration than UC parasites, a profile similar to that obtained with GFP-MPK10-K51A. We observed a higher percentage of cell death of GFP-MPK10 NM (41±5%) compared to UC parasites. This difference is largely attributed to the over-expression of MPK10, as there is only a slight yet statistically significant difference between the percentage of cell death of untransfected parasites and transgenic parasites expressing the empty vector during axenic amastigote differentiation (supplementary Figure S7). This finding, which was not observed in promastigotes, indicates that the over-expression of MPK10, regardless of its activity, is somewhat detrimental to axenic amastigotes.
We next compared the phenotype of strains over-expressing GFP-MPK10 to those over-expressing GFP-MPK10-Y192F, -T190A or -T190A_Y192F (Figure 4B, lower panels). All mutant strains showed a 24 hours growth delay, which corresponded to a higher percentage of cell death compared to that observed for GFP-MPK10 NM. At 24 h, strains expressing GFP-MPK10-Y192F and -T190A_Y192F showed a percentage of cell death of 51±7% and 48±6.1%, respectively (Figure 4B, lower panel right), but recovered thereafter and showed growth characteristics similar to the GFP-MPK10 NM used as control (Figure 4B, lower panel left). Parasites expressing GFP-MPK10-T190A showed a more severe phenotype with 61±7.6% of cell death (Figure 4B, lower panel right), from which they recovered slower than the other mutants (Figure 4B, lower panel left). This dominant negative effect, observed during the first 48 h after pH and temperature shift, corresponds to the period where MPK10 is the most active. Altogether, these data indicate that MPK10 could be important only transiently during differentiation. Moreover, these findings show that not only T190 phosphorylation is essential for the catalytic activity of MPK10 in promastigotes but also for axenic amastigote viability. Strikingly, increased cell survival can be restored in the GFP-MPK10-T190A mutant by Y192F mutation.
We then measured the kinase activity of the purified mutant proteins from parasites at 48 h of differentiation, when the percentage of cell death was the highest, and at 96 h after differentiation, when parasites have recovered. As shown in Figure 4C, we obtained similar results from amastigotes at 48 h (left panel) and 96 h (right panel) to those obtained from promastigotes, i.e. no kinase activity was detected for GFP-MPK10-K51A, -T190A, -Y192F or -T190A_Y192F. We did not find a clear correlation between over-expression phenotype and MPK10 kinase activity, as all mutant kinases were inactive, yet only the over-expression of GFP-MPK10-T190A was detrimental for the amastigotes, suggesting that the effect on growth is not entirely due to whether the protein is active or not. We only observed this phenotype in axenic amastigotes but never in promastigotes. This phenomenon may be due to the level of GFP-MPK10 expression, which was two- to five-fold lower in promastigotes compared to axenic amastigotes (data not shown). Thus, the level of transgenic MPK10 NM and MPK10 mutant protein could be too low to efficiently compete with the endogenous MPK10 in promastigotes, masking a dominant-negative effect. Why and how MPK10 levels are regulated in promastigotes and not in axenic amastigotes remains to be established. This observation seems independent from the vector, as expression of other kinases was not regulated the same way (data not shown). Altogether, these data confirm the essential role of both residues of the TxY motif for MPK10 activity.
Our observation that auto-phosphorylation activity of recombinant MPK10 is enhanced after removal of the 46 C-terminal amino acids primed us to investigate the role of this domain in the regulation of MPK10 NM in its physiological context using our transgenic system. Compared to GFP-MPK10 NM transgenic parasites, GFP-MPK10-ΔC over-expression had neither an effect on promastigote growth (Figure 5A, left panel) nor cell death (Figure 5A, right panel). We next measured the kinase activity of this mutant. As judged by quantification using imageJ, a 3.5 fold increase in phosphorylation of MBP by GFP-MPK10-ΔC compared to that obtained by GFP-MPK10 was observed (Figure 5B), demonstrating that transgenic GFP-MPK10-ΔC has a higher activity than GFP-MPK10. Strikingly, GFP-MPK10-ΔC showed a strong increase of MBP phosphorylation but a moderate increase of auto-phosphorylation, suggesting that part of the increase in activity could be due to a better affinity for the substrate rather than enhanced phosphotransferase activity. This finding supports the hypothesis that the C-terminal domain has an auto-inhibitory function in situ. Surprisingly, Y192 phosphorylation was reduced by 80% in GFP-MPK10-ΔC compared to GFP-MPK10 NM (Figure 5C). Thus, although only 20% of GFP-MPK10-ΔC showed Y192 phosphorylation and thus can be considered active, the truncated kinase still phosphorylated MBP 3.5 fold more efficiently than GFP-MPK10 NM, suggesting a dramatic increase in kinase catalytic function after removal of the C-terminal domain.
We investigated in the following the effect of GFP-MPK10-ΔC over-expression on axenic amastigotes. Contrary to promastigotes, the over-expression of GFP-MPK10-ΔC caused an important reduction of cell growth after temperature and pH shift (Figure 5D, left panel), which resulted from an increase in the percentage of parasite death during the differentiation process, reaching 78±8.5% at 48 h. In contrast to the over-expression of the other mutants (Figure 4B), GFP-MPK10-ΔC transgenic parasites did not recover and maintained a high level of cell death. This result indicates that GFP-MPK10-ΔC is toxic for axenic amastigotes and as a consequence we were not able to investigate the kinase activity of GFP-MPK10-ΔC at this parasite stage. However, based on the enhanced activity detected in promastigotes, we hypothesized that the toxicity could be the consequence of a non-physiologically high level of MPK10 kinase activity. To test this possibility we generated two double mutants, GFP-MPK10-ΔC_K51A and -ΔC_Y192F that lack activity by different means. Over-expression of GFP-MPK10-ΔC_Y192F was no longer toxic for axenic amastigotes, suggesting that rendering GFP-MPK10-ΔC inactive is sufficient to rescue the parasites from the lethal phenotype as the percentage of cell death was significantly reduced compared to the active kinase (Figure 5D, right panel). GFP-MPK10-ΔC_K51A also showed a reduction in toxicity but not as complete as GFP-MPK10-ΔC_Y192F. This discrepancy was due to a higher percentage of cell death from 48 h to 144 h, which was not observed with the over-expression of GFP-MPK10-ΔC_Y192F. We next investigated whether this difference between GFP-MPK10-ΔC_K51A and -ΔC_Y192F could be explained by a difference in kinase activity but, as shown in Figure 5B and expected from the results obtained with the mutants of full length MPK10, both mutants showed a weak or no activity towards MBP, respectively. Thus, the different phenotypes observed with these two mutants are not linked to a difference in kinase activity. We further showed that in contrast to hyperactive GFP-MPK10-ΔC, the level of Y192 phosphorylation of inactive GFP-MPK10-ΔC_K51A corresponds to wild-type level (Figure 5C), thus revealing a negative feedback loop between MPK10 activity that controls the level of Y192 phosphorylation. Overall, these findings demonstrate that the toxicity following over-expression of GFP-MPK10-ΔC in amastigotes is due to the high activity of the truncated kinase, which can be compensated by reduction of tyrosine phosphorylation.
After demonstrating the importance of the C-terminal domain for auto-inhibition of MPK10 activity and parasite viability, we investigated the potential mechanisms involved in its regulation. Several mechanisms have been described to release kinases from auto-inhibition, including protein phosphorylation [45]. We performed a qualitative phospho-peptide analysis using L. donovani axenic amastigote extracts (48 h of axenic differentiation) to test this possibility. Our analysis identified the expected phospho-peptides encompassing the TxY motif (supplementary Figure S8A), but revealed a novel phospho-peptide located in the C-terminal domain of MPK10 showing a single phosphorylation at residue S395 (supplementary Figure S8B). This serine is conserved across MPK10 orthologs in trypanosomatids (Figure 6A, gray arrow), and part of the conserved sequence motif DHMxRTxSxME (underlined, Figure S3A), which suggests an important role for MPK10 function. If this serine were implicated in the regulation of MPK10 auto-inhibition, we would expect its phosphorylation to be stage regulated. To test this hypothesis we took advantage of the previously described quantitative phospho-peptide analysis. SRM analysis confirmed the presence of phosphorylated S395, which was mostly identified in promastigotes. Besides T190 of the activation loop this is the second stage-specifically regulated residue in MPK10 and thus may be crucial to control kinase activity (Figure 6B). As this promastigote-specific phosphorylation occurs when MPK10 is least active, dephosphorylation of S395 could be important to release MPK10 from the auto-inhibition.
We next investigated this possibility and studied the role of this residue in MPK10 regulation by generating transgenic parasites over-expressing GFP-MPK10-S395A. No difference was observed between GFP-MPK10 NM and -S395A with respect to promastigote growth (Figure 6C, left panel) or percentage of cell death (Figure 6C, right panel). Comparison of the kinase activities of GFP-MPK10 NM and -S395A purified from promastigotes did not reveal any difference based on quantification of the signals by a scintillation counter as detailed in Figure 3A (Figure 6D). We next compared the kinase activity of GFP-MPK10 NM and -S395A purified from parasites at different time points during axenic amastigote differentiation (Figure 6D). At 48 h, GFP-MPK10 NM phosphorylated MBP more efficiently than GFP-MPK10-S395A (100% and 54% respectively), whereas the activity of the mutant protein was slightly stronger than that of GFP-MPK10 NM at 96 h (127% and 100% respectively). These findings provide evidence for a delay in GFP-MPK10-S395A activation, supporting a role of this residue in proper kinase regulation.
The mean values with standard deviation of three independent experiments assessing activity for GFP-MPK10 NM and -S395A of parasites at 48 h and 96 h during axenic amastigote differentiation is shown in Figure 6E. Statistically significant differences were observed at 48 h with a reduction by twofold of GFP-MPK10-S395A activity (p-value<0.001), and at 96 h where on the contrary GFP-MPK10-S395A activity is increased by 75% (p-value<0.05). No significant difference was observed in promastigotes (p-value of 0.06). These data strongly suggest that in axenic amastigote, GFP-MPK10-S395A is differentially regulated compared to GFP-MPK10 NM, which could cause the observed effect on parasite viability. Parasites expressing GFP-MPK10-S395A presented a delay of 24 h in cell growth compared to GFP-MPK10 NM but thereafter resumed growth with kinetics slightly slower than that of parasites overexpressing GFP-MPK10 NM (Figure 6F, right panel). This delay is caused by a significantly higher percentage of cell death (69±16%, p-value<0.01) observed for parasites expressing GFP-MPK10-S395A at 48 h of axenic amastigote differentiation compared to GFP-MPK10 NM (44±12%, Figure 6F, right panel). In conclusion, our findings demonstrate the importance of S395 residue for axenic amastigote survival. The fact that the GFP-MPK10-S395A phenotype resembles that obtained with the over-expression of GFP-MPK10-T190A attributes similar importance for kinase regulation to S395 in the C-terminal domain as to T190 in the activation loop.
While the biological processes underlying Leishmania stage differentiation are poorly understood, the regulation of parasite development by environmental cues is firmly established. In eukaryotes, the MAPK pathway transmits environmental signals to trigger a wide range of cellular response. Thus studying the Leishmania MAPK pathway can provide new insight into molecular mechanisms underlying parasite differentiation and parasite-specific kinase biology. Here we uncover novel mechanisms regulating the Leishmania MAP kinase homolog MPK10 by utilizing two complementary approaches to study MPK10 regulation. First, recombinant expression and purification of MPK10 enabled us to identify a potential auto-inhibitory domain at the C-terminus of the protein. Second, transgenic expression and purification of MPK10 kinase allowed us (i) to validate MPK10 auto-inhibition in situ, (ii) to demonstrate the importance of the C-terminal domain for axenic amastigote survival, and (iii) to identify T190 and S395 as key regulatory residues. We propose that auto-inhibition of active MPK10 provides a fast signaling switch likely involved in environmentally induced pro- to axenic amastigote conversion and stage-specific regulation of MPK10 activity.
MAPKs are proline-directed serine/threonine kinases that phosphorylate substrates containing proline in the P+1 site [46]. Classically, MAPKs are inactive enzymes that are solely activated by M2Ks, which phosphorylate both the threonine and the tyrosine of the TxY motif present in the activation loop [9], [46]. This phosphorylation allows conformational changes that lead to the alignment of the R and C spines required for the activation of the kinase [29], [35], [36]. Moreover, the phosphorylation of the tyrosine residue of the TxY motif is important to permit the formation of the proline-directed P+1 specificity site required for substrate recognition and restriction of specificity [46]. The threonine residue possesses a structural role by stabilizing MAPK conformation and improving the geometry of the active site [46], [47]. MPK10 retains certain characteristics of classical MAPKs such as the conserved motif typical of MAPKs, TxYxxxRxYRxPE, including the TxY motif and the (P+1)-specificity pocket [16]. We demonstrated the importance of the T190 residue for the catalytic activity of MPK10, as alteration of this site abrogates phosphotransfer and severely reduces axenic amastigote survival. We also showed the essential role of the Y192 residue for MPK10 activity although its mutation does not have an impact on axenic amastigote survival unlike mutation of T190.
Aside these conserved MAPK features, MPK10 regulation presents many non-classical characteristics. We have previously shown that based on its structural conformation, MPK10 appears to be in an active conformation, without the need for dual-phosphorylation of the TxY motif, as the replacement of the DFG motif by a DFN motif results in the alignment of the R and C spines, similar to eukaryotic MAPKs after phosphorylation of the TxY motif by M2Ks [29]. Our study supports these findings revealing alternative modes of regulation of this intrinsic MPK10 activation state. We first showed that phosphorylation of Y192 is largely dissociated from MPK10 activity as its phosphorylation state does not show any significant stage-specific change, even though the activity of MPK10 increases by about twofold between log phase promastigotes and axenic amastigotes at 48 h after initiation of differentiation. In classical MAPKs, kinase activity correlates with phosphorylation of both the threonine and the tyrosine residues of the TxY motif, which is tightly regulated by environmentally induced upstream M3Ks and M2Ks [9]. By contrast, our data suggest that Leishmania MPK10 is mostly constitutively phosphorylated on Y192 in promastigotes and during amastigote development and proliferation. Consequently, this residue seems not to be implicated in regulating the observed stage-specific activation of MPK10. These findings were confirmed by proteomics identification of a mono-phosphorylated T190-H-pY192 activation loop in both pro- and axenic amastigotes, whereas dual pT190-H-pY192 phosphorylation of the activation loop was mainly identified in axenic amastigotes, where MPK10 is most active. However, we observed a decrease in Y192 phosphorylation in axenic amastigotes at stationary phase, which correlated with a decrease in MPK10 activity, suggesting that inactivation of MPK10 at this stage requires Y192 dephosphorylation.
Altogether, these findings indicate that the two residues of the TxY motif are differentially regulated, raising the question on how MPK10 is phosphorylated by M2Ks. There are at least five possibilities: (i) Y192 could be auto-phosphorylated in cis, a possibility that we rule out since kinase dead MPK10-K51A still shows phosphorylation on this residue. (ii) Classically, the phosphorylation of the TxY motif by M2Ks is sequential, initiating with the tyrosine residue [48]. M2Ks could constitutively phosphorylate the tyrosine residue of MPK10, but require additional signals or interactions to complete phosphorylation of the adjacent threonine residue. (iii) Each regulatory residue could be phosphorylated independently by two different M2Ks as it was demonstrated for human JNK kinase, whose threonine of the TxY motif is preferentially phosphorylated by MKK7, whereas its tyrosine is preferentially phosphorylated by MKK4 [9]. (iv) Without activation of the MAPK cascade, only Y192 could be accessible for phosphorylation, whereas access to T190 could be blocked by the C-terminal domain. (v) Both sites could be phosphorylated by the same M2K but only T190 would be constantly dephosphorylated until the MAPK pathway is fully activated. Three types of phosphatases target MAPKs, the dual-specificity phosphatases also called the MAP kinase phosphatases (MKPs), the tyrosine phosphatases such as PTP-SL and STEP, and finally the serine/threonine phosphatase such as PP2A [49]. PP2A could likely be responsible for this dephosphorylation; a possibility supported by recent findings that T. cruzi PP2A blocks axenic amastigote differentiation [50].
Dissociation of Y192 phosphorylation from stage-specific induction of MPK10 activation during amastigote differentiation raises the question of its importance for MPK10 function. We showed that similar to mammalian MAPKs, the Y192 residue is essential for MPK10 enzymatic activity. However, the dominant negative effect generated by overexpression of MPK10-T190A on axenic amastigote growth and viability was more detrimental than the one observed by overexpression of MPK10-Y192F, suggesting a more important role for the T190 residue for MPK10 activity. We can only speculate on the role of the largely constitutive phosphorylation of the Y192 residue. As described in the literature, phosphorylation of the tyrosine of the TxY motif is important to form the proline-directed (P+1)-specificity pocket [46], [49]. Consequently, maintaining Y192 mostly phosphorylated could allow a constant and dynamic interaction of MPK10 with its substrates allowing for faster phosphorylation once the pathway is activated.
A second striking feature of the regulation of MPK10 is represented by its C-terminal domain. Whether we used the recombinant or the transgenic protein, removal of this domain resulted in a significant increase in kinase activity. While this increase translates mainly into higher auto-phosphorylation levels for the recombinant kinase, we observed a dramatic increase in substrate phosphorylation for the transgenic kinase. One main difference between the two kinases is the absence of Y192 phosphorylation of the recombinant protein, which could explain the inability of recombinant kinase to efficiently transfer phosphate onto the substrate (Figure S5). Thus, our findings firmly establish regulation of MPK10 through auto-inhibition. This is a common feature in the regulation of kinase activities, and has been described for example for Twitchin kinase or CaMKI (Ca2+/calmodulin-dependent kinase I) but is a rare feature for MAPKs [48]. Although five human MAPKs have a long C-terminal domain (ERK3, ERK4, ERK5, ERK7 and ERK8), only that of ERK5 has an auto-inhibitory function [9]. The authors have shown that deletion of the last 100 amino acids of ERK5 leads to an increase of its kinase activity and they postulated that the deletion of the C-terminal domain could facilitate its activation by the upstream kinase [51]. Evidence for a different auto-regulatory mechanism of MPK10 arises from our analysis using transgenic parasites expressing the hyper-active MPK10 C-terminal deletion mutant: the strongly reduced phosphorylation levels of the regulatory Y residue in this mutant despite its significantly increased phosphotransferase activity demonstrates that the pool of active MPK10 is reduced, while the activity of each single active kinase protein is dramatically increased. Thus the deletion of the C-terminal domain increases the basal activity of the truncated kinase rather than allowing better access for the M2K. How the C-terminal domain regulates MPK10 activity is still elusive since no conserved pattern or domains such as SH2 or SH3 have been identified. One possible explanation could be that the C-terminal domain masks the active site acting as a pseudosubstrate such as described for Twitchin kinase [52]. It is interesting to note that in T. brucei, a hybrid kinase between a CDK (Cyclin dependent kinase) and a MAPK termed TbECK1 is also auto-inhibited by its C-terminal domain [32]. Procyclic parasites expressing the truncated protein showed growth defects and the parasites presented aberrant karyotypes. The expression of truncated TbECK1 was toxic to the bloodstream form, which is reminiscent to the toxic effect of MPK10-ΔC in axenic amastigotes.
Our data show that activating MPK10 via T190 is important for axenic amastigote survival, especially during the first 48 h after induction of axenic amastigote differentiation by temperature and pH shift. However, inhibition of MPK10 activity is equally essential as absence of auto-inhibition caused massive cell death in axenic amastigotes expressing MPK10-ΔC. These results therefore demonstrate the transient requirement for MPK10 during axenic amastigote differentiation and reveal a new mechanism to regulate MAPK activity by auto-inhibition, which is crucial for amastigote viability.
Finally, we identified a highly conserved, trypanosomatid-specific serine at position 395 inside the MPK10 C-terminal domain as a novel, stage-specific regulatory residue that is mainly phosphorylated in promastigotes and whose mutation causes an important dominant negative effect on axenic amastigote survival. Because kinases can be released from auto-inhibition by phosphorylation or dephosphorylation [45], we hypothesized that S395 could be implicated in the regulation of the auto-inhibition of MPK10 by its C-terminal domain. We did not confirm this hypothesis but we have two pieces of evidence suggesting that this residue could have an important role in the regulation of MPK10 activity. First, we showed that S395 is conserved in all trypanosomatids and is part of a sequence motif inside the C-terminal domain, which is also conserved in all trypanosomatids. Second, we clearly established the importance of the phosphorylation of S395 for axenic amastigote survival, which is as important as the regulation of T190, and partially mimics the viability defect caused by MPK10-ΔC. The function of this residue is still elusive and may be studied in the future by using an anti-phospho-S395 specific antibody to investigate the link between the phosphorylation kinetics of this residue and MPK10 activity, and address the question whether the release from auto-inhibition is regulated by environmental signals or whether it is alleviated before signal sensing, leaving the kinase in a semi-activated state, relying only on T190 phosphorylation for activation.
Because MPK10 shows a transient peak of activity during the first 48 h of axenic amastigote differentiation, its activity seems to be tightly regulated. The decrease in MPK10 activity after 48 h of induction of axenic amastigote differentiation is concomitant with the decrease of Y192 phosphorylation, suggesting that dephosphorylation of Y192 and probably T190 are required for MPK10 inactivation. We have evidence suggesting that MPK10 could directly or indirectly regulate its activity through the phosphorylation or the dephosphorylation of its TxY motif. Indeed, we have shown that to reduce toxicity caused by the over-expression of the hyperactive MPK10-ΔC, its level of Y192 phosphorylation was decreased. The level of pY192 was restored to wild-type level only after rendering the truncated kinase inactive, suggesting the existence of a feedback loop, where MPK10 could either inactivate the M2K that phosphorylates Y192 or activate the phosphatase that dephosphorylates this residue. This kind of feedback regulation between MAPKs and M2Ks or MKPs has not been extensively studied, as only three types of feedback loop have been described that are linked to regulation of MKPs, involving (i) rescue of MKP from degradation through phosphorylation by MAPK [53], (ii) activation of MKP by MAPK binding [54], and (iii) induction of MKP transcription by activated MAPK [55], [56], a possibility that can likely be discarded as transcription is constitutive in Leishmania [57]. In addition, Mody et al. have reported the phosphorylation of MKK5 by ERK5, revealing the existence of a potential feedback loop between a MAPK and its upstream activating kinase [58]. Future identification of the M2K and MKP that regulate the phosphorylation dynamics of the TxY motif of MPK10 will provide more insight into the mechanism controlling this feedback loop.
In conclusion, our transgenic study identifies novel mechanisms of MPK10 regulation that are unusual for MAPKs and document once more the stunning capacity of Leishmania to adapt highly conserved signaling proteins to its parasitic life style. Our data propose a model in which MPK10 is not inactivated but partially active in promastigotes as judged by tyrosine phosphorylation and structural conformation (Figure 7). At this stage the kinase is kept in a standby configuration by auto-inhibition. During the first 48 h of axenic amastigote differentiation, MPK10 is released from auto-inhibition, which correlates with T190 phosphorylation and S395 dephosphorylation. This activity seems to be controlled by a feedback loop where MPK10 regulates its own tyrosine phosphorylation levels. Thereafter, MPK10 activity is decreased likely due to dephosphorylation of the TxY motif and phosphorylation of S395. These regulatory residues may fine tune MPK10 regulation according to environmental signals and differentiation state through activating and inhibitory mechanisms. Future studies combining null mutant analysis and complementation assays for a detailed structure-function analysis of these residues and the auto-inhibitory C-terminal domain will uncover the contribution of these sequence elements in regulating MPK10 functions relevant for parasite differentiation and infectivity.
Leishmania donovani strain 1S2D (MHOM/SD/62/1S-CL2D), clone LdB was cultured and axenic amastigotes were differentiated as previously described [4] [59] [60]. Briefly, 106 logarithmic promastigotes per mL were grown at 26°C in M199 media (supplemented with 10% heat-inactivated FCS, 25 mM HEPES pH 6.9, 4.2 mM NaHCO3 7.5%, 2 mM glutamine, 8 µM 6-biopterin, 1× RPMI 1640 vitamin mix, 10 µg/mL folic acid, 100 µM adenine, 30 µM hemin, 100 U/mL of Penicillin/Streptomycin (Pen/Step)) and differentiated in axenic amastigotes by incubation at 37°C and 5% CO2 with RPMI media (supplemented with 20% of heat-inactivated FCS, 28 mM MES, 2 mM glutamine, 1× RPMI 1640 amino acid mix, 10 µM folic acid, 100 µM adenine, 100 U/mL of Pen/Step). Parasites were harvested at different time points between 12 h to 144 h after induction of axenic amastigote differentiation. For kinetic analysis, one flask (Corning) containing 200 mL of culture medium was used for each time point to avoid internal variations.
Leishmania mexicana MNYC/BZ/62/M379 promastigotes and axenic amastigotes were grown as described previously [61]. Cultures were incubated at 27°C until late-log phase (4–5×107 parasites/ml), and either harvested or differentiated into amastigotes by inoculation in Schneider's Drosophila medium (PAN Biotech, Aidenbach, Germany) supplemented with 20% heat-inactivated FCS (PAN Biotech), 2 mM L-glutamine, 100 U/ml Pen/Strep, and 20 mM 2-morpholinoethanesulfonic acid monohydrate [MES] (Serva, Heidelberg, Germany) for a final pH of 5.5. Cultures were incubated at 34°C, 5% CO2, for 72 h. Parasites were harvested by centrifugation at 2,000× g at 4°C, and washed consecutively in ice-cold HEPES and ice-cold HEPES with protease and phosphatase inhibitors (1 mM Na-orthovanadate, 0.1 µM okadaic acid, 10 mM NaF, 10 mM o-phenanthroline, EDTA-free protease inhibitors (Roche)). Parasite pellets were snap-frozen in liquid nitrogen and stored at −80°C.
Episomal tranfectants were generated by electroporation of 5×107 L. donovani LdB promastigotes from logarithmic culture with 20 µg of plasmid [62]. Transfected cells were selected in liquid media containing 20 µg/ml geneticin (Invitrogen) and resistant parasites were expanded in liquid culture at drug concentrations up of to 100 µg/ml of geneticin. Parasites were then frozen one passage after selection and all experiments were performed with parasites issued from the same electroporation. To avoid any potential bias due to adaptation or compensatory responses, parasites were used for all the experiments at passage 2 after selection. Pools of transfectants were used to avoid clonal variation that can bias transgenic studies.
Cultured parasites were incubated for 15 min with 10 µg/ml propidium iodide (Sigma-Aldrich) and diluted in PBS (Gibco). Cells were analyzed with a FACSCalibur flow cytometer (Beckman Coulter) to determine the incorporation of propidium iodide (excitation wave length λex = 488 nm; emission wave length λem = 617 nm). The percentages of cell death and cell growth were calculated using FlowJo (v7.6) software (Tree Star, Inc., San Carlos, CA).
109 parasites were washed with ice cold RPMI and lysed in 1 ml lysis buffer containing 150 mM NaCl, 1% Triton X-100, 50 mM Tris HCl pH 8 and inhibitor cocktails for proteases (Complete Mini EDTA-free tablets, Roche Applied Science, IN) and phosphatases (Phosphatase Inhibitor Cocktails I and II, Sigma–Aldrich, MO). Clear lysates were obtained after sonication and centrifugation at 12 000× g for 10 min and stored at −80°C.
Purified GFP-MPK10 wild-type and mutant proteins were isolated from crude cell extracts of respective transgenic parasites using the μMACS Epitope Tag Protein Isolation Kit, according to the manufacturer's specifications (Miltenyi Biotec Inc., CA). Briefly, equal amounts of total proteins were incubated with 50 µl of magnetic bead-conjugated mouse monoclonal anti-GFP antibody for 1 hour at 4°C, immuno-complexes were immobilized on the μMACS separator, washed four times with 150 mM NaCl, 1% Igepal CA-630 (formerly NP-40), 0.5% sodium deoxycholate, 0.1% SDS, 50 mM Tris HCl (pH 8.0), and once with 20 mM Tris HCl (pH 7.5). Bound GFP-MPK10 protein was eluted in 75 µl PBS after removing the columns from the magnetic field.
Purified proteins were separated by SDS–PAGE (NuPAGE gel 4–12% Bis-Tris, Invitrogen) and visualized either by Coomassie staining or SYPRO Ruby Protein Gel Stain (Invitrogen) using a Typhoon 9400 scanner (Amersham Biosciences) with λex = 457 nm and λem = 610 nm. Alternatively, proteins were separated by SDS–PAGE on NuPAGE 4–12% Bis-Tris gels (Invitrogen) and blotted onto polyvinylidene difluoride (PVDF) membranes (Pierce). Proteins were revealed using the following antibodies at the indicated dilutions: i) polyclonal anti-MPK10 antibody, generated by rabbit immunization using recombinant MPK10 protein produced in E. coli transformed with pGEX-Strep3-MPK10 plasmid (Eurogentec), 1∶10,000; ii) anti-phospho-tyrosine antibody 4G10 Platinum from Millipore, 1∶1,000; and iii) secondary goat anti-rabbit-HRP and anti-mouse-HRP antibodies from Thermo Scientific, 1∶20,000. The visualization was performed on X-ray film (Roche) at various exposure times.
E. coli BL21 Rosetta (VWR) transformed with pQE80-His6-MPK10 or pGEX-GST-Strep3-MPK10 were grown at 37°C and induced with IPTG (0.2 µM final) overnight at RT. Cells were harvested by centrifugation at 15,000× g for 10 min at 4°C.
For the purification of GST-Strep3-MPK10, bacteria were resuspended in a pre-chilled buffer containing 25 mM Tris-HCl pH 8, 150 mM NaCl, 1 mM DTT, 10 µg/mL apoprotein, 1 µM leupeptin, 1 µM pepstatin, 1 mM PMSF. Samples were sonicated (Bioruptor system, Diagenode) for 2 min at 20 V setting on ice (10 s on/10 s off cycle). After addition of 500 µg/mL of lysozyme and 500 U of benzonase, lysates were incubated on ice for 30 min with 140 µM EDTA, 0.0035% Triton-X100 and centrifuged at 15,000× g for 30 min at 4°C. The supernatant was immediately subjected to GST affinity chromatography (GSTrap, GE Healthcare Life Sciences, Waukesha, WI, USA). GST-tagged proteins were washed with a buffer containing 50 mM Tris pH 8, 100 mM NaCl, 1 mM DTT and eluted with the same buffer supplemented with 30 mM L-glutathione. Appropriate fractions were pooled and the GST-tag was cleaved by incubation of the fractions with 5 µg/mL Xa factor in presence of 1 mM CaCl2. The reaction was stopped by adding 5 µg/mL of Glutamyl-glycyl-arginine chloromethyl ketone GGACK (Calbiochem). A second purification step was performed using a StrepTrap column (Strep-tactin, GE Healthcare Life Sciences, Waukesha, WI, USA). Elution was performed with E Strep buffer (100 mM Tris-HCl pH 8, 150 mM NaCl, 1 mM EDTA and 2.5 mM Desthiobiotin) and appropriate fractions were collected, and stored at 4°C until used.
For His6-MPK10 purification, the bacterial pellet was resuspended in PBS containing 60 mM β-glycerophosphate, 1 mM sodium vanadate, 1 mM sodium fluoride, 1 mM disodium phenylphosphate, 150 mM sodium chloride, 10 mM imidazole supplemented with protease inhibitor cocktail (Complete EDTA free tablets, Roche Applied Science). The sample was sonicated for 2 min at 20 V setting on ice (10 s on/10 s off cycle). Triton X-100 (0.1% final) was added, the sample was incubated for 30 min at 4°C (shaking) and centrifuged at 15,000× g for 30 min at 4°C. The supernatant was purified on Co-NTA agarose (Pierce). The beads were washed with PBS containing 60 mM β-glycerophosphate, 1 mM sodium vanadate, 1 mM sodium fluoride, 1 mM disodium phenylphosphate, 300 mM sodium chloride, 30 mM imidazole, 1% Triton X-100 at pH 7.5. Elution was performed with 300 mM imidazole in elution buffer pH 7.5 (PBS containing 60 mM β-glycerophosphate, 1 mM sodium vanadate, 1 mM sodium fluoride, 1 mM disodium phenylphosphate). The eluate was supplemented with 15% glycerol and stored at −80°C.
Ten percent of the GFP-MPK10 purified protein was incubated on a shaker for 30 min at 37°C with 25 µg myelin basic protein (MBP) substrate, 200 µM of ATP, 50 mM of MOPS pH 7.5, 100 mM NaCl, 10 mM MgCl2 and 1 µCi [γ-32P] adenosine-triphosphate (ATP) (3000 Ci/mmol) in final volume of 20 µl. The phosphotransferase reaction was then stopped by adding Laemmli loading buffer. Reaction mixtures were separated by SDS–PAGE, which was stained by Commassie and dried. 32P incorporation was monitored by exposing the dried gel on an X-ray sensitive film (Roche) at −80°C. After exposure, the bands corresponding to MPK10 or MBP were excised from dried gels and radioactivity was quantified by a scintillation counter.
Recombinant His6-MPK10 and respective mutants were assayed with 36 µg dephosphorylated casein, 12 µg histone H1, 9 µg of Ets1 or 9 µg MBP as substrates in a Tris buffer at pH 7.5 (50 mM Tris-Cl pH 7.5, 10 mM MnCl2 and 100 mM NaCl) in 20 µl final volume and in the presence of 15 µM [γ-33P]-ATP. After 30 min incubation at 37°C, the reaction was stopped by adding an equal volume of 2× electrophoresis loading buffer to the 20 µl reaction mix. Incorporated 33P was monitored by auto-radiography.
50 µg of Strep3-MPK10 were digested with 0.25 µg trypsin at RT. Aliquots were taken at 0, 2.5, 5, 15, 30, 60, and 150 min and the reaction was stopped by adding Laemmli loading buffer. The polypeptides were then separated by SDS-PAGE, transferred to PVDF membrane, stained with amidoblack, and N-terminal sequencing was performed. For the mass determination of cleavage products, pH of the cleavage reaction was lowered to 5.0 and mass determination was performed by SELDI-TOF analysis after immobilizing the samples on a H4 ProteinChip Array (C16 reversed phase surface).
A Leishmania amastigote cell pellet was submitted for qualitative phosphoproteomic analysis using titanium dioxide phosphopeptide enrichment followed by an iTRAQ labeling experiment for analysis of phosphopeptides by LC-MS/MS. Briefly, 400 µg of Leishmania proteins were reduced with DTT and the free cysteines were alkylated with iodoacetamide for 30 min at 37°C in darkness. Proteins were then digested overnight at 37°C using Porcine trypsin (sample:enzyme ratio of 50∶1). Following the digestion, the peptides were acidified, concentrated and de-salted using a Waters HLB Oasis SPE cartridge. The peptides were then enriched for phosphopepitdes using a TiO2 affinity column and splited into two 100 µL aliquots. Each aliquot was then reduced and alkylated with MMTS followed by iTRAQ labeling (114, 116). The 116 labeled sample was then treated with FastAPalkaline phosphatase before the samples were combined, fractionated by SCX chromatography and analyzed by ESI-Q-TOF MS/MS. Sample were then analyzed by reversed phase nanoflow (300 nL/min) HPLC with nano-electrospray ionization using a quadrupole time-of-flight mass spectrometer (QSTAR Pulsar I, Applied Biosystems) operated in positive ion mode and a 2 hour gradient.
1×107 parasites were lysed in a solution of 7 M urea, 2 M thiourea, 40 mM Tris, 1% n-octyl-β-D-glycopyranoside, 1 mM MgCl2, 1 mM o-phenanthroline, 300 U benzonase, 1 mM Na-pervanadate (Na-orthovanadate activated in 18% H2O2), EDTA-free protease inhibitors (Roche) and phosphatase inhibitor cocktails (P2850 and P5726 from Sigma), and sonicated for 3×15 s on ice. Lysates were incubated at −80°C for 30 min prior to reduction (DTT, 20 mM, incubation at room temperature for 60 min) and alkylation (iodoacetamide, 40 mM, incubation at room temperature in the dark for 45 min). Proteins were precipitated in 8 fold excess of ice-cold acetone-ethanol (1∶1, v/v) by overnight incubation at −20°C. Proteins were reconstituted in 6 M urea/2 M thiourea and diluted in 50 mM NH4HCO3 for digestion with trypsin at a 75∶1 substrate-enzyme ratio over night. For selected reaction monitoring (SRM) analyses, 3×330 µg digested protein from whole cell lysates of promastigotes, axenic amastigotes were subjected to TiO2 enrichment as described in Rosenqvist et al. [40]. The TiO2 eluates were pooled prior to SRM analysis. The discovery analyses were conducted in triplicates for each of the sample types using LTQ Orbitrap XL mass spectrometers (Thermo Fisher Scientific, Bremen, Germany). The discovery data were processed with DTASuperCharge (Mortensen, P. DTASuperCharge, an MSQuant application. http://msquant.alwaysdata.net/msq/) and searched against a customized L. mexicana 6-frame translation library as well as a predicted protein list (GeneDB http://www.genedb.org) using an in-house Mascot server (version 2.2.06, Matrix Science, London, UK). For SRM analyses, the discovery data were processed in ProteomeDiscoverer, and the MPK10 phosphopeptides identified were imported into Pinpoint, version 1.0.0 (Thermo Scientific).
Leishmania spp and Trypanosoma spp MPK10 gene and protein sequences were retrieved from the web databases GeneDB (www.genedb.org) and TriTrypDB (http://tritrypdb.org/tritrypdb/) [63]. Homology searches were carried out using BLAST Program with the default BLOSUM-62 substitution matrix [64], and pattern recognition analysis using the program PRATT v2.1 [65]. Multiple sequence alignments were performed using built-in algorithm ClustalXv2. Additional sequence analyses were carried out using the BioEdit Program suite (Tom Hall, North Carolina State University). Statistical analysis and data plotting were performed using Rstudio software (http://www.rstudio.org/) and R language (R Development Core Team (2005). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL: http://www.R-project.org). Statistical analyses were performed using the t-test on mean values when samples followed a Normal distribution, otherwise the Mann-Whitney Rank Sum Test was used. Differences were considered significant when p value<0.05.
|
10.1371/journal.pcbi.1000346 | The Role of Environmental Transmission in Recurrent Avian Influenza
Epidemics | Avian influenza virus (AIV) persists in North American wild waterfowl, exhibiting
major outbreaks every 2–4 years. Attempts to explain the patterns of
periodicity and persistence using simple direct transmission models are
unsuccessful. Motivated by empirical evidence, we examine the contribution of an
overlooked AIV transmission mode: environmental transmission. It is known that
infectious birds shed large concentrations of virions in the environment, where
virions may persist for a long time. We thus propose that, in addition to direct
fecal/oral transmission, birds may become infected by ingesting virions that
have long persisted in the environment. We design a new host–pathogen
model that combines within-season transmission dynamics, between-season
migration and reproduction, and environmental variation. Analysis of the model
yields three major results. First, environmental transmission provides a
persistence mechanism within small communities where epidemics cannot be
sustained by direct transmission only (i.e., communities smaller than the
critical community size). Second, environmental
transmission offers a parsimonious explanation of the 2–4 year
periodicity of avian influenza epidemics. Third, very low levels of
environmental transmission (i.e., few cases per year) are sufficient for avian
influenza to persist in populations where it would otherwise vanish.
| Avian influenza viruses (AIVs) in wild waterfowl constitute the historic source
of human influenza viruses, having a rich pool of genetic and antigenic
diversity that often leads to cross-species transmission. Although the emergence
of H5N1 avian influenza virus onto the international scene has captured the most
attention, we do not as yet understand the mechanisms that underpin AIV
persistence and dynamics in the wild. We developed a novel
host–pathogen model intended to describe the epidemiology of low
pathogenic AIV in temperate environments. Our model takes into account
seasonality in migration and breeding together with multiple modes of
transmission. AIVs have been detected in unconcentrated lake water, soil swabs,
and mud samples. Laboratory experiments show that AIVs persist and remain
infectious in water for extended periods. However, so far, the possibility of
environmental transmission of AIV has been largely overlooked. Our work shows
that environmental transmission provides a parsimonious explanation for the
patterns of persistence and outbreaks of AIV documented in the literature. In
addition to their scientific importance, our conclusions impact the design of
control policies for avian influenza by emphasizing the dramatic and long-term
role that environmental persistence of pathogens may play at the epidemic
level.
| Many important infectious diseases persist on a knife-edge: rapid rates of
transmission coupled with brief infectious periods generate boom-and-bust epidemics
that court extinction. Such violent epidemic behavior has been observed in measles
[1]–[4], plague [5],
cholera [6],
meningitis [7],[8], and pertussis [9], among
others. Several distinct mechanisms have been proposed to explain the long-term
dynamics and persistence of these pathogens. For example, measles persistence is
primarily determined by the rate at which the susceptible pool is replenished,
leading to Bartlett's concept of critical community size,
the minimum population size above which an infectious disease remains endemic [4]. In
contrast, plague is enzootic in rodents and their fleas and thus its persistence in
human populations is explained by intermittent reintroduction from the animal
reservoir [10]. King et. al [6] argue that rapid loss of
immunity to cholera may replenish the human susceptible pool so quickly that large
amplitude cholera outbreaks can be observed semiannually. Finally, rich strain
polymorphism allows echoviruses –responsible for aseptic
meningitis– to circumvent host immunity and thus reinvade the population
[7],[8]. These examples
illustrate the need for understanding alternate persistence/re-invasion mechanisms
of infectious diseases for effective management and control.
In this paper, we investigate the persistence and dynamics of low pathogenic avian
influenza virus (AIV) in North American bird populations. Avian influenza viruses in
wild waterfowl constitute the historic source of human influenza viruses [11], with
a rich pool of genetic and antigenic diversity [11],[12] that often leads to
cross-species transmission. Perhaps the best-known and most topical example is the
transmission of H5N1 avian influenza virus to humans [13]. Human infection with
H5N1 is associated with a significant risk of mortality; to date, approximately
50% of infected individuals have died from the infection (see [13] and
references therein). Developing a better understanding of the ecology of avian
influenza viruses is, therefore, very timely.
AIVs infect more than 90 species of birds from 13 orders, mostly Anseriformes (ducks)
and Charadriiformes (shorebirds). Long-term studies of AIV prevalence in North
America [14],[15] have gathered time series of annual estimates
that extend over 26 years for Anseriformes and 20 years for Charadriiformes. The
data is stratified over influenza subtype: H3, H4, and H6 were the most prevalent
subtypes isolated from Anseriformes. Most interestingly, the prevalence of infection
with these subtypes as well as the aggregate prevalence exhibit recurrent outbreaks
in duck populations at 2–4 year intervals.
It is well established that birds infected with avian influenza are infectious for
approximately a week (range 6–10 days), during which they continuously
shed vast concentrations of viral particles in their feces [11],[16],[17]. These
virions are then ingested by susceptible birds, completing the fecal/oral
transmission route [18],[19]. However, attempts to recover the patterns of
periodicity and persistence in avian influenza epidemics in waterfowl from simple
modeling principles using only this essentially direct transmission mechanism are
unsuccessful (see, for example, Text S1 and Discussion). We propose that the
missing ingredient in direct transmission models is the additional indirect
contribution made to transmission by the ingestion of infectious virions that
persist in the environment. It has been demonstrated, for example, that the avian
influenza strain H2N4 (A/Blue-winged teal/TX/421717/01) can persist for extended
periods in the environment, with an estimated one log decay time of 490 days in
water at temperature 4°C, pH 7.2, salinity 0 ppt [16],[20].
Additionally, these persistent virions are known to be infectious [16],[20],[21], arguing
for a potentially significant epidemiological contribution by environmental
transmission.
Here we examine whether environmental transmission provides a more parsimomious
explanation for the observed patterns of avian influenza epidemics. The phenomenon
of environmental transmission is known to be significant for viral infections in
humans (e.g., gastroenteritis [22]) and animals (e.g., rabbit haemorrhagic disease
[23]), water-borne pathogens (e.g., cholera [6],[24] and
avian cholera [25]), some bacterial infections (e.g., tetanus [26],
salmonella [27] and epizootics of plague [28]), prion diseases (e.g.,
chronic wasting disease [29] and bovine spongiform encephalopathy [30]) and
zoonoses (e.g., Nipah and Hendra viral diseases [31]). Despite these
examples, the epidemiological consequences of environmental transmission remain
poorly understood [32]–[38].
Here we propose a new host-pathogen model that combines within-season transmission
dynamics, with a between-season component that describes seasonal migration,
reproduction and environmental variation. Analysis of deterministic and stochastic
versions of this model shows that environmental transmission plays a critical role
for the persistence of avian influenza and its inter-annual epidemics. We conclude
that environmental transmission may provide a parsimonious explanation of the
observed epidemic patterns of avian influenza in wild waterfowl.
Our model is designed to represent a typical population
(∼5,000–10,000 individuals) of ducks (Anseriformes) that migrates
twice a year between a northern breeding ground and a southern wintering ground. As
shown in Figure 1, the model
assumes two geographically distinct sites linked by rapid migration (thick black
arrows). The duration of the breeding and the wintering seasons are assumed to be
the same. At the beginning of each breeding season, new susceptible chicks are added
to the flock (Figure 1, open
thick arrow); i.e., we assume pulsed reproduction.
Given the uncertain and possibly complex patterns of cross-immunity in wild ducks, we
focus on the dynamics of a single subtype. Hence, we assume that after recovery from
infection, ducks acquire life-long immunity. Thus, within each season, the
epidemiological dynamics are of the familiar type with two transmission routes: direct and environmental. To
derive the environmental transmission functional form, we denote the probability
that a duck escapes infection when exposed to virions by ; note that must decrease with and . Next, we consider a bird that is exposed to virions in two steps: first virions and then virions . Therefore, where is the conditional probability that the duck will escape infection
when exposed to virions after escaping infection when exposed to virions. It is assumed that there is no immunological consequence
of unsuccessful exposure; that is, the probability of escaping infection is
independent of past AIV challenges that did not result in infection (). Thus, we obtain the exponential Cauchy equation [39]
. Since is a decreasing probability function defined on all non-negative
real numbers, the only acceptable solution is where is a constant with unit of . Therefore, environmental infection is modeled using a continuous
Markov chain with a constant rate . Note that the parameter is related to the empirically determined (the dose at which there is a 50% probability of
infection) by the following equation , giving . However, a bird is exposed to virus in the environment via
continuous ingestion of lake water. To model this, we introduce a constant rate that expresses the per capita fraction of the virions ingested per unit time. Thus , which we call exposure rate, is given by the
per capita consumption rate scaled by the lake volume. The
transmission rate per susceptible due to environmental contamination is given by .
Infected ducks shed virus in the environment where the virus persists. We assume that
the viral population is large enough so that these two processes can be captured by the
following differential equation(1)where is the number of infecteds, is the shedding rate and is the decay rate of the virus in the environment. If we divide
the above equation by and use the variable instead of then the equation no longer contains the parameter . Using instead of amounts to measuring the number of virions in units of per shedding rate (i.e., ) which is the unit that we adopt for the rest of the paper. The
environmental transmission rate now becomes , where is the number of susceptibles. Thus, the dynamics of the model
depends on and through their product , which is a re-scaled environmental infectiousness.
Model variables and parameters are presented in Tables 1 and 2, respectively. We use capital subscripts to
denote the season (i.e., for the breeding season and for the wintering season) and lower case superscripts for
geographical location (i.e., for the breeding grounds and for the wintering grounds). For a deep understanding of the
system, we develop two versions of the model: (i) a deterministic system, with
continuous state variables, and (ii) a hybrid framework that consists of discrete
population variables, and stochastic demographic and transmission transition
probabilities together with deterministic virus kinetics. The transmission dynamics
within the continuous model are expressed as coupled ordinary differential equations
and are useful in examining the underlying deterministic clockwork of the system.
Not surprisingly, however, this framework often predicts biologically unrealistic
fractional numbers of infecteds (Mollison's so-called
“atto-fox” phenomenon [40]). Since we are
particularly interested in the processes of extinction and persistence of AIV, we
further refined our study by constructing a stochastic model, where the host
population variables are integer-valued.
The model proceeds as follows.
In this model, the bird population variables are discrete, evolving through a
continuous-time Markov chain integrated using Gillespie's direct method
[41]. The processes that take place throughout a season and their
corresponding rates are summarized in Table 3. The algorithm of the model is as
follows.
We note that a continuous-time Markov chain where all the variables , and are evolved using point processes can be easily constructed by
adding birth (i.e., with rate for and and rate 0 for and ) and death processes (i.e., with rate ) for to the list presented in Table 3. First, it can be shown that if the
variables of this Markov chain are approximately uncorrelated and normally
distributed, then their expectations satisfy the equations of the continuous
model presented in the previous section [42]; i.e., that the
mean-field approximation of this Markov chain is the continuous model
represented by Eqs. (2)–(11). Second, our hybrid model is a good
approximation of the continuous-time Markov chain when the variables are large and the sum of their rates is much larger than the
sum of all the other rates. Indeed, under these conditions, most processes are
births and deaths of virions and other processes occur only sporadically. In
between these processes, the stochastic dynamics of the viral load provided by
the continuous-time Markov chain can be satisfactorily approximated by the
deterministic equations of the hybrid model. We thus conclude that in the case
where virus is abundant the continuous model represents the mean-field
approximation to our hybrid model described above.
As a baseline, we first explored a simplified model that includes fecal/oral
transmission, migration, seasonality and pulsed reproduction, without
environmental transmission. Whether stochastic or deterministic, this model is
unable to reproduce the recurrent pattern of avian influenza epidemics. The
continuous model shows unrealistic infected populations as low as
10−8 individuals (see Text S1)
while the stochastic model undergoes rapid extinction when the infected
population drops to zero.
Figure 2 shows numerical
results for a typical orbit of our deterministic model with both direct and
environmental transmission mechanisms (for definitions of the technical terms in
this section the reader is referred to [43]–[46]).
The orbit rapidly settles to an attractor with a period of two years. Figure 2A–C show the
number of susceptibles, infected and recovered versus time, respectively. The
Fourier power spectrum density of the infected time series is presented in Figure 2D; a peak at is easily noted. Figure 2E and 2F show bifurcation diagrams versus the direct
transmisibility and the re-scaled environmental infectiousness , respectively. The orbits are sampled annually at the end of
the wintering season when, each year, the number of infected is the lowest.
Panel (e) shows a period doubling and an inverse period doubling bifurcation,
while no bifurcations are present in Panel (f). The position of the orbit
presented on the left is marked with dotted lines.
However, the continuous model for the parameters of avian influenza in
populations of 5,000 to 10,000 individuals regularly predicts numbers of
infecteds less than one. Thus, the epidemic would often go extinct as the number
of infected would reach zero. This phenomenon is not captured by a
continuous-state model. Furthermore, note that in Figure 2E model dynamics are predicted to be
rigidly biennial, in contrast to the erratic 2–4 year outbreaks
observed in the wild.
To understand the extinction and persistence dynamics of avian influenza we
integrated the stochastic version of the model. Figure 3A–C show the number of
susceptibles, infected and recovered versus time, respectively, in a simulation
of our stochastic model. In this case, the infected population often goes
extinct and the epidemic is then reignited by environmental transmission. In
direct contrast to the predictions of the deterministic model, a major epidemic
does not occur every two years as such an event is sparked probabilistically
(Figure 3B). In general,
the periodicity of stochastic orbits is larger than that of corresponding
deterministic orbits. If an epidemic does not occur then susceptibles continue
to build up and the next epidemic will thus be more severe. Note that the
incidence peaks of the sporadic epidemics of the stochastic model are higher
than those of the biannual epidemics of the continuous model by about a factor
of three. The Fourier power spectrum density of the infected time series clearly
shows a sequence of peaks corresponding to the annual inflow of susceptibles;
Figure 3D. A peak around is still visible; however, the peak is now very flat, covering
a broad frequency range. The Fourier transform does not appear to provide a very
insightful characterization of the epidemic dynamics owing to tall and narrow
prevalence peaks that do not occur at very regular intervals.
A more useful approach to revealing periodic patterns in the stochastic time
series is a wavelet spectral decomposition. Here we use the
Difference-of-Gaussians (DoG) wavelet since it fits well the tall and narrow
prevalence peaks of the time series (see Text S1). Figure 3E and 3F show the global spectral
decomposition of stochastic orbits in DoG wavelets versus the direct
transmissibility and the re-scaled environmental infectiousness , respectively. Each spectrum is an average over 100 wavelet
transforms of individual stochastic realizations of the orbit. The white solid
lines in Figure 3E and 3F
trace the positions of the local peaks in the spectra versus the corresponding
system parameters. Note that stochastic time series show periodicity larger than
one year (i.e., at ∼2 years and above) for a significantly broader range
of than deterministic time series. Also, note that the dominant
periodicity of the stochastic time series changes very little with , similar to the findings presented in Figure 2F.
It is important to distinguish the parameter sets for which AIV is endemic. While
many model parameters have empirically-established ranges (e.g., host breeding
traits and the duration of infectiousness [21]), the values of
other key parameters, such as the direct transmission rate and the environmental infectiousness are less certain. Therefore, we explore the plane (, ) with all the other parameters of the model given in Table 1. For the continuous
model, the disease-free state is a periodic attractor with period of one year.
This disease-free state loses stability through a transcritical bifurcation
which marks the disease-free/endemic transition. Since the bifurcation is
codimension one, the transition occurs on a line segment in the (, ) plane; see Figure 4. The segment was obtained by numerically solving for the
value of where the transcritical bifurcation of the continuous model
with a given value of occurs.
For the stochastic model, the disease-free/endemic transition is defined in a
more subtle way. The disease-free region is defined by all the parameter sets
for which, in all of the realizations of the model, the number of infected reaches zero in finite time and stays at zero for all
subsequent time, irrespective of the initial conditions. The epidemic region is
defined by all the parameter sets for which there exist realizations of the
model such that, for any moment of time , is not zero for all time once . In the disease-free region the probability of an epidemic is
zero; however, in the endemic region, the probability of an epidemic increases
from zero (close to the boundary with the non-epidemic region) toward one.
Therefore, in the case of the stochastic model, it is more difficult to
numerically obtain a precise border between the disease-free and the endemic
regions. Here we computed the time-average of the infected over 200 years in 100
realizations of the model for a region in the (, ) plane; see Figure 4. (A transient of 100 years was discarded for each
stochastic realization. Numerical analysis reveals that the results are robust
and accurate at these parameters.) Thus, dark blue region corresponds to an
epidemic probability of less than ∼1% and encloses the
disease-free region. Note that for the probability of a sustained epidemic to be
larger than 1%, the re-scaled environmental infectiousness must exceed 103. Simulations did not show sustained
epidemics for low or absent environmental transmission.
Because of the way in which the disease-free/endemic transition is defined for
the stochastic model, it is difficult to compare the epidemic threshold of the
stochastic model with that of the deterministic model. In our case, however, we
may expect that they disagree. The mean-field approximation of a stochastic
model is obtained in two steps. First, one derives an infinite set of ordinary
differential equations that describes how the moments of all orders of the
stochastic variables change with time. Second, under the assumption that all
stochastic variables are uncorrelated and normally distributed, the set of
equations is truncated at the first moment (i.e., moment
closure) which is the expectation [42]. Disagreement
between a stochastic model and its mean-field approximation is expected if the
assumptions on normality or correlations are violated. This typically happens
when any of the population compartments is small. Here, the disagreement at low
numbers of infecteds might be particularly enhanced because of the fact that the
continuous model allows for the number of infected birds to be less than one so
that we always have two different transmission routes of avian influenza. When
the epidemic is at its nadir in the continuous model, the direct transmission
rate does not vanish (the number of infected always stays larger than zero even
though it may be substantially smaller than one) and thus the the chain of
transmission is maintained by both direct and environmental transmission
mechanisms. In contrast, in the stochastic model the numbers of infecteds often
reaches zero. Therefore, AIV maintenance is exclusively due to environmental
transmission. We thus expect that the disease-free region of the stochastic
model is larger than that of the deterministic model.
In Figure 5A and 5B we
present the time-averages of the direct and environmental transmission rates,
respectively. Note that the environmental transmission rate is two orders of
magnitude smaller than the direct transmission rate, yet critical in maintaining
the epidemic. The time-average of the direct transmission rate increases with and , following the pattern of the time-average of the number of
infected in Figure 4.
However, the time-average of environmental transmission rate has a very
different pattern, attaining high values at low values of and decreasing at high ; Figure
5B. Another picture of these contrasting patterns is Figure 6. At low , the environmental transmission rate is relatively high and
increases with as more infected individuals shed more virus in the
environment. A turning point in this scenario happens at when direct transmission starts to dominate. As the direct and
environmental mechanisms of transmission compete for susceptibles, a marked
increase in direct transmission results in a decrease of environmental
transmission.
A fundamental feature of environmental transmission is the fact that it persists
(i.e., does not vanish) even when the number of infecteds (and hence the rate of
direct transmission) is zero. As a result, environmental transmission may
reignite the epidemic. To contrast the persistence characteristics of direct and
environmental transmission, we calculated the average (over stochastic
realizations) fraction of time when direct transmission does not vanish (Figure 5C) and environmental
transmission does not vanish (Figure 5D). The direct transmission rate vanishes when either or while the environmental transmission rate vanishes when either or, quite unlikely, . (Here we assumed that the environmental transmission is
virtually zero when . Computations with yield very similar results.) From Figure 5C, we obtain that direct transmission
at is non-zero at most 30% of the time with a
relatively prominent peak at . In contrast, the environmental transmission is non-zero at
most 70% of the time and the peak is much more shallow over the
chosen range of . Therefore, even though much smaller than the direct
transmission, environmental transmission is much more persistent and may
re-ignite the epidemic when there are no infected left.
An investigation of the time-averaged environmental transmission rate when the
epidemic is reignited was performed as follows (Figure 7). Given a stochastic realization of
the model, we selected the events where the number of infected increases from
zero to one. Say that these events occurred at times and that the corresponding preceding events occurred at times (i.e., for every , the event at time is immediately followed by the event at time ). For each event , we integrated the environmental transmission rate over the time interval (,). Then, the time-averaged transmission rate when the epidemic
is reignited is given by(16)where is the number of susceptibles in the time interval (,) and is a constant. In the analysis presented in Figure 7 we further averaged
over 100 realizations of the stochastic model. The pattern in Figure 7 is comparable to that
in Figure 5B. Note that the
environmental transmission rate that re-ignites the epidemic is less than a
factor of two larger than the average.
In this paper, we have explored the epidemiological dynamics and persistence of avian
influenza viruses, with a view to understanding the respective roles of
environmental transmission and demographic stochasticity. We have found that an framework that includes seasonal migration, pulsed reproduction
and fecal/oral, but not environmental transmission is unable to reproduce the
documented recurrent pattern of avian influenza epidemics. The continuous version of
the model predicts unrealistic infected populations, with values as low as
10−8 individuals (see Text S1), while
the stochastic analogue predicts rapid extinction (similar to the depletions of
infected in Figure 3B). The
unrealistically low infection prevalence is also observed in the continuous model
with added environmental transmission Figure 2E and 2F. Including the interaction between the deterministic
clockwork of the continuous system and demographic noise is fundamental in obtaining
realistic dynamics (with periodicity of 2–4 years), as it is for other
infectious diseases; e.g., see [47],[48] and references therein.
In our full hybrid model, we observe that even small levels of environmental
transmission (a few cases per year) facilitate AIV persistence. Environmental
transmission rates are –on average– hundreds of times smaller
than direct transmission rates, yet they appear critical in sustaining the virus.
The ability of the pathogen to survive in the environment for a long time before
infecting susceptible hosts may thus have profound epidemiological consequences.
The relative influence of environmental transmission for epidemic persistence depends
on the population size. If the population is substantially larger than the critical
community size, then the number of infecteds does not go to zero in between
recurrent epidemics [4],[43],[49] and direct transmission dominates the course of
the epidemic. If, however, the population is small and the number of infecteds goes
to zero, then environmental transmission is a key factor in sustaining the epidemic.
Thus, environmental transmission provides an epidemic persistence mechanism within
populations smaller than the critical community size.
Our results hold for low pathogenicity AIV. The extension to high pathogenicity AIVs,
as evidenced by outbreaks in tufted ducks and pochards [50], awaits additional
empirical information. Another limitation of our model is that we have restricted
our consideration to a single immunological subtype that confers life-long immunity.
We note that partial cross-immunity in a multi-serotype model would enhance the
effective number of susceptibles and, therefore, should be expected to promote
persistence. In reference [51], we address the conditions under which
environmentally and directly transmitted pathogens may coexist.
The actual mechanism of persistence of avian influenza in wild waterfowl may be
complex, including a number of other factors such as spatial and age structures,
waning immunity and strain polymorphism leading to immune escape. Several studies
address the role of spatial heterogeneity in a general framework. For example, Lloyd
and May [52] show in a metapopulation model that persistence of
epidemics (asynchrony of within-subpopulation dynamics) occurs only if the
immigration in between the subpopulations is small. A more recent and thorough
analysis by Hagernaas et al. [53] discussing both oscillatory and non-oscillatory
population dynamics arrives at the same conclusion. Further modeling work is needed
in order to evaluate the relative contribution of other possible persistence
mechanisms.
Further work is also needed to explore our modeling assumption that host populations
form (nearly) closed systems. Empirical evidence suggests that the interaction
between the Eurasian and American clades of migratory birds is so small (despite
overlap in their Alaskan migratory routes) that their exchange of full genome
influenza viruses has yet to be documented [54]. While this observation
supports our modeling assumption, the data on the smaller scale interaction between
flocks of migratory birds within the American continent is insufficient for
validation. Alternate modeling assumptions could be explored theoretically.
Using mathematical modeling, we have investigated the role of environmental
transmission for the pattern and persistence of avian influenza in wild waterfowl
and demonstrated that indeed environmental transmission is a fundamental ingredient
for the modeling of this epidemic. The persistence mechanism induced by enviromental
transmission raises novel problems of epidemic control since traditional strategies
may prove ineffective in the presence of an environmental viral reservoir [55]. Thus,
environmental transmission remains a topic of increasing interest in theoretical
epidemiology.
|
10.1371/journal.pcbi.1003917 | Membrane Partitioning of Anionic, Ligand-Coated Nanoparticles Is Accompanied by Ligand Snorkeling, Local Disordering, and Cholesterol Depletion | Intracellular uptake of nanoparticles (NPs) may induce phase transitions, restructuring, stretching, or even complete disruption of the cell membrane. Therefore, NP cytotoxicity assessment requires a thorough understanding of the mechanisms by which these engineered nanostructures interact with the cell membrane. In this study, extensive Coarse-Grained Molecular Dynamics (MD) simulations are performed to investigate the partitioning of an anionic, ligand-decorated NP in model membranes containing dipalmitoylphosphatidylcholine (DPPC) phospholipids and different concentrations of cholesterol. Spontaneous fusion and translocation of the anionic NP is not observed in any of the 10-µs unbiased MD simulations, indicating that longer timescales may be required for such phenomena to occur. This picture is supported by the free energy analysis, revealing a considerable free energy barrier for NP translocation across the lipid bilayer. 5-µs unbiased MD simulations with the NP inserted in the bilayer core reveal that the hydrophobic and hydrophilic ligands of the NP surface rearrange to form optimal contacts with the lipid bilayer, leading to the so-called snorkeling effect. Inside cholesterol-containing bilayers, the NP induces rearrangement of the structure of the lipid bilayer in its vicinity from the liquid-ordered to the liquid phase spanning a distance almost twice its core radius (8–10 nm). Based on the physical insights obtained in this study, we propose a mechanism of cellular anionic NPpartitioning, which requires structural rearrangements of both the NP and the bilayer, and conclude that the translocation of anionic NPs through cholesterol-rich membranes must be accompanied by formation of cholesterol-lean regions in the proximity of NPs.
| The increasing applications of nanotechnology in medicine rely on the fact that engineered nanomaterials, such as diagnostic and therapeutic nanoparticles, are able to translocate across the cellular membrane and reach their site of action without toxic effects. One of the first steps into assessing the NP cytotoxicity requires a thorough understanding of the nanoparticle-membrane interaction mechanism. We have computationally investigated, using unprecedented spatiotemporal effort, the structure and dynamics of anionic NP partitioning in explicit cholesterol-containing membranes. Our results show that NP partitioning in the membrane is accompanied by the rearrangement of the NP surface ligands and causes the re-organization of the lipids and cholesterol in its vicinity. In this context, our study is an early step towards novel strategies for tailored decoration of NPs aiming to selectively target specific cells based on their cholesterol content.
| Understanding the interaction mechanisms between nanoparticles (NPs) and cell membranes is of critical importance for their use in medical applications [1]–[5]. In these applications, engineered nanostructures are required to contact target cells without damaging essential tissues. The ability of NPs to reach intracellular compartments depends on their morphology and surface chemistry as well as on environmental factors such as cell type, pH, etc. A number of experimental (for a review see [6]) and simulation [7]–[15] studies focused on the effect of NP physico-chemical properties on their interaction with membranes and other liquid/liquid interfaces. In a striking example, Stellacci and co-workers [16] prepared gold NPs (AuNPs) coated with hydrophobic and hydrophilic ligands, which assemble into well-defined striped domains depending on ligand composition. Subsequent in vivo studies on cells suggested both endocytotic and direct translocation mechanisms for striped NPs, whereas NPs with random hydrophilic surfaces could translocate only via endocytosis [17]. Recently, the existence of the striped domains on the surface of these NPs has been challenged by an alternative interpretation of the experimental results, prompting new studies to understand the mechanisms of interactions between NPs and biological membranes [18].
Why do certain NPs easily translocate through biological membranes and others do not? Answering this question would enable us to assess cytotoxicity of various nanostructures and harvest their properties for tailored applications. Unfortunately, current toxicological knowledge about NPs is limited and does not allow for a complete understanding of the effect of nanostructure morphology, composition, and aggregation-dependent interactions with biological systems. Molecular simulations can help rationalize experimental findings by providing a microscopic-level description of the NP-membrane interactions. Such a microscopic-level picture could lead to the creation of predictive models for estimating NP cytotoxicity. This, however, is an immensely challenging task.
Firstly, in vivo processes are too complex to be directly considered in molecular simulations. A cell membrane is a multilayer entity featuring a complex composition of lipids, proteins, and other components, while its environment is a complex solution containing ions, proteins and other species. Therefore, one must resort to simplified models to assess and decouple the influence of various factors on NP cytotoxicity. Secondly, the processes of interest may take place on spatiotemporal scales, which are difficult to access with atomistic simulations. For this reason, simulation studies of NP-membrane interactions are usually based on Coarse-Grained (CG) models, where a group of several atoms is represented by one effective interaction bead.
Not surprisingly, one of these CG models, MARTINI [19], has been extensively applied to the study of NP-membrane interactions [9]–[11], [15], [20]–[22]. Previously, we aimed to understand the role of hydrophobic and hydrophilic domains in the NP translocation process, assuming that these domains do form and remain intact regardless of the NP environment [15]. Therefore, ligand chains on the NP surface were not modeled explicitly, but were represented with CG beads of hydrophilic and hydrophobic types according to the MARTINI model, arranged on the NP surface forming domains of a particular geometry. It was shown that the free energy barrier for NP translocation across the bilayer could be manipulated by controlling the size of hydrophobic domains, onto which lipid molecules tend to self-assemble. In the same study we considered an NP with ligand chains represented explicitly using the same CG model and although the domains on the NP surface were initially designed to follow a particular pattern (i.e. stripes) NP surface ligand flexibility allowed the ligand chains to re-arrange, resulting in a surface chemistry and geometry significantly different from the designed underlying surface pattern as a response to the environment.
In a different study, Lin et al. modeled the interaction of AuNPs coated with flexible ligands with two different types of lipid bilayers [7]. The systems consisted of ∼1,000 lipids and were simulated for a few nanoseconds in unbiased and biased MD. Based on their biased simulations, the authors found that electrostatic interactions between the charged AuNP ligands and the lipid head groups govern the binding of AuNPs to bilayers and that, upon penetration, defective areas and a water pore are induced in the bilayer, while the lipids close to the NP are considerably disordered.
The importance of surface ligand flexibility has also been studied in two articles by Van Lehn and co-workers [23], [24]. The authors also observed that the initial pattern of ligands on the surface of a AuNP is likely not important in the consequent behavior of the NP as the ligands tend to re-arrange within the lipid bilayer so that their polar/charged heads “snorkel” to the surface. In their work, however, the bilayer was modeled implicitly and therefore, the model could not capture the process of lipid reorganization around the NP. Furthermore, only the free energy difference between the NP in the water phase and in the bilayer core was reported, leaving the intermediate states of the system and dynamics of the process outside of the scope of the study.
The above-mentioned studies highlight the importance of a sufficient level of realism of the model to capture the phenomena of interest. An integral aspect of the membrane complexity that has been so far neglected is the presence of other components in the membrane. In particular, cholesterol can substantially influence membrane physical properties, such as fluidity, and induces the formation of lipid rafts, which play an important role in signal transduction and thus in several diseases [25]. Recently, it was shown that cells from certain cancer tissues contain higher concentrations of cholesterol compared to healthy cells [26]. This variation of the cholesterol concentration can be exploited in accurate targeting of cancer cells in advanced drug delivery strategies.
Herein, we investigate the partitioning of anionic NPs into explicit cholesterol-containing dipalmitoylphosphatidylcholine (DPPC) bilayers using biased and unbiased MD simulations. Our results demonstrate that the timescale of spontaneous NP fusion and translocation exceeds the unbiased MD simulated timescale (10 µs). We show that NP partitioning in the bilayer causes rearrangement of the NP surface ligands to facilitate the “snorkeling” of the charged groups towards the lipid head-groups, while its hydrophobic chains remain buried in the bilayer core. Embedding of an NP into the bilayer also forces lipids and cholesterol to re-organize. Specifically, we observe that cholesterol concentration is lower in the vicinity of the NP, while the bilayer structure is more disordered and corresponds to the liquid phase. We conclude with a discussion of the NP decoration as a tuning parameter controlling NP translocation mechanism through the formation of cholesterol-lean patches.
We consider an anionic NP with a core diameter of 4.3 nm coated with regular, striped patterns of hydrophobic (octanethiol, OT) and hydrophilic, negatively charged (11-mercapto-1-undecanesulphonate, MUS) domains in a 2∶1 MUS∶OT ratio, inspired by recent studies [17], [24], [27]. In our model, the bilayer, water phase, and NP are modeled with a CG representation using the MARTINI force field (Figure 1) [19]. The NP is coated with surface ligands that are represented explicitly as flexible chains (Figure S1 in Text S1, Supporting Information (SI)). The modeled NP is negatively charged with a charge density of 1.19 e/nm2 and total charge of −134. Charged ligands are often used as capping agents on NPs to keep them separated via electrostatic repulsion and prevent their aggregation, which has been linked to cytotoxic effects [28].
In general, the burial of a charge in a low dielectric medium such as the lipid bilayer is associated with a substantial energy penalty. For example, atomistic simulations show that in this process the ion remains solvated and its translocation requires formation of water defects, which is accompanied by barriers of 91.7 kJ/mol for Na+ and 98.8 kJ/mol for Cl− [29]. Given the free energy barrier required for a single ion to translocate across a lipid bilayer, we were intrigued by the recent study of van Lehn et al. [24], who predicted that AuNPs decorated with anionic and hydrophobic ligands should prefer, for the majority of systems they explored, to be located inside the lipid membrane compared to the water phase. Depending on the size of the particle, the free energy change between the water phase and the bilayer core varied between slightly positive values (corresponding to the core of the bilayer not being the favorable location for the NP) to as low as −715 kJ/mol for 3.5 nm 2∶1 MUS∶OT NPs and −1,205 kJ/mol for 4.5 nm 1∶1 MUS∶OT NPs [23]. The magnitude of these free energy minima implies that these specific NPs should be trapped in the core of the bilayer indefinitely, although particles of other sizes studied by the authors and featuring not as deep free energy minima in the bilayer core may actually be able to translocate into the cell interior. The authors argued, that this behavior, being similar to one of a purely hydrophobic NP, is due to the ability of the flexible ligands to rearrange, thus increasing the contact area between the hydrophobic residues and bilayer core.
Therefore, we set out to investigate in more detail the origin of this behavior for the anionic NP translocation by exploring the mechanism of association of the NP with the lipid bilayer. To establish a relationship between the NP structure and its interaction mechanism with membranes of different composition, six different systems were considered: a cholesterol-free bilayer and bilayers containing 10%, 20%, 30%, 40%, and 50% mol. cholesterol, each with approximately 8,000–10,000 lipids in total except for the 50% mol. bilayer, which consisted of 14,000 lipids (the exact system sizes are shown in Table S1 in Text S1). The total simulation time was 10 µs for all systems except the system with 50% mol. cholesterol, where the simulation was performed for 8.5 µs. For the modeling of the systems we employed the MARTINI force field (the details of the simulations are provided in the methodology and Text S1) [19]. The parameters of the force field, including the mapping of the atoms to coarse-grained particles, are calibrated to reproduce the free energy differences for partitioning between polar and apolar phases for several reference species. We note, however, that in general the CG MARTINI force field does not reproduce the water defect formation upon translocation of charged groups across lipid bilayers. For example, unlike in atomistic simulations of charged residues, in CG MARTINI representation these residues tend to lose their hydration shell at 0.7 nm away from the bilayer center, which is mostly due to the first hydration shell being included in the coarse-grained representation of charged MARTINI particles [19]. We estimated that the free energy barrier for the translocation of a sodium ion across a DPPC lipid bilayer is 60 kJ/mol (Figure S2 in Text S1, also see Text S1 for details of the calculation). Although this value is lower than the results from atomistic simulations, it is consistent with other CG simulation studies that report energy barriers of 69.0 kJ/mol and 69.2 kJ/mol for Na+ and Cl− translocation, respectively [29]. The barrier for a sodium ion to translocate through a cholesterol-containing bilayer (50% mol.) is even higher (80 kJ/mol, Figure S2 in Text S1); this increase in the free energy barrier is expected as the addition of cholesterol to a fluid phase bilayer decreases the permeability of water and ions [30], [31]. One might conjecture this increase in the energy barrier across cholesterol-containing membranes also for NPs decorated with charged ligands, and go even further to speculate that the details of ligand rearrangement on the surface of NP upon its insertion in the membrane must depend on the composition of the membrane.
Initially, we performed unbiased simulations placing the NP in the water phase 4 nm away from the bilayer surface. Within the first 50 ns of each simulation, the NP partitions at the surface of the bilayer for all different bilayer systems. Figures S3 and S4 in Text S1 show the final simulation snapshots, where the NP adopts a position close to the bilayer-water interface. Within the simulation time we observe no evidence of possible lipid or ligand rearrangement that would propose NP fusion with the bilayer. The number density maps (see Text S1 for the definition) in Figure S4 in Text S1 show no reorganization of the negatively charged end-terminal groups of the NP ligands over the last 500 ns of the simulation.
Therefore, to investigate the structure of the NP-bilayer system upon NP partitioning, we performed a series of unbiased 5 µs CG-MD simulations with the NP placed inside the hydrophobic core of preassembled and equilibrated bilayers. After equilibration, in all considered systems, the NP positions itself either in the core of the bilayer or close to the bilayer-water interface (Figure S5 in Text S1). We observe that the hydrophobic ligands on the NP surface ligands rearrange in order to associate with the bilayer interior, maximizing hydrophobic and minimizing polar contacts. At the same time, the negatively charged MUS termini form salt bridges with the positively charged choline group of the DPPC molecules inducing the so-called “snorkeling effect” (Figure S5 in Text S1). This effect was also observed in the study of Ref. [24]. In Figure 2, the “snorkeling” effect is presented using the number density maps of the negatively charged end-terminal groups of the NP ligands over the last 500 ns of the simulation for each system.
Our observations from both simulation setups (NP initially in water and in the bilayer center) bring to light the microscopic features of the NP-membrane systems described above and most importantly the effect of ligand flexibility in bilayer-NP interactions. It should be noted that the explicit representation of flexible ligands is an important attribute of the system and cannot be neglected as in previous studies, where the NPs were designed as smooth objects with surface patterns [11], [15]. Visual inspection of the MD trajectories indicates that the “hairy” nature of the NP hinders lipid aggregation around it, as the disordered environment of the flexible chains prevents direct access of the hydrophobic lipid tails to the NP surface contrary to what was observed in the case of the smooth NPs [15]. Additionally, the NP anionic surface charges associate with the choline groups of the DPPC bilayer. These features render the “hairy” NP less prone to be incorporated within the bilayer compared to the smooth NPs. Indeed, spontaneous insertion of the specific NP has not been observed in any of our simulations, indicating that longer timescales are required for this process.
Equilibrium MD simulations did not show spontaneous penetration of the NP into the bilayer. Intuitively, this result should have been expected as the translocation of a highly charged NP from water into the bilayer must entail a substantial energy penalty to bury the exposed anionic heads of the ligands into the hydrophobic medium of the bilayer. To assess the free energy barriers associated with this process and elucidate the underlying molecular mechanisms of interaction, we performed Potential of Mean Force (PMF) calculations. Given the intrinsic complexity of PMF calculations of large and slowly evolving systems, an extensive investigation on the sampling time that is necessary for the convergence of the PMF has been performed and is presented in the SI (Figures S6 and S7 in Text S1). It is interesting to note that the free energy for the insertion of the NP from the water to the core of the bilayer fluctuates only between 27 kJ/mol and 29 kJ/mol between sampling times of 50 up to 400 ns for the cholesterol-free membrane (Figure S7 in Text S1). For the 50% mol. cholesterol system, the fluctuation of the barrier is also small, between 49 kJ/mol and 56 kJ/mol between sampling times of 50 up to 300 ns, however its increasing trend does not allow us to conclude on the convergence of the calculations. The PMF profiles with respect to the distance from the bilayer midline are shown in Figure 3 for the cholesterol-free system and Figure S8 in Text S1 for the 50% mol. cholesterol system.
For the cholesterol-free membrane, an energy minimum can be observed at a distance of ≈4.5 nm from the center of the bilayer. At about 4 nm distance from the bilayer center, the polar ligands on the NP surface interact with the positively charged choline groups of the lipids. The NP has to overcome a barrier of , in order to translocate from the bilayer-water interface (minimum) inside the bilayer, indicating a non-spontaneous process on the simulated timescale (Figure 3). In Figure S8 in Text S1, we present the PMF profile calculated from 300 ns per window for the 50% mol. cholesterol membrane. Based on our convergence analysis discussed above, this result should be considered with caution. However, qualitative information provided by comparing the two free energy profiles, i.e. 0% and 50% mol. cholesterol, is consistent with what one would expect: the higher the cholesterol concentration, the more difficult the NP translocation across a membrane.
Although inspired by the same experimental system, the employed models and methodologies are very different in our study and in the article by van Lehn and co-workers, since we use an explicit bilayer, use the MARTINI force field instead of a bespoke force field and for the calculation of the free energy difference we perform Umbrella sampling calculations while van Lehn et al use an approach based on the free energy decomposition [24]. Nevertheless, here we attempt to establish an overarching link to their observations. Some of their systems exhibited zero free energy difference between water phase and the bilayer core for the NP location (and even slightly positive values, depending on the NP size), while the majority of their systems exhibited a substantial free energy minimum in the bilayer core. Our studies reveal positive free energy barrier to NP translocation across the bilayer with the minimum lying at the NP-bilayer interface. The difference in van Lehn's and our results indicate that the free energy penalty is a strong function of whether a charged group is buried inside the bilayer core. The “snorkeling” mechanism allows flexible ligands to avoid this penalty by relocating their charged termini to the surface of the bilayer. The efficiency of this “snorkeling” mechanism should, however, depend on the length of ligands (or, alternatively, on the size of the NP core), their flexibility, and packing. In our model, a fraction of the charged groups remain entangled in the bilayer core, unable to find a pathway to the surface of the bilayer (Figure S5 in Text S1). This entrapment leads to an energy penalty and positive free energy barrier. In some of the systems studied by van Lehn et al., a larger portion (if not all) of ligand charged groups are able to position themselves at the surface of the bilayer, while the hydrophobic residues are only exposed to the bilayer center and interact favorably with lipid molecules. This arrangement leads to a substantially favorable free energy difference for the transport of the NP from the water phase to the bilayer core. For NPs of other sizes this may not be the case, resulting in less negative, zero and even positive free energy differences, in agreement with the observations here. We also emphasize that the process of a NP exiting the membrane may involve a different pathway from the NP entering the membrane. For example, as has been recently observed in our own studies, while a “naked” NP particle may attach to the membrane followed by fusion with it, NP exit from the membrane can occur only in association with lipid molecules attached to it [15]. This invariably complicates the definition of a consistent reaction coordinate and thus PMF calculation. In the present study, we only consider the process of a pristine NP entering (or fusing with) the membrane from the bulk water phase.
We further examined the re-organization of the lipid bilayer structure upon NP inclusion in the unbiased MD simulations, where the NP is initially placed in the bilayer core. Lipid bilayers composed of a single phospholipid species undergo a well-defined phase transition in which the lipid chains change from an ordered or gel state (Lg) to a fluid or liquid-disordered (Ld) state. Within the physiologically relevant temperature range (30–40°C), addition of cholesterol in a concentration above 30% mol. eliminates the Lg-Ld phase transition, and a new, distinct state of the bilayer is observed, termed the liquid-ordered (Lo) phase [32], [33]. This new phase is characterized by an intermediate fluidity between those of the gel and the fluid phases [34].
To describe the structure of the lipid bilayer in the vicinity of the NP, we calculated the spatially averaged second-rank lipid order parameter, , which is a metric of the order of the lipid hydrocarbon chains and characterizes the alignment of lipid molecules with the bilayer normal (Figure 4). The square brackets denote an ensemble average and α is the angle between the bond formed by two CG beads and the bilayer normal. A value of corresponds to perfect alignment with the bilayer normal, to perfect anti-alignment, and to a random orientation of molecules with respect to the bilayer normal. The Ld phase is characterized by order parameters ranging between 0.2–0.5 and the Lo phase order parameters are in the range 0.7–1 [35] (see also the color bar scale of Figure 4). In Figure 4 it is shown that the presence of the NP induces a local increase in the disorder of lipids in all systems. Remarkably, for the systems with high cholesterol concentration (more than 30% mol.), we find that in a spherical segment area of approximately 2–3 nm (depending on cholesterol concentration) radius around the NP, the bilayer is in the liquid-disordered (Ld) phase for all the systems as indicated by the average lipid tail order parameter of ∼0.4.
To further characterize the local structure of the lipid bilayer in the vicinity of the NP, we focused on two additional structural metrics. Radial concentration profiles, c(d), describe the concentration of various molecular species as a function of distance from the NP center of mass. Plotted in Figure 5, these profiles are normalized with the bulk concentration, , of respective species. In addition, Radial Distribution Functions (RDFs) between the MUS terminal groups on the NP surface and various groups on DPPC and cholesterol molecules are shown in Figure S9 in Text S1. The radial concentration profiles of DPPC molecules do not exhibit a significant variation across different systems (Figure 5). However, the behavior of the cholesterol radial concentration profiles is different from that of DPPC groups in several aspects. According to these profiles, the normalized cholesterol concentration is lower in the vicinity of the NP, compared to DPPC concentration, and the effect of the NP presence on cholesterol distribution is still evident even beyond 6 nm away from the NP centre. Furthermore, these radial concentration profiles exhibit substantial variation across different systems, compared to the analogous DPPC profiles. The RDF analysis supports this picture. As seen in Figure S9 in Text S1, cholesterol density is depleted in the presence of the NP, compared to the bulk value, and these effects are seen up to and, depending on the concentration, beyond a 6-nm separation distance from MUS groups. In addition, Table S2 in Text S1 shows the concentration of cholesterol in the 3 nm vicinity of the negatively charged MUS terminal group in comparison with the bulk concentration. This data provides further evidence that the NP prefers to interact with DPPC molecules rather than cholesterol, leading to a local depletion of cholesterol concentration in the vicinity of NP and formation of a region around NP with characteristic features of the liquid-disordered (Ld) phase.
To quantify the preference between the NP and the DPPC/cholesterol molecules, we calculated interaction energies between different species in the system (per NP). In Table 1 interaction energies normalized with respect to the number of molecules within the cut-off distance of 1.2 nm, are shown. At low concentrations of cholesterol, cholesterol molecules tend to move away from the NP, leading to lower interaction energies. As the concentration increases, crowding effects force cholesterol molecules to be closer to the NP, leading to higher values of interaction energies per molecule. This behavior is consistent with the concentration as a function of the distance from the NP center, c(d), and the 2D RDF analysis presented in Figures 5, 6, and Figure S6 in Text S1. In Figure 6 is shown the radial bilayer thickness relative to the NP center (Figure S10 in Text S1 provides an additional metric and analysis of the bilayer thickness). Close to the NP (i.e. at distances of up to 3–4 nm), a thinning of the membrane is observed with respect to the bulk thickness of the bilayer (distances over 4 nm, where the bilayer thickness reaches a plateau value). The strongest effect on the bilayer thickness is observed in the 50% mol. cholesterol bilayer case (Figure 6 and Figure S10 in Text S1). This bilayer thinning can be correlated to the increase of lipid disorder in the vicinity of the NP due to cholesterol depletion; lipid disordering is known to lead to reduced membrane thickness [34]. Moreover, the positively charged phosphate groups of lipid heads are attracted by the negatively charged MUS ligand ends, and are dragged towards the inner part of the bilayer (Figure S11 in Text S1). This effect contributed to the local thinning of the membrane caused by cholesterol depletion and increase in lipid disorder.
The partition mechanism of a striped anionic NP in lipid bilayers is studied herein using state-of-the-art biased and unbiased MD simulations. At the beginning of the translocation process, when the NP is on the surface of the bilayer, the ligand-decorated, “hairy” nature of the NP hinders the self-assembly and aggregation of lipids on the surface of the NP obstructing it from entering the bilayer. As a result, in contrast to previous studies of smooth NPs with similar size and polarity [9], [15], unbiased MD simulations did not capture the actual process of NP fusion and translocation through the membrane on the 10 µs timescale. We note here that recently insights into why such events may require even longer simulation times to be directly detected in molecular dynamics have emerged [36].
NP insertion in the bilayer, studied here by biased and unbiased MD simulations, causes the NP surface ligands to rearrange and “snorkel” the charged groups towards the lipid head-groups, while at the same time the ligand hydrophobic chains remain buried inside the bilayer core. This ligand re-arrangement also leads to formation of contacts and more enhanced interactions between DPPC molecules and hydrophobic ligands on the surface of the NP. Although spontaneous insertion of a NP into the bilayer is not observed in equilibrium MD simulations here, we note that the free energy barriers observed for the NP are much lower than that for the translocation of a single ion. We conclude, that it is this snorkeling mechanism which is responsible for these relatively low free energy barriers. Comparison of these observations with the results of van Lehn et al. [24] suggests, however, that the snorkeling mechanism and the resulting free energy profiles should strongly depend on ligand flexibility and length, as well as on the model of the lipid bilayer. At the same time, this variation in the behavior also implies new opportunities for the design of NPs, based on the structure of ligands as the tuning parameter.
The entrapment of the NP in the bilayer core leads to an energy penalty and positive free energy barrier. Depending on the NP size or effective size, the results presented by van Lehn et al. start from small negative values of the free energy difference, go through deep negative minima and then reach small positive values. In some cases reported in Ref. 24, a large portion (or even all) of ligand charged groups are able to position themselves at the surface of the bilayer, while the hydrophobic residues are only exposed to the bilayer center and interact favorably with lipid molecules due to simulation methodology used. This arrangement leads to a substantially favorable free energy difference of certain NPs between the water phase to the bilayer core. For NPs of other sizes this is not the case, resulting in less negative, zero and even positive free energy differences, in agreement with the observations here.
While the snorkeling of the charged groups towards the surface of the lipid bilayer could be an important mechanism of transport facilitation, it may be hampered by the presence of cholesterol that rigidifies and orders the membrane. To study the effect of cholesterol concentration in the NP translocation, six different bilayers with cholesterol concentrations ranging from 0% to 50% mol. are investigated with unbiased MD simulations. We observe that flexible ligands on the NP surface induce formation of a region corresponding to liquid-disordered phase in the vicinity of the NP in all studied systems. This phenomenon is due to the NP preferential interaction with the DPPC lipids, which leads to effective expulsion of cholesterol around the NP. As a result of the existence of cholesterol lean area around the NP, we could not observe any definite impact of the cholesterol presence on the ability of ligands to snorkel (see Figure S12 in Text S1). This however requires further investigation.
In the absence of unbiased MD simulations capturing the actual process of fusion and translocation, here we attempt to infer possible scenarios prompted by the relative mobility of the membrane components. The DPPC lipid and cholesterol self-diffusion coefficients were calculated using the unbiased MD simulations. The Mean Square Displacement (MSD) and the technical details of the diffusion coefficient calculations are presented in section F of Text S1 (Figure S13 in Text S1 and Table S3 in Text S1). At 40% mol. cholesterol, the cholesterol self-diffusion coefficient is , same as the one for cholesterol (see also Table S3 in Text S1). In contrast, at 10% mol. cholesterol, the cholesterol self-diffusion coefficient is , significantly higher than that for the 40% case. Hence, the mobility of cholesterol strongly depends on the cholesterol concentration of the membrane and thus, in cholesterol-rich membranes, the local re-arrangement of the membrane is possibly hindered by slow diffusion of cholesterol molecules, and the whole fusion and translocation process may become diffusion-limited.
These preliminary considerations implicate yet another possible scenario. A situation is possible where several NPs are already incorporated in the membrane form and stabilize between them larger patches of cholesterol-lean membrane. These patches may feature much higher permeation rates, compared to other regions of the membrane, due to greater mobility of molecules in these patches, higher disorder and lower thickness of the membrane. Such a co-operative adsorption mechanism in the presence of multiple NPs could explain the nonlinear cellular uptake with concentration as has been recently observed [6].
This strong dependence of structural reorganization of the membranes on the presence and concentration of cholesterol significantly affects the drug delivery process in various contexts. With certain cancer cells containing higher levels of cholesterol, it is clear that design of efficient NPs for cancer therapy must be based on a better understanding of the impact of cholesterol on the nanomaterial cellular uptake mechanisms. Therefore, strategies for tailored decoration of NPs aiming to selectively target specific cells based on their cholesterol content, are required. This study is an early step in this direction; moreover, it highlights the importance of a realistic representation of the system under investigation and long simulation timescales, required to capture NP translocation in its full complexity.
In the present study, all membranes are described as two-component lipid bilayers using the MARTINI CG model for lipids, cholesterol, and water [19]. The MARTINI model was also employed for the construction of the NP using an approach previously introduced by us [15]. We emphasize that the NP under investigation, although inspired by previous studies, is purely representative, and its dimension, pattern, and charge are exactly not those of any specific experimentally studied nanoparticle. The CG simulations presented in this article were performed with the GROMACS simulation package, version 4.5.5 [37].
The lipid bilayers were constructed using ∼4,000–8,000 1,2-dipalmitoyl-sn-glycero-3-phosphocholine (DPPC) lipid molecules and the respective cholesterol molecules to achieve the desired cholesterol concentrations. To setup each system we performed the following procedure: (a) mixing system components starting from random initial positioning of lipids, cholesterol, and water. The mixture was simulated for several tens of nanoseconds until the eventual formation of the bilayer was observed, (b) several properties of the bilayer were then checked in order to ensure almost perfectly symmetric distribution to within 10 molecules of cholesterol and lipids among the two leaflets since an asymmetric cholesterol distribution affects membrane curvature [38], and (c) in addition, occasional runs with larger simulation boxes were performed to assess finite size effects. In all instances, bilayers were simulated in excess water, with approximately 100 water molecules per lipid (or, equivalently, 28 CG water molecules per lipid), a ratio well above the degree of hydration observed in multilamellar vesicles for fluid (Ld) bilayers and the even less hydrated gel (Lg) bilayers [39]. The water phase has been almost doubled from the original lipid∶CG water ratio of 1∶15 suggested by Ref. 7 to allow for the complete inclusion of the NP and in order to ensure that individual uncoupled bilayers separated by bulk water will be simulated, thus avoiding artifacts caused by interacting periodic images of the system.
Each system was initially equilibrated for 1 µs. Within this time, tensionless membranes were formed and thermodynamic properties converged to equilibrium values. During the simulation course the lipids remain in the bilayer and do not break in cylindrical micelles. Following equilibration, an NP was placed at the center of the bilayer using VMD [40].
MD simulations were performed with constant pressure, temperature, and number of particles (NPT ensemble) [37]. The temperature was kept constant at 323 K using the Berendsen thermostat with a relaxation time of 1.25 ps [41]. The pressure of the system was semi-isotropically coupled and maintained at 1 bar using the Berendsen algorithm with a time constant of 0.22 ps and a compressibility of 3e-5bar−1. The non-bonded potential energy functions were cut off and shifted at 12 Å, smoothly decaying between 9 and 12 Å for van der Waals forces and throughout the whole interaction range for the treatment of electrostatic forces. The simulations were performed using a 20 fs integration time step. Six different cholesterol ratios, ranging from a cholesterol-free bilayer to a 50% ratio were considered; the exact composition of the different membranes is presented in Table S1 in Text S1.
The Potential of Mean Force (PMF) as a function of the distance between the NP and the center of the lipid bilayer has been calculated using the Umbrella Sampling protocol of Gromacs for the cholesterol-free membrane and the membrane containing 50% mol. cholesterol concentration. The distance between the initial NP position and the center of the bilayer was 7 nm, which was divided into 36 windows of 0.2 nm each. In each window, a different initial configuration was set up with the NP placed at the corresponding distance from the center of the bilayer. Then, the system was subjected to four subsequent annealing simulations at 450 K, 400 K, 380 K, and finally 323 K. The temperature coupling was restarted at the 0 ps value after 400 ps, 1000 ps, and 2000 ps. Then, the biasing potential was applied in order to restrain the NP at the given position, and the system was left to equilibrate at the restrained position for 200 ns. This procedure was followed in order for the lipids to re-arrange correctly around the NP. Subsequently, the system was simulated for another 400 ns in the case of 0% mol. cholesterol and 300 ns in the case of 50% mol. cholesterol, with the biasing potential applied to restrain the center of mass of the NP at a required distance from the center of the bilayer. A single PMF profile required 36 simulations (36 windows) and a total simulation time of 21.6 µs (18 µs in the 50% mol. cholesterol membrane). A force constant of 750 kJ mol−1 nm−2 was applied, following the approach by Gkeka et al. [15] To examine the PMF sampling and convergence in further detail, we investigated the PMF change when increasing the sampling time for each window starting with 50 ns up to 400 ns (or 300 ns for the cholesterol-containing membrane) with 50 ns intervals. The initial equilibration period was 200 ns in all cases.
The simulated system features a large enough water phase to avoid possible effects associated with the system size and NP-NP interactions over periodic boundaries, as tested by considering both larger and smaller systems. Note that for the 50% cholesterol system, we doubled the size of system in order to avoid any tension coupling between the periodic image of the NP, so the final system contained ≈14,000 lipids and ≈400,000 CG water particles (∼2,000,000 water molecules) corresponding to a system size of 648,302 CG particles. This system size was used to avoid finite size effects and artifacts from the nano-object interacting with its periodic images. In Figure S14 in Text S1 we present a representative simulation snapshot at zero NP-bilayer distance, which demonstrates that despite the size of the membrane, minimal buckling was involved. The NP was left free to rotate around its restrained center of mass. In order to obtain the unbiased PMFs, we used the weighted histogram analysis method (WHAM) [42] with 200 bins and a tolerance of 10−7 for the convergence of WHAM equations.
|
10.1371/journal.ppat.1002486 | Human Herpesvirus 8 (HHV8) Sequentially Shapes the NK Cell Repertoire during the Course of Asymptomatic Infection and Kaposi Sarcoma | The contribution of innate immunity to immunosurveillance of the oncogenic Human Herpes Virus 8 (HHV8) has not been studied in depth. We investigated NK cell phenotype and function in 70 HHV8-infected subjects, either asymptomatic carriers or having developed Kaposi's sarcoma (KS). Our results revealed substantial alterations of the NK cell receptor repertoire in healthy HHV8 carriers, with reduced expression of NKp30, NKp46 and CD161 receptors. In addition, down-modulation of the activating NKG2D receptor, associated with impaired NK-cell lytic capacity, was observed in patients with active KS. Resolution of KS after treatment was accompanied with restoration of NKG2D levels and NK cell activity. HHV8-latently infected endothelial cells overexpressed ligands of several NK cell receptors, including NKG2D ligands. The strong expression of NKG2D ligands by tumor cells was confirmed in situ by immunohistochemical staining of KS biopsies. However, no tumor-infiltrating NK cells were detected, suggesting a defect in NK cell homing or survival in the KS microenvironment. Among the known KS-derived immunoregulatory factors, we identified prostaglandin E2 (PGE2) as a critical element responsible for the down-modulation of NKG2D expression on resting NK cells. Moreover, PGE2 prevented up-regulation of the NKG2D and NKp30 receptors on IL-15-activated NK cells, and inhibited the IL-15-induced proliferation and survival of NK cells. Altogether, our observations are consistent with distinct immunoevasion mechanisms that allow HHV8 to escape NK cell responses stepwise, first at early stages of infection to facilitate the maintenance of viral latency, and later to promote tumor cell growth through suppression of NKG2D-mediated functions. Importantly, our results provide additional support to the use of PGE2 inhibitors as an attractive approach to treat aggressive KS, as they could restore activation and survival of tumoricidal NK cells.
| Natural Killer (NK) cells are part of the innate immune response against virus infections and tumors. Their activation is the net result of signals emanating from a panel of inhibitory and activating receptors recognizing specific ligands on target cells. Human Herpes Virus 8 (HHV8) is an oncogenic virus responsible of Kaposi Sarcoma (KS), a multifocal angiogenic tumor. How NK cells contribute to the control of infection by HHV8 infection and development of KS, is unclear. In this paper, we show different strategies used by HHV8 to escape NK cell response. Patients with asymptomatic infection or KS have down-modulated expression of NKp30, NKp46 and CD161 receptors. In addition, patients with active KS show additional down-modulation of the NKG2D activating receptor, associated with impaired NK-cell cytotoxicity against target cells. Resolution of KS correlates with regained NKG2D expression and cytotoxic function. We present evidence that down-modulation of NKG2D is mediated by inflammatory prostaglandin E2 (PGE2), known to be released by KS cells, and show that PGE2 acts by preventing IL-15-mediated activation of NK cells. These results strongly support the use of PGE2 inhibitors as an attractive approach to treat active KS.
| Human herpesvirus 8 (HHV8), although known as Kaposi's sarcoma-associated herpesvirus (KSHV), is a γ herpes virus able to establish a predominantly latent, life-long infection in host's monocytes, dendritic cells (DCs), B lymphocytes, and endothelial cells. HHV8 is the etiological agent of Kaposi's sarcoma (KS), a multifocal angiogenic tumor consisting of spindle-shaped cells of endothelial origin and infiltrating leukocytes [1], [2]. HHV8 lytic cycle generally occurs following primary infection, and rapidly the virus enters the latent state. Reactivation leads to the initiation of the lytic cycle, which is necessary for virus propagation and survival. Within KS lesions, HHV8 infection is predominantly latent. KS is the most common neoplasm in untreated AIDS patients. It also occurs in immunosuppressed organ transplant recipients, and in some African or Mediterranean populations in the absence of overt immunosuppression (classical KS). The marked decline in the incidence of AIDS-KS since the advent of antiretroviral therapy (ART), and the frequent resolution of transplant-related KS after reduction of immunosuppression, highlight the key role of cellular immune responses in the control of HHV8 infection. We and others recently demonstrated the crucial role of HHV8-specific cytotoxic T lymphocytes (CTL) in controlling HHV8 replication, preventing malignancies in latently infected subjects, and conferring genuine resistance to persistent infection [3], [4]. The multiple mechanisms elaborated by herpesviruses to escape immune responses prompted us to explore further other immune cells involved in the control of HHV8 infection.
NK cells play a key role in early control of viral infections, through direct lysis of infected cells and secretion of cytokines controlling viral replication. NK cells also influence specific T-cell priming through direct cross talk with DCs, and thus participate to the establishment of antiviral adaptive responses [5]–[7]. NK cells are able to prevent and control the development and dissemination of tumors [8]. NK cells are therefore likely to represent critical obstacles HHV8 needs to effectively overcome, not only very early during infection prior to de novo viral gene transcription, but also later for the maintenance of persistent infection and establishment of tumors. NK cell activity is tightly regulated by a fine balance between activating and inhibitory signals [9]. The latter are mostly generated by the binding of HLA-1 molecules to the Killer cell Immunoglobulin-like Receptors (KIRs) and CD94/NKG2A, which guarantees that healthy self cells will be spared from NK cell-mediated lysis. Activating receptors, in particular NKG2D and the natural cytotoxicity receptors (NCR) NKp30, NKp44 and NKp46, detect the presence of stress-induced or infectious non-self ligands on abnormal cells. Other receptors, such as CD94/NKG2C, DNAM-1, NKp80, 2B4 and CD161, can modulate NK cell effector functions.
HHV8 has exploited several evasion mechanisms to avoid immune recognition [10]. In particular, the K3 and K5 ubiquitin ligases, expressed during the early lytic cycle, downregulate HLA-1 molecules on infected cells to avoid virus-specific CTL recognition [11]. This will at the same time sensitize the virus to missing-self recognition by NK cells, unless other evasion mechanisms operate simultaneously. Interestingly, K5 also downmodulates NKG2D ligands and the ICAM-1 adhesion molecule [12], thus helping HHV8 to evade NK cell surveillance in the early phase of HHV8 infection before establishing latency, and later during reactivation and viral replication. NK cells from HIV-viremic AIDS patients with active KS have a decreased cytolytic capacity against HHV8-infected body cavity B-cell lymphoma (BCBL-1) cells [13]. However, little is known about how HHV8 by itself, in the absence of HIV infection, prevents NK cells from killing latently-infected endothelial cells and KS tumor cells.
To address these questions, we investigated NK cell phenotype and functions in a large series of HHV8-infected subjects. Our data revealed substantial alterations in the expression of several NK cell receptors, even in asymptomatic HHV8 carriers. In addition, NK cells from patients with active KS showed a significant decrease of NKG2D expression, which was associated with impaired cytotoxic capacity. We identified PGE2, a known tumor-derived inflammatory molecule, as a factor responsible for NKG2D down-modulation, and for inhibition of IL-15-mediated activation and survival of NK cells. These observations are consistent with sequential immunoevasion mechanisms that may allow HHV8 to escape NK cell recognition at early stages of infection in order to establish latency, and later to promote tumor cell growth. Importantly, they give further support to the idea that PGE2 inhibition based therapy might provide an effective way to treat the active KS lesions.
We addressed the putative influence of HHV8 infection on NK cell phenotype in a cohort of 70 HHV8-infected individuals, including 25 asymptomatic carriers (HHV8+ KS−) and 45 KS patients (HHV8+ KS+) in comparison with 45 HHV8-negative controls, all sub-grouped according to the presence or absence of HIV co-infection. To avoid any confounding effect of HIV replication on the NK cell repertoire, all HIV+ subjects from the different subgroups were HIV-aviremic, and were matched for age, CD4 T cell count, CD4 nadir, and duration of disease and ART.
No gross abnormality in NK cell distribution was observed in the different patient groups, with levels of total NK cells, and relative proportions of CD56bright and CD56dim NK cell subpopulations being comparable to those in controls (Figure 1a). Accumulation of dysfunctional CD56-negative NK cells was recently reported in HIV-viremic patients, with suppression of HIV-viremia upon ART restoring normal proportions of CD56+ NK cells [14]–[16]. In line with the fact that all study subjects were HIV-aviremic, we did not observed any abnormal expansion of CD56-negative NK cells in HHV8-infected patients (Figure S1).
We next analyzed HLA class 1-specific NK cell receptors, including those belonging to the KIR family (KIR2DL1/S1, KIR2DL2/L3/S2, KIR3DL1/S1, KIR2DS4), and the HLA-E-specific CD94/NKG2A and CD94/NKG2C receptors (Figure 1b). Although large inter-individual variations were observed, the mean percentage of cells expressing individual KIRs or NKG2A was overall similar in the different study groups. The fraction of NKG2C+ NK cells varied within a wide range in infected patients (<0.1% to 54%), but the difference with uninfected controls was not statistically significant. CMV seropositivity has been associated with high frequencies of NKG2C+ NK cells [17], [18]. We thus wondered if the enrichment of NKG2C+ NK cells in some HHV8-infected patients was related to CMV co-infection. Notably, 68 out of the 70 HHV8- and/or HIV-infected patients were CMV IgG positive, and the two remaining CMV IgG-negative patients had very low NKG2C+ NK cell frequency (1.56% and 0.25%, respectively). Therefore, it is likely that previous CMV infection has driven expansion of NKG2C+ NK cells in HIV- or HHV8-infected patients.
We next studied expression of NK cell receptors that recognize virus-associated or stressed-induced molecules (Figure 2a). Levels of NKp30, NKp46 and CD161 were significantly decreased in HHV8-infected patients compared to controls, whatever the presence or absence of KS, indicating that HHV8 is able to skew the NK cell receptor repertoire in otherwise healthy individuals, before overt tumor transformation. DNAM-1 expression was not significantly different between the groups. NKp44 was never detected (not shown). Notably, the profile of NKG2D expression was clearly distinct from that of other receptors, as NKG2D levels were decreased in patients with classical KS (HIV− HHV8+ KS+), but neither in HIV+ KS+ patients nor in asymptomatic HHV8-infected subjects (HHV8+ KS−). Because all HIV+ KS patients were HIV-aviremic upon ART for more than 1 year and showed complete clinical remission of KS at time of study, we wondered if NKG2D expression was correlated with the KS activity. Indeed, NKG2D levels were twofold lower in patients with active KS (all classical KS) than in patients with resolved lesions or in healthy controls (Figure 2b,c). By contrast, expression of other receptors was similarly decreased in KS patients, whatever disease activity (data not shown).
The variable degrees of NK cell receptor modifications among patients prompted us to examine whether there was any correlation between phenotypic changes. We observed highly significant correlations between expression levels of NKp30, NKp46 and CD161, the three receptors downmodulated in asymptomatic HHV8-infected subjects (Figure 3), suggesting that a common mechanism sustained these alterations. At contrast, there was no correlation between expression of any of these receptors and expression of NKG2D.
Taken together, these results show that HHV8 infection selectively imprints the NK cell receptor repertoire even at an asymptomatic stage, with a coordinate decrease in NKp30, NKp46 and CD161 expression. At a more advanced stage, a specific down-modulation of NKG2D occurs in patients progressing to KS, which is likely mediated by a distinct, KS-specific mechanism.
As a first step to study the molecular interactions underlying NK cell recognition of infected cells, we analyzed the phenotype of HHV8-infected endothelial cells, which are considered as one of the precursors of KS tumor cells. Primary infection of the microvascular endothelial cell line HMEC with a recombinant virus, rKSHV.152, expressing the green fluorescent protein (GFP) and neo (conferring resistance to G418) [19] results in the establishment of latent HHV8 infection, with a very few percentage of cells undergoing lytic replication, a situation thought to mimic in vivo replication. HHV8-infected HMECs exhibited a two to threefold decreased expression of HLA-1 and ICAM-1 compared to uninfected HMECs (Figure 4a left panel), although the remaining expression was still very consistent. We thought that the partial dowmodulation of HLA-1 and ICAM-1 might be explained by the presence of low levels of early lytic immunoregulatory proteins such as K3 or K5 in these latently-infected cells, as previously reported [20], [21]. Indeed, qPCR analysis demonstrated some expression of K3 and K5 mRNAs in infected cells (Figure 4b). In line with the relatively conserved expression of HLA-1 molecules, HHV8-infected HMECs showed a consistent expression of the non-classical HLA class 1b molecule HLA-E, which is dependent on signal sequences from classical HLA-1 molecules for its stabilization [22]. Because the CMV-encoded UL40 protein can also contribute a peptide cargo for HLA-E [23], we searched for potential HLA-E binding peptide motifs in the HHV8 proteins. We did not find any relevant peptide motif, indicating that HHV8 by itself is unlikely to stabilize HLA-E expression (data not shown).
We next analyzed whether HHV8-infected cells expressed ligands for NK cell receptors. We did not observe any expression of NKp30 or NKp46 ligands, or of LLT1, the ligand of CD161, on infected nor uninfected cells. When looking at NKG2D ligands, it appeared that MICA and MICB were up-regulated on HHV8-HMECs compared to uninfected cells, while ULBP-1 was not detected, and ULBP-2 and ULBP-3 were similarly expressed. The ligands of DNAM-1 (CD155/PVR and CD112/Nectin-2) were strongly expressed on both uninfected and infected cells, with a stronger expression of CD155 on HHV8- HMECs.
In attempt to study a model of HHV8-infected cells more clinically relevant to KS than microvascular endothelial cells, we generated an immortalized HIV-negative KS-derived cell line. These SV2G cells exhibited phenotypic characteristics of endothelial cells, as demonstrated by expression of CD146, CD131 and CD141/thrombomodulin (data not shown). However, HHV8 genome was lost early after the first 2 passages, as in most KS-derived cell lines [24]. We thus infected SV2G cells in vitro with rKSHV.152, used above for infecting HMECs. The resulting HHV8-SV2G cell line showed predominantly latent infection, with some expression of K3 and K5 mRNAs, like HHV8-HMECs (Figure 4b). Comparing the phenotype of SV2G and HHV8-SV2G cells, we observed that, for most markers analyzed, modifications paralleled those observed in HHV8-HMEC cells, except for an increased expression of ULBP-2 and ICAM-1 and a weak but significant detection of NKp30 ligand in HHV8-SV2G cells compared to uninfected cells (Figure 4a, right panel).
Collectively, these data show that persistently HHV8-infected cells, which show the same latency program as KS spindle cells, express a variety of ligands that should allow engagement of activating NK cell receptors such as NKG2D, DNAM-1 and NKp30. At the same time however, they show a decreased, but yet strong expression of HLA-1 molecules that likely protects them from NK cell lysis.
To determine whether NK cell receptor/ligand interactions occur at the tumor site, we performed immunohistochemical staining of KS biopsies. In line with the flow cytometric data on infected cells, we readily detected expression of HLA-1 and MICA/B molecules in tumor cells (Figure 4c). We did not observe any staining using the NKp30-Fc or NKp46-Fc reagents. Ligands of DNAM-1 and CD161 could not be analyzed due to the lack of commercially available markers working in paraffin-embedded tissues. To our surprise, although large inflammatory infiltrates were observed in most KS samples, we did not detect any CD56-positive cell, suggesting that NK cells did not reach tumor lesions, or could not survive in the tumor microenvironment.
We next evaluated the putative consequences of receptor/ligand modifications on NK cell functions. NK cells from patients with active or resolved classical KS and healthy controls were used as effector cells in CD107a degranulation and intracellular IFNγ production assays in the presence of uninfected or HHV8-infected KS-derived target cells. As anticipated from the relatively strong HLA-1 expression on these target cells, NK cell degranulation was weak and not different between controls and KS patients, whatever the KS activity (Figure 5a). However, degranulation was slightly but significantly higher in the presence of HHV8-infected compared to uninfected targets. Whether this was related to the observed up-regulation of NKG2D, DNAM-1 and NKp30 ligands on infected cells could not be determined. Responses to PMA/ionomycin, used as positive control, were not impaired in patient NK cells (mean CD107a+ cells 40.5% compared to 45.1% in controls), indicating that the granule-exocytosis pathway was intact. Intracellular IFNγ accumulation was almost undetectable in all conditions, both in controls and patients (data not shown).
We then evaluated the ability of NK cells to recognize K562 target cells, which not only lack HLA-1 molecules, but also express ligands for NKG2D, DNAM-1, and NKp30 (personal data and [25]). Compared to healthy controls, patients with resolved classical KS showed intact NK cell degranulation in spite of decreased expression of NKp30, NKp46 and CD161. At contrast, patients with active KS showed significantly impaired NK cell degranulation (Figure 5b), supporting our hypothesis that the NKG2D down-modulation observed in these patients might alter NK cell lytic capacity. Indeed, monoclonal antibody-mediated masking of NKG2D sharply reduced K562-induced degranulation of NK cell from healthy controls and resolved KS patients, but had no effect on NK cells from patients with active KS (Figure 5c). Moreover, we observed a positive correlation between CD107a degranulation and expression of NKG2D (Figure 5d), but not of NKp30 or NKp46. Lastly, comparative analysis of NK cells obtained before and after successful treatment of active classical KS in 3 patients showed that regression of KS was associated with significant restoration of NK cell degranulation, and a parallel increase in expression of NKG2D, but not of NKp30 and NKp46 (Figure 5e).
Taken together, these results suggest that NKG2D may play an important role in the control of KS progression in HHV8-infected individuals.
Shedding of NKG2D ligands constitutes a major countermechanism of tumors to subvert NKG2D-mediated immunosurveillance. Thus, soluble MICA released from tumor cells by proteolytic cleavage drives down-regulation of NKG2D and is associated with compromised immune response and progression of disease in cancer patients [26]–[28]. We quantified soluble MICA in the serum of KS patients and controls, but found similar low levels in both groups. In addition, we did not detect soluble MICA in culture supernatants of HHV8-infected or uninfected cells (data not shown).
HHV8-infected endothelial cells secrete a variety of pro- and anti-inflammatory cytokines, growth factors and angiogenic factors [29]–[31]. In addition, Cyclooxygenase-2 (COX-2) and its metabolite prostaglandin E2 (PGE2), two pivotal proinflammatory molecules, have been shown to play crucial roles in the establishment and maintenance of HHV8 latency, and in inflammatory, angiogenic and invasive events during HHV8 infection [32]–[34]. In attempt to identify the factors responsible for phenotypic changes of NK cells in HHV8-infected patients, we focused on IL-10, TGFβ IL-8, VEGF and PGE2, because they have been involved in the modulation of NK cell functions [35]–[40]. Serum levels of IL-10 and TGFβ were low and not significantly different between patients and controls (data not shown). By contrast, levels of VEGF, IL-8 and PGE2 were significantly increased in KS patients, particularly in those with active classical KS (Figure 6a).
To determine if these factors could modify the NK cell phenotype, control PBMCs were exposed in vitro for 48 h to VEGF, IL-8, or PGE2, after which NK cell receptor expression was evaluated (Figure 6b). TGFβ used as positive control, decreased expression of NKG2D and NKp30 as described [35], but had no significant effect on NKp46, DNAM-1 or CD161. IL-8 and VEGF did not modify the NK cell phenotype. Notably, PGE2 induced a significant decrease of NKG2D expression, but no reproducible modification of the other NK cell receptors. The PGE2-induced down-modulation of NKG2D was dose-dependent (Figure 6c), and was already observed at concentration comparable with those found in the serum of some KS patients. In line with this observation, NKG2D levels on patient NK cells negatively correlated with PGE2 serum levels (r = −0.70, p 0.01).
Taken together, our results suggest that PGE2 may specifically alter NKG2D expression on NK cells, thus preventing NKG2D-mediated elimination of KS cell precursors and favoring the development and/or progression of KS in persistently infected patients.
IL-15 plays a pivotal role in the activation, function and survival of NK cells. IL-15 is a surface-bound cytokine, presented by dendritic cells via its high-affinity receptor, IL-15Rα to the neighboring NK cells that express IL-2/IL-15Rβ and γ chains. PGE2 has been reported to suppress cytotoxicity of NK cells through down-regulation of IL-15Rγ [39]. Notably, IL-15 is also a potent inducer of NKG2D expression [41], and IL-15 signaling potentiates NKG2D-mediated cytotoxicity of NK cells [42]. To determine if PGE2 could inhibit IL-15-induced up-regulation of NKG2D, control PBMCs exposed for 48 h to 5 ng/ml of IL-15 in the presence or absence of PGE2 (10–1,000 ng/ml), after which NK cell phenotype was analyzed (Figure 7a). IL-15 alone strongly up-regulated expression of NKG2D, NKp30 and CD161, but had no or minor effect on NKp46 and DNAM-1. PGE2, even at low concentration (10 ng/ml), fully abrogated the IL-15-induced up-regulation of NKG2D and NKp30, partially inhibited up-regulation of CD161, and had no effect on expression of NKp46 and DNAM-1.
Because immunochemistry did not detect any CD56-positive cells within KS lesions, we next wondered if PGE2 could mediate a defect in survival of NK cells. We analyzed IL-15-induced NK cell proliferation (expression of Ki67) and survival (expression of the anti-apoptotic protein Bcl-2), together with surface expression of IL-15Rβ and IL-15Rγ (Figure 7b). IL-15Rβ was expressed on almost all NK cells, and was not significantly modified by PGE2. Conversely, PGE2 fully prevented the IL-15-induced up-regulation of IL-15Rγ, as already described. Furthermore, PGE2 was able to abolish the proliferative and pro-survival responsiveness of NK cells to IL-15. Staining for annexin V indicated that PGE2, at the concentrations used (up to 10 µg/ml), did not affect NK cell viability (not shown).
Taken together, our results indicate that PGE2 inhibits IL-15 signaling in NK cells through down-regulation of IL-15Rγ, thus preventing IL-15 from promoting NKG2D signaling.
Infection by the oncogenic virus HHV8 raises issues of control of latent infection and control of tumor growth. HHV8 must overcome host's immune responses not only very early during infection prior to de novo viral gene transcription, but also after latent viral gene expression and later on during tumor transformation. Furthermore, the virus must simultaneously avoid innate and adaptive responses, using strategies that are sometimes mutually exclusive. Multiple evasion mechanisms have been dedicated by herpesviruses to thwart NK cell responses [43]–[45]. They can downmodulate ligands for NK cell activating receptors, provide competitors for their ligands, interfere with their translation, or directly target the activating receptors. It is likely that several viral inhibitor mechanisms play in concert to simultaneously or sequentially prevent NK cell activation. Our study shows for the first time to our knowledge, that distinct NK cell modifications are observed at the different stages of HHV8 infection, suggesting that selective pressure allows the virus to evade the successive waves of host immune responses.
First, asymptomatic HHV8 carriers, as well as KS patients, exhibited significant down-regulation of the NKp30 and NKp46 activating receptors. Whether ligands of these receptors were expressed on HHV8-infected cells could not be ascertained in the absence of specific antibodies, although our data suggest that at least NKp30 ligand was present at the surface of infected cells. Thus, elimination of NKp30 ligand-expressing cells might be compromised in case of NKp30 down-modulation on effector cells. Expression of CD161 was also reduced in asymptomatic HHV8 carriers. This observation was rather surprising, given CD161 is described as an inhibitory receptor in NK cells [46], [47]. We did not evidence expression of its ligand LLT1 on HHV8-infected cells, implying that CD161 may not directly participate in the elimination of these cells. Interestingly, LLT1 is expressed on TLR-activated DCs and B cells [48], suggesting that it may contribute to their resistance to NK cell-mediated lysis. The loss of CD161 might thus result in the accumulation of a population of NK cells with a lower activation threshold that may be more easily triggered, which could lead to the elimination of activated DCs and participate in the defective establishment of antiviral adaptive responses. Because HHV8 infection of monocytes/macrophages and B lymphocytes has been demonstrated in KS patients [49], [50], further studies are required to explore LLT1 expression on these cells, and address whether reduced expression of CD161 modifies NK-DC interactions during HHV8 infection. Notably, expression of DNAM-1, another key receptor in NK-cell mediated recognition of several tumors [51]–[54], was not significantly altered in HHV8-infected individuals. Since DNAM-1 ligands were strongly expressed on HHV8-infected cells, it is possible that at least the DNAM-1 recognition pathway may operate in NK cells during HHV8 infection.
Our observation of a co-modulation of NKp30, NKp46 and CD161 in HHV8 infected subjects suggested us that common microenvironmental factors, acting early during asymptomatic stage of infection, might be involved. Multiple inflammatory and angiogenic factors are produced by HHV8-infected cells. We could not identify a unique factor capable to induce concomitant modifications of these receptors. Although IL-8, VEGF and PGE2 were present in high amounts in patient sera, as reported in AIDS-related KS patients [2], [55], [56], they had no significant effect on expression of these NK cell receptors. TGFβ down-regulated the surface expression of NKp30 as reported [35], [57], but not that of NKp46, or CD161. We cannot exclude that intercellular contact indirectly contribute to alter NK cell phenotype, for instance by inhibiting the functional maturation of DCs, and therefore compromising the DC-NK crosstalk. An effect of the microenvironment on NK cell precursors could also determine their skewed maturation during HHV8 infection, leading to the prevalent expansion of mature NK cells with an altered phenotype. However, we did not observe any skewing in the distribution of CD56bright, CD56dim and CD56-negative NK cell subpopulations, as reported in other chronic infections (HIV, HCV). Finally, the possibility that HHV8 infection of NK cells themselves may modify their phenotype cannot be excluded so far [58]. Various leukocyte subsets support HHV8 latency, including B cells, monocytes, macrophages, DCs and even CD34+ hematopoietic progenitors [49], [50], [59]–[62]. Studies in HHV8-infected NOD/SCID mice demonstrated the presence of LANA+ NK cells in the spleens, suggesting that HHV8 can also target NK cells [63], but this has not been confirmed in the human.
Secondly, in addition to the decreased expression of NKp30, NKp46 and CD161, patients with active classical KS exhibited a specific down-modulation of NKG2D and a parallel defect in NK cell lytic capacity. Resolution of KS after treatment correlated with restoration of NKG2D levels and NK cell activity. A decreased NK cell activity was previously reported in AIDS patients with progressive KS. NK cell function was restored upon ART treatment, but whether this was only related to HIV clearance was not determined [13]. Our results indicate that HHV8 by itself is responsible for the down-modulation of NKG2D in HIV-negative patients with classical KS. Consequently, NKG2D-mediated NK cell cytotoxicity is hampered. A small fraction of KS cells express the cascade of lytic cycle genes, in particular K5, which down-modulates HLA-1, ICAM-1 and certain MICA/B molecules [12]. Loss of surface MICA/B may help HHV8 to evade NK cell surveillance in the early phase of lytic HHV8 infection before establishing latency. This mechanism is clearly not operational in persistently infected KS cells, which express high levels of MICA/B molecules. Instead, the decreased expression of NKG2D appears as an efficient mean for HHV8-infected cells to evade anti-viral immunity and develop their tumoral program. A similar strategy of decreasing NKG2D expression is also adopted by another persistent virus, hepatitis C virus, to evade NK-cell mediated responses in chronically-infected patients [64]. Our results thus reinforce the notion that NKG2D plays an important role not only in control of viral infections, but also in surveillance of tumor development, by protecting the host from tumor initiation and growth [65]–[67].
Soluble NKG2D ligands and TGFβ are known mechanisms for down-regulating NKG2D expression [26], [27] in some cancer patients, but were not involved in KS patients. Chronic expression of NKG2D ligands on tumor tissues also induces the down-regulation of NKG2D [68]. That MICA was strongly expressed in situ within KS lesions may sustain the hypothesis that it is an efficient mechanism to repress antitumor activity by inducing NKG2D down-modulation on intra-tumoral NK cells. However, it does not easily explain why NKG2D was reduced on circulating NK cells. Notably, we observed that PGE2 was able to decrease NKG2D expression on NK cells in vitro, and PGE2 levels in patient sera negatively correlated with NKG2D expression on NK cells.
PGE2 is a major inhibitory factor produced by tumor cells or their surrounding microenvironment [69]. The rate-limiting enzyme in PGE2 synthesis is COX-2, which is over-expressed in many cancers, leading to an over-production of PGE2 often linked to an adverse clinical outcome [70], [71]. COX-2 and PGE2 play crucial roles in the establishment and maintenance of HHV8 latency program [32]. In addition, HHV8-induced PGE2 regulates VEGF, which controls cell growth, adhesion, angiogenesis, proliferation and differentiation. COX-2/PGE2 expression is induced during the early stages of infection of primary HMEC cells [32] and in latently-infected human umbilical vein endothelial cells [34], and abundant COX-2/PGE2 expression is detected in KS tissues [33], [34], [72]. We did not detect the presence of PGE2 in the supernatant of rKSHV.152-infected cells (data not shown), preventing us from reproducing the effect of synthetic PGE2 on NK cells with infected cell supernatants. It must be noted that, although these cells expressed low levels of the early lytic proteins K3 or K5, our attempts to switch latent infection into lytic cycle were always unsuccessful. In KS lesions, KSHV-infected cells show predominantly latent infection, and occasionally undergo lytic reactivation.
Interestingly, PGE2 was shown to downregulate IL-15Rγ chain on NK cells, thus suppressing IL-15-activated NK cell functions [39]. IL-15 is critical for NK cell-dependent clearance of several viral infections, in particular infections by human herpesviruses [73]. Since IL-15 up-regulates expression of NKG2D, and potentiates NKG2D signaling through Jak3-mediated phosphorylation of the NKG2D adaptor DAP10 [41], [42], it is conceivable that PGE2 may profoundly affect NKG2D-dependent NK cell activities. We found that PGE2 not only decreased NKG2D expression on resting NK cells in vitro, but also fully prevented IL-15-induced up-regulation of NKG2D, NKp30 and CD161. Thus, PGE2 overproduction by KS cells may preclude activation of NK cells during HHV8 infection, and promote a progressive drift towards hyporesponsive NK cells. We were surprised by the absence of any CD56+ NK cell within KS lesions, suggesting a defect in homing or survival of NK cells in the vicinity of tumor cells. Indeed, we found that PGE2 inhibited IL-15-induced proliferation and expression of the pro-survival protein Bcl-2 in NK cells, as previously shown in CD4 T cells [74]. Altogether, our results strongly suggest that, by inhibiting IL-15-induced proliferation, activation and NKG2D-mediated function in NK cells, PGE2 appears as a critical factor in preventing immune surveillance of KS development in HHV8-infected individuals. Our results also corroborate recent studies showing the deleterious effect of PGE2 released from mesenchymal stem cells or melanoma-derived fibroblasts on IL-2-induced NK cell activation [37], [75].
In conclusion, our study provides additional clues of the multifactorial complexity of HHV8-host interactions governing KS progression. In addition to previously reported alterations of HHV8-specific CD8 T cell responses in KS patients, we now report how HHV8 stepwise modifies NK cell-mediated activities. These changes may not only affect the early control of HHV8 infection at an asymptomatic stage, but also preclude efficient prevention and immunosurveillance of KS. We also provide new evidence that HHV8 utilizes the inflammatory PGE2 to its advantage in the KS microenvironment, not only for maintaining latency as previously reported, but also for inhibiting NK cell activation, function and survival in response to proinflammatory cytokines.
These results strongly support the potential for COX-2/PGE2 inhibitors in treating KS, as they could simultaneously control latency gene expression and chronic inflammation, reduce angiogenesis and cell adhesion, promote NK cell survival and restore IL-15-induced priming of NKG2D-mediated cytotoxicity.
The study was performed in accordance with the Declaration of Helsinki and French legislation, and received approval of the Saint-Louis Hospital Ethical Committee (P040105). All participants provided written informed consent.
The HHV8-infected group consisted of 70 individuals (mean age 56 years), including 25 asymptomatic HHV8 carriers (HHV8+ KS−) and 45 patients with a history of KS (HHV8+ KS+). Because HHV8 infection frequently occurs in the context of HIV infection, patients were sub-classified as follows: HIV− HHV8+ KS− asymptomatic carriers (n = 10, recruited from a cohort of ketosis-prone type 2 diabetes patients [76]); HIV+ HHV8+ KS− asymptomatic carriers (n = 15); HIV− HHV8+ KS+ patients (n = 31 classical KS, including 10 active KS and 21 resolved KS); HIV+ HHV8+ KS+ patients (n = 14 AIDS-related KS, all with resolved KS following antiretroviral therapy). The HHV8-negative control group consisted of 45 age-matched subjects, including 38 healthy blood donor volunteers (HIV− HHV8−), and 7 HIV+ individuals (HIV+ HHV8−). All HIV+ subjects were on stable antiretroviral therapy and had undetectable HIV load for at least 1 year before study. In addition, they were matched for age, CD4 T cell count at time of study, CD4 nadir, duration of disease and duration of ART in the different subgroups. Determination of IgG antibodies against latent and lytic HHV8 antigens was performed by indirect immunofluorescence. Cytomegalovirus (CMV)-specific IgG were detected by ELISA.
Blood samples were processed within 2 h of collection and PBMCs were separated by Lymphoprep gradient centrifugation (Abcys). When indicated, NK cells were freshly purified from PBMCs by negative selection using magnetic microbead separation (StemCell Technologies) with purity higher than 95%. Cells were incubated for 20 min at 4°C with combinations of the following antibodies: FITC-conjugated anti-CD3; PE-conjugated anti-CD56, anti-NKp30, anti-CD158e1e2 (KIR3DL1/S1), anti-CD158i (KIR2DS4); APC-conjugated anti-CD56, anti-CD158ah (KIR2DL1/S1), anti-CD158bbj (KIR2DL2/L3/S2), PE-Cy7-conjugated anti-CD56, Pacific Blue-conjugated anti-CD3 (all from Beckman Coulter); PE-conjugated anti-CD3; PercP-conjugated anti-CD3, APC-conjugated anti-NKp46, anti-CD161; FITC-conjugated anti-CD94, anti-DNAM-1 (BD Pharmingen); PE-conjugated anti-NKG2D (eBioscience); FITC-conjugated anti-CD122 (IL-15Rβ) APC-conjugated anti-NKG2A, anti-NKG2C, anti-CD132 (IL-15RγR&D Systems). For intracellular detection of Ki67 and Bcl-2, cells were fixed in 1% formaldehyde, permeabilized with 0.2% saponin and stained with FITC-conjugated anti-Bcl-2 and PE-conjugated anti-Ki67 (BD Pharmingen). Cells were analyzed on FACSCalibur or LSRFortessa (BD Biosciences), collecting a total of 100,000 events in a live gate. Data were analyzed using FlowJo software.
SV40-immortalized human microvascular endothelial cells (HMEC) were infected with rKSHV.152, a recombinant virus expressing the green fluorescent protein (GFP) and neo (conferring resistance to G418), and able to establish cells containing only latent HHV8 [19]. HMEC and HHV8-latently infected HMEC cells (thereafter called HHV8-HMEC) were cultured in MCDB131 medium supplemented with 10 ng/ml epidermal growth factor (EGF) and 1 µg/ml hydrocortisone (Sigma), 10% fetal calf serum (FCS), 2% glutamine, penicillin (100 U/ml) and streptomycin (100 U/ml). In addition, HHV8-HMEC medium contained G418 at 700 µg/ml.
SV2G is an SV40-immortalized HIV-negative KS-derived cell line. Because HHV8 genome was lost early after the first 2 passages, SV2G was also infected in vitro with rKSHV.152. The resulting HHV8-SV2G cells line also showed predominantly latent infection. SV2G cells were cultured in RPMI 1640-10% FCS, while HHV8-SV2G cells were cultured in 50% HMEC medium/50% SV2G medium. Cells were seeded at subconfluent density, and were recovered by trypsine/EDTA treatment. Culture supernatants were collected and kept frozen. Cell viability (ViaProbe, Pharmingen) and phenotype were analyzed by flow cytometry after cell surface staining with antibodies specific for HLA class I (W6/32), MICA and MICB [77]; ULBP-1, ULBP-2, ULBP-3, PVR, LLT1 (all from R&D System), Nectin-2 (BD Pharmingen), HLA-E (clone 3D12, eBioscience) and ICAM-1 (AbD Serotec). Expression of NCR ligands was investigated using NKp30-Fc and NKp46-Fc fusion proteins (R&D) and FITC-anti-human Fc antibody (Jackson Immunoresearch). These reagents were validated for staining NKp30 and NKp46 ligands on K562 and Hela cell lines, previously reported to express these ligands [25], [78], [79].
Purified NK cells (105 per U-bottom well) were incubated with target cells at 1∶1 effector:target ratio for 6 hr. FITC-conjugated anti-CD107a (20 µg/mL, BD Biosciences) was added directly. After 1 hour at 37°C in 5% CO2, brefeldin A (1 µg/ml) and monensin (6 µg/ml, Sigma) were added for additional 5 hr, and cells were stained with CD56-APC and CD3-PE antibodies, and Viaprobe. Where indicated, NK cells were preincubated with NKG2D blocking antibody (20 µg/ml, Coulter Immunotech) or isotype control. For intracellular IFNγ analysis, cells were fixed following staining with anti-CD3 and anti-CD56, permeabilized with 0.2% saponin and stained with IFNγ FITC antibody (BD) for an additional 30 min.
Recombinant TGFβ, VEGF, IL-8, and IL-15 were purchased from R&D Systems. PGE2 (Cayman Chemicals) was dissolved at 100 mg/ml in 95% ethanol and further diluted with RPMI 1640. The final concentration of ethanol had no effect on NK cell viability and function. ELISA was used to quantify serum IL-8, IL-10, VEGF and TGFβ (R&D Systems), PGE2 (Cayman Chemicals and soluble MICA ([80]).
Total RNA was extracted from HHV8-infected or uninfected HMEC and KS-derived cells (RNeasy system; Qiagen) and retrotranscribed to cDNA with the use of Superscript III reverse transcriptase and random primers (Invitrogen). For real-time quantitative polymerase chain reaction, Light Cycler 480 SyBR Green I Master and Light Cycler 480 detection system (Roche) were used. The level of K3 and K5 amplified transcripts was determined using a 25-fold dilution of each cDNA, and normalized to glyceraldehyde-3-phosphate dehydrogenase (GAPDH) mRNA levels. Primers used for quantifying the expression of K3 and K5 mRNAs were as follows: 5′-gCAAACCCTgTggAAggATA-3′ (forward) and 5′-AAgCTgCAgggTACAAggAA-3′ (reverse) for K3; 5′-ACCACCACAgACATCAgCAA3′ (forward) and 5′-gTAgggAAgAggTggggAAC-3′ (reverse) for K5.
Paraffin-embedded KS biopsy samples from 5 patients were obtained from the Pathology Department. After antigen retrieval in 10 mM citrate buffer pH 6.0 at 100°C, sections were blocked with hydrogen peroxide and PBS containing 10% pooled human AB serum, 1∶30 goat serum and incubated at 20°C for 1 hour with antibodies directed to HLA-1 (W6/32), MICA/B (SR99 [77]) and CD56 (clone 1B6, MBL), or control isotypes, and staining was visualized by using the Envision+ AEC System (Dako). For detection of NCR ligands, sections were incubated with NKp30-Fc or NKp46-Fc fusion proteins (R&D, 8 µg/ml in PBS containing 0.3% normal goat serum), followed by biotin-goat anti-human Fcγ (1∶2,000; Jackson Immunoresearch).
All statistical tests were performed with Instat 3 (GraphPad software). Comparisons between two groups were performed using the Wilcoxon and the Mann-Whitney t tests for paired and unpaired groups, respectively. Multiple comparison analyses were performed using Kruskall-Wallis test (non parametric ANOVA) with Dunn's multiple comparison test. Two-sided p values less than 0.05 were considered significant. Correlation analysis was performed using non parametric Spearman rank correlation.
Name: NKG2D, Accession number (Swissprot): P26718, Entry name: NKG2D_HUMAN.
|
10.1371/journal.pgen.1008226 | Profound analgesia is associated with a truncated peptide resulting from tissue specific alternative splicing of DRG CA8-204 regulated by an exon-level cis-eQTL | Carbonic anhydrase-8 (CA8) is an intracellular protein that functions as an allosteric inhibitor of inositol trisphosphate receptor-1 (ITPR1) critical to intracellular Ca++ release, synaptic functions and neuronal excitability. We showed previously that murine nociception and analgesic responses are regulated by the expression of this gene in dorsal root ganglion (DRG) associated with a cis-eQTL. In this report, we identify an exon-level cis-eQTL (rs6471859) that regulates human DRG CA8 alternative splicing, producing a truncated 1,697bp transcript (e.g., CA8-204). Our functional genomic studies show the “G” allele at rs6471859 produces a cryptic 3’UTR splice site regulating expression of CA8-204. We developed constructs to study the expression and function of the naturally occurring CA8-204G transcript (G allele at rs6471859), CA8-204C (C allele at rs6471859 reversion mutation) and CA8-201 (full length transcript). CA8-204G transcript expression occurred predominantly in non-neuronal cells (HEK293), while CA8-204C expression was restricted to neuronal derived cells (NBL) in vitro. CA8-204G produced a stable truncated transcript in HEK293 cells that was barely detectable in NBL cells. We also show CA8-204 produces a stable peptide that inhibits pITPR1 and Ca++ release in HEK293 cells. These results imply homozygous G/G individuals at rs6471859, which are common in the general population, produce exclusively CA8-204G that is barely detectable in neuronal cells. CA8 null mutations that greatly impact neuronal functions are associated with severe forms of spinal cerebellar ataxia, and our data suggest G/G homozygotes should display a similar phenotype. To address this question, we show in vivo using AAV8-FLAG-CA8-204G and AAV8-V5-CA8-201 gene transfer delivered via intra-neural sciatic nerve injection (SN), that these viral constructs are able to transduce DRG cells and produce similar analgesic and anti-hyperalgesic responses to inflammatory pain. Immunohistochemistry (IHC) examinations of DRG tissues further show CA8-204G peptide is expressed in advillin expressing neuronal cells, but to a lesser extent compared to glial cells. These findings explain why G/G homozygotes that exclusively produce this truncated functional peptide in DRG evade a severe phenotype. These genomic studies significantly advance the literature regarding structure-function studies on CA8-ITPR1 critical to calcium signaling pathways, synaptic functioning, neuronal excitability and analgesic responses.
| Carbonic anhydrase-8 (CA8) inhibits IP3 binding to the inositol trisphosphate receptor-1(ITPR1), which regulates intracellular calcium signaling critical to neuronal functions. Recessive CA8-null mutants are associated with spinocerebellar ataxia and neurodegenerative disorders. We have previously demonstrated that nociception and analgesic responses are associated with a DRG cis-eQTL that regulates murine expression of this gene. This study focuses on a human DRG exon-level cis-eQTL (rs6471859) that regulates CA8-204 alternative splicing producing a truncated 1,697bp transcript. Herein, we demonstrate the “G” allele at rs6471859 produces a cryptic 3’UTR splice site regulating tissue-specific CA8-204 expression. In vitro studies show the “G” allele (CA8-204G) produces a stable peptide that inhibits ITPR1 activation and Ca++ release in non-neuronal cells, but not in neuronal cells. However, using AAV8 gene transfer in vivo we show CA8-204G peptide is expressed in both glial and to a lesser extent in neuronal cells, producing profound analgesia and anti-hyperalgesia using inflammatory pain models, similar to the full length CA8-201 positive control. These data significantly extend our understanding of CA8 structure-function, demonstrating the truncated peptide may represent a novel therapeutic candidate.
| Car8 and its human ortholog (CA8) are enzymatically inactive isoforms of carbonic anhydrase that inhibit activation of neuronal inositol 1,4,5-trisphosphate receptor type-1 (ITPR1) by phosphorylation (pITPR1), and intracellular calcium release in mice. ITPR1 converts inositol trisphosphate (IP3) signaling to intracellular calcium signaling [1], essential to the regulation of neuronal excitability, synaptic morphology [2]. Previously, Hirota et al., (2003) determined essentially all 8 exons of Car8 (amino acids 45–291 of Car8-201) are minimally required for the direct binding to the modulatory domain of ITPR1 (amino acids 1,387–1,647) [3]. Furthermore, they showed Car8 acts as an allosteric inhibitor of IP3 binding to ITPR1, thereby inhibiting calcium signaling. Humans with a CA8 null mutation occurring in exon three (S100P), demonstrate a severe form of spinal cerebellar ataxia associated with quadrupedal locomotion and mental retardation; further defining a critical role for CA8 in neuronal development and functioning [4].
Previously, we reported Car8/CA8 and its role in calcium signaling are critical to nociception, inflammatory and neuropathic pain [5, 6]. Additionally, we reported Car8/CA8 gene therapy was analgesic and anti-hyperalgesic in animal models of pain [5, 6]. In order to quantify analgesic responses observed in animal models after Car8/CA8 gene therapy, we used allometric conversions indicating profound analgesia that exceeded an oral dose of 250 mg of morphine in an average sized adult [7]. Using integrative genomics to understand the impact of Car8 genetic regulation on nociceptive responses we identified a murine gene-level cis-expression quantitative trait (eQTL) in DRG associated with variable Car8 gene expression. This Car8 DRG cis-eQTL was shown to regulate nociception and analgesic responses [8].
In this report, we identify a 3’UTR single nucleotide polymorphism (SNP)(rs6471859) that functions as an exon-level cis-eQTL in human DRG. Through our functional genomic studies, we found the “G” allele to be associated with probe PSR0801622.hg1 (P = 0.0095) exclusively recognizing a truncated 1,697 bp transcript (e.g., CA8-204). CA8-204 comprises coding sequence for exons 1–3 with a retained intron with a self-contained polyadenylation signal and coding for an extended exon 3 associated with a new stop. It is believed that CA8-204 is non-coding in current public databases. Based on the work of Hirota et al., (2003) this truncated peptide is not expected to bind ITPR1 directly, impact IP3 binding, nor influence calcium signaling [3]. We also report that this eQTL directs tissue-specific alternative splicing, where CA8-204 is predominantly expressed in non-neuronal cells in vitro. Additionally, 1000 Genomes Project Phase 3 data show G/G homozygous individuals at rs6471859 are prevalent in the general population, with their frequencies differing by ethnic background. While the G allele is on average more prevalent in populations of East Asian ancestry (G: 75%; G/G homozygotes 56%); and less prevalent on average in populations of African ancestry (G: 15%; G/G homozygotes 2%); there is a near equal average prevalence in populations of European ancestry (G: 58%; G/G homozygotes 34%). Thus, G/G homozygotes are common in the general population and are expected to exclusively express the no-coding CA8-204 transcript predominantly in non-neuronal DRG cells. Given CA8 deficiency is associated with severe spinal cerebellar ataxia and that we have reported CA8 is critical to nociception, inflammatory and neuropathic pain [5, 6], neuronal DRG wildtype CA8 (transcript CA8-201) deficiency associated with G/G genotype should display a deleterious phenotype. Therefore, it was imperative that we clarify for the literature the significance of this cis-eQTL on DRG CA8 tissue-specific alternative splicing, and the role of CA8-204 in calcium signaling, nociception and analgesic responses.
Herein using QPCR, we show that rs6471859 regulates tissue-specific alternative splicing of CA8-204. Specifically, in vitro we found that the ‘G’ allele produces a cryptic splice site regulating expression of this transcript primarily in non-neuronal HEK293 cells. We also determined that CA8-204 codes for a truncated protein product, which inhibits steady state pITPR1, much like wild type CA8 (CA8-201). Similarly, like CA8-201, this truncated CA8-204 peptide also negatively regulates intracellular calcium release in vitro. Finally, using gene transfer containing AAV8-CA8-204 viral particles in vivo we show this peptide is expressed in both neuronal and glial cells within transduced DRG producing analgesia and anti-hyperalgesia. Collectively, these studies explain why G/G homozygotes at rs6471859 that exclusively produce this truncated peptide in DRG evade a severe phenotype. These genomic studies significantly advance the literature regarding structure-function studies on CA8-ITPR1 critical to calcium signaling pathways, synaptic functioning, neuronal excitability and analgesic responses.
Table 1 shows the association between rs6471859 with exon-level probe PSR08016202.hg1 (P = 0.0095) regulating expression of alternative transcript CA8-204. CA8-204 is comprised of the first three exons of CA8-201 with an extended exon 3 due to a retained intron containing a new stop (Fig 1). These data suggest that rs6471859 represents an exon-level cis-eQTL that resides at nucleotide position 1,416 within the retained intron of the CA8-204 transcript regulating expression of this alternatively spliced transcript in DRG. The exact mechanism by which rs6471859 regulates the alternative splicing of CA8-204 remains unknown.
To better understand how rs6471859 regulates alternative splicing of CA8-204 we developed two constructs (CA8-204G or CA8-204C) and transfected HEK293 (renal-derived, non-neuronal) and NBL (neuroblastoma, neuronal-derived) cell lines with pCMV-N-FLAG-tagged vectors coding for the naturally occurring CA8-204G transcript or CA8-204C produced by reversion mutation of CA8-204. We then ran qPCR (4 replicates from at least 3 separate cultures) to quantify the production of these transcripts. QPCR of the cDNA from CA8-204G and CA8-204C showed selective expression in HEK293 and NBL cells, respectively (Fig 2). Results from qPCR revealed enriched or unique expression of CA8-204G in non-neuronal cells (HEK293) compared with low levels of exogenous expression of CA8-204C. The reverse results occurred in NBL cells, where CA8-204C expression (reversion mutation as a control to test for the lack of alternatively spliced transcript) was greatly enriched over CA8-204G expression (Fig 2).
To determine whether CA8-204 produced a stable peptide, we transfected HEK293 cells with CA8-204G, CA8-201 or empty vector and collected cell lysates for immunoblotting with anti-FLAG or anti-V5 antibodies, normalizing these data to ß–tubulin. Results are shown in (Fig 3) suggested that like CA8-201, CA8-204G produced a peptide of the expected size of 153 amino acids (26-27kDa). However, this protein was not detected using similar immunoblotting methods with NBL cell lysates. IHC analyses in vitro using HEK293 and NBL cells, shows no detectable CA8-204G found in NBL cells (S1 Fig).
We then investigated the role of CA8-204G in the regulation of ITPR1 activation (pITPR1) by measuring forskolin-induced phosphorylation in transfected HEK293 cells followed by western blotting. Our results demonstrate CA8-204G, similar to CA8-201, inhibits the forskolin-induced phosphorylation in vitro (Fig 4). In contrast, CA8-201, but not CA8-204G inhibited phosphorylation in forskolin-induced NBL cells (S2 Fig).
We next investigated whether CA8-204G affects ATP-stimulated ITPR1-dependent cytosolic calcium release. Once again, HEK293 and NBL cells were harvested after transfection with CA8-204G, CA8-201 or empty vector. ATP (1μM) was used to induce free calcium release in vitro, and cells were monitored in real time by measuring calcium release (Fura2-AM dye). Empty vector had no effect on ATP-mediated free calcium release. HEK293 cells transfected with either CA8-201 or CA8-204G demonstrated reduced cytoplasmic free calcium after ATP-stimulation (Fig 5A). In contrast, only CA8-201 inhibited ATP-stimulated calcium release in NBL cells (Fig 5B). Collectively, these results demonstrate tissue-specific transcription, translation and functions of CA8-204G on calcium signaling in vitro is restricted to non-neuronal cells.
Due to the tissue-specific transcription, translation and functions of CA8-204G in non-neuronal cells, we investigated the impact of this alternative transcript using a pain model in vivo. AAV-mediated gene transfer to DRG cells including glial and sensory neurons through SN injections is an effective route in the treatment of chronic pain [5, 6]. We chose AAV8 in this study because it has been shown to transduce both neuronal and glial DRG cells after peripheral nerve injection. SN in wildtype C57BL/6J mice (N = 8, each group) was carried out using AAV8-V5-CA8-201 (positive control), AAV8-V5-CA8-201MT (negative control) and AAV8-FLAG-CA8-204G. Thermal behaviors were monitored daily and maximum latencies were recorded through day 15 for mice receiving SN injections. Mice receiving CA8-201 or CA8-204G, but not those injected with CA8-201MT, demonstrated elevations in their thermal latencies starting at Day 13 reaching a maximum by Day 15 (two-way repeated measures ANOVA test, post-hoc test, overall P-value <0.001; CA8-201 vs. CA8-201MT, P-value<0.001; CA8-201MT vs. CA8-204G, P-value<0.001) of approximately 3.0 seconds above baseline. This increase in thermal withdrawal latencies in C57BL/6J mice exceeds a dose of 80 mg of oral morphine in a 60 kg adult, after allometric conversion [7]. Following carrageenan injection thermal paw withdrawal latencies reached a minimum on Day 16. Mice injected previously with CA8-201, CA8-204G, but not CA8-201MT showed rapid recovery through Day 22 (Fig 6) (P-values<0.001) and once again demonstrating profound analgesia with significant increases over baseline in thermal withdrawal latencies after Day 20 (CA8-201MT vs. CA8-201, P-value<0.001; CA8-201MT vs. CA8-204G, P-value<0.001).
Lumbar DRG sections (L4 and L5) were collected from each group (N = 8) after perfusion and immunohistochemistry (IHC) was performed using primary anti-FLAG and -V5 antibodies, with fluorescent (Alexa Fluor 488; green) secondary antibody. Because DRG are comprised of both neurons and glial cells, to measure the overexpression of CA8 peptides after SN, we used anti-S100 (glia) and anti-advillin (sensory neurons) antibodies with fluorescent secondary antibody (Alexa Fluor 594; red) to define the tissue-specific expression of each CA8 peptide. Total cells were counted using DAPI staining. The percentage of overexpression in neuronal and glial cells was measured separately for V5-CA8-201MT, V5-CA8-201 and FLAG-CA8-204G (Fig 7A, 7B, 7C and 7D, S3 Fig). While overexpression calculated for total population of cells (neurons and glia) in CA8-201 and CA8-204G showed a very similar pattern (Fig 7(i)), surprisingly, and in contrast to our in vitro data, significant CA8-204G, like CA8-201, was present in both neuronal and glial cells. However, while CA8-204G expression was not significantly different from CA8-201 in glial cells, expression was reduced (about 50%) when compared with CA8-201 in neuronal cells (Fig 7(ii)). Nonetheless, this level of neuronal expression in vivo appears to be enough to produce analgesia and anti-hyperalgesia in this animal model. Thus, moderate levels of the functional CA8-204 peptide appear to be protective in vivo. Thus, the “all” or “none” expression based on cell type in vitro was not replicated in vivo. Presumably, even this moderate level of expression of CA8-204 in neuronal cells explains how G/G homozygotes evade a deleterious phenotype.
We performed a bioinformatic analysis using Human Splicing Factor (HSF 3.1) web server to understand the role of rs6471859 as a cis-acting regulatory element within the intronic or exonic sequences of CA8-204 [9]. Human Splicing finder or HSF3.1 (http://www.umd.be/HSF3/HSF.shtml) predicts splice sites (donor or acceptor) by calculating maximum entropy for the target nucleotide sequence. Consensus values range from 0 to 100 for HSF (Human Splicing Factor). The threshold is defined at 65 for HSF, which means that every signal with a score above this threshold is considered to be a splice site (donor or acceptor). The webserver also predicts the nature of the splice site.
The program was run for sequence CA8-204G including the naturally occurring “G” allele at position 1,416 in the 3’UTR; and on CA8-204C (reversion mutation) by replacing the “G” with a “C” nucleotide at rs6471859. Table 2 provides results of HSF3.1 analyses. The “G” allele at rs6471859 is predicted to create a donor site within the sequence “CAGGTCTTG” with CAG as potential new splice site (71.92% consensus). This splice site is lost with the “C” allele (CACGTCTTG) at rs6471859. These findings suggest a potential mechanism associated with the observed alternative splicing of the CA8 gene.
In this report we identify the following for the first time: (1) a human DRG exon-level cis-eQTL (rs6471859) that regulates alternative splicing of CA8; (2) a cryptic splice site is produced by the “G” allele at position 1,416 (rs6471859) within the 3’UTR of the CA8-204 transcript; (3) CA8-204G (naturally occurring transcript) and CA8-204C (reversion mutation) expression constructs demonstrate tissue-specific splicing in vitro; where CA8-204G is transcribed predominantly in non-neuronal cells (HEK293), CA8-204C is transcribed predominantly in neuronal-derived cells (NBL); (4) Surprisingly, we also show that CA8-204G is translated into a stable truncated peptide in HEK293 cells; (5) and in contrast to the work of Hirota et al., (2003), CA8-204 inhibits forskolin-induced ITPR1 activation (pITPR1) and ATP-stimulated calcium release [3]. Collectively, our in vitro data demonstrate tissue-specific transcription; translation and functions of CA8-204G are restricted predominantly to non-neuronal cells. Therefore, it appears that homozygous G/G individuals exclusively produce CA8-204, predominantly in non-neuronal DRG cells. These results raise numerous important questions because according to 1000 Genomes Project Phase 3 allele frequencies for rs6471859; G/G homozygotes are common across in almost all populations. To further address the presumed discordance between these in vitro results with what is known about the deleterious effects of CA8 deficiency, we show: (6) AAV8-mediated gene transfer of CA8-204 via direct sciatic nerve injection to transduce DRG in vivo showing CA8-204 overexpression of CA8-204, like CA8-201, produces profound analgesia after about 13 days; (7) overexpression of CA8-201 and CA8-204G showed very similar patterns in DRG neuronal and glial cells (Fig 7 (i)); and (8) surprisingly, while CA8-204G expression was seen in neuronal cells, but to a lesser extent (about 50% compared to CA8-201)(Fig 7(ii)). Importantly, this level of neuronal expression in vivo appears to be enough to produce anti-hyperalgesia and profound analgesia in the inflammatory pain model tested. Thus, it is critical to appreciate that despite its moderate levels, the functional CA8-204 peptide in somatosensory cells appears to be protective in vivo. This is consistent with the described CA8-S100P null mutation that only produces a phenotype in recessive homozygotes [4].
GTEX data reported for rs6471859 show it represents a multi-tissue eQTL including both neuronal and non-neuronal tissues as follows: lung (4.0e-23); whole blood (8.8e-6); brain-hippocampus (2.7e-4), -cortex (3.5e-3), -cerebellum (2.3e-3), and -pituitary (3.2e-7). However, no splice eQTL data are reported (https://gtexportal.org/home/snp/rs6471859). It is also important to note that our in vitro data could have been misleading if we wrongly concluded that there is an “all” or “none” expression of CA8-204 based eQTL data by cell type. Only the in vivo pain model data presented enabled us to explain the discordance between the population genetic data for rs6471859 and the lack of a severe phenotype in G/G homozygotes due to the splice variants reported. Dorsal root ganglia (DRG) comprises both of neuronal and neuroglia population of cells. Markers such as S100 and advillin were used to identify glial and neuronal cell population. Glial cells in form of satellite glial cells and sensory ganglia surely regulate the afferent signaling and neuronal excitability, with a possible role in chronic pain formation as shown in literature [10–12]. Again, the moderate level of expression of CA8-204 in neuronal cells in vivo explains how G/G homozygotes evade a deleterious phenotype. Our overall findings are consistent with reports suggesting that CA8 null mutations producing severe spinal cerebellar ataxia are recessive, and heterozygotes are phenotypically normal. Thus, heterozygotes producing approximately 50% of normal cellular CA8 evade a severe phenotype [4]. Surprisingly, CA8-204 transcript is stably expressed in a tissue-specific manner in vitro. This may likely occur because of this retained intron harboring a polyadenylation signal. Retained introns with self-contained 3’UTR polyadenylation signals are prevalent in animals [13–15]. Several alternatively spliced gene products prove to be beneficial and provide potential therapeutic strategies in treatments of diseases such as cancer, spinal muscular atrophy and tauopathies leading to neurodegenerative diseases [16–19]. Alternative Polyadenylation (APA) generates protein isoforms with truncated/shortened protein products such as vascular epidermal growth factor regulator (VEGFR) and SMN (Spinal Motor Neuron 1 and 2) resulting from intron retention [20–22]. 3′ UTR length is also a critical factor towards regulating transcript stability and protein expression. In general, the longer the 3′ UTR, the greater the likelihood for sequence-specific motifs to be recognized by splicing and translational regulatory factors [21, 23]. Positioning of the intron adjacent to an alternatively spliced exon increases the likelihood that it is retained. Recent studies suggest that immortal cell lines have differential recruitment and highly altered RNA-splicing genes representing an underappreciated hallmark of tumorigenesis [24–26]. Differential recruitment of specific splicing factors may explain our tissue-specific splicing results in vitro (Fig 2).
Remarkably, our results demonstrated that CA8-204 also produces a stable functional peptide, which contradicts the earlier findings of Hirota et al. (2003), which showed that virtually the entire CA8 peptide (exons 1–8) is required for the negative regulation of IP3 binding to ITPR1 [3]. Although, we didn’t show physical binding data, it is clear from our results that CA8-204, like CA8-201 inhibits ITPR1 activation by phosphorylation and intracellular calcium release. We speculate that the difference between our findings and theirs may be attributed to the use of purified ITPR1 for the IP3-CA8 binding. It is likely that the functions of ITPR1 are highly dependent associations with Homer, TRPC1, 80 K-H, IRAG, cGKI, IRBIT, CIB1, Na+-K+-ATPase and protein 4.1N, which are shown to dynamically impact channel activation [27–34].
Our results clearly demonstrated that CA8-204 overexpression is protective of inflammatory pain in vivo, suggesting that this peptide is fully functional and may serve as a validated therapeutic candidate as a long acting local anesthetic using gene therapy.
Director of Compliance, IACUC/IBC/ESCRO, Office of Research at University of Miami, Miller School of Medicine approved the protocol for the use of animal subjects in research study (IBC #17–075). Before SN injections all mice were anesthetized using intraperitoneal injection of a cocktail of ketamine, xylazine and acepromazine (VEDCO, Saint Joseph, Mo). Perfusion was carried out only once mice were anesthetized with isoflurane. The tests for assessing neuropathic pain were performed using IACUC approved protocols and the thermal beam projected for 15 seconds taken as cut-off threshold to prevent potential injury.
CA8 exon-level DRG analyses were run essentially as described elsewhere [35]. Briefly, bilateral L4 and L5 DRG were collected from a total of 214 brain-dead human subjects following asystole after consent of family members and snap-frozen. Genomic DNA was isolated and genotyped using Illumina’s Infinium Human Omni Express Exome-8 v1.2 chip (≈ 1M probes) and analyzed using Illumina Genome Studio 2011.1, as described. RNA was isolated from the same DRG samples frozen in TRIzol reagent (Qiagen, Austin, TX). Total RNA was analyzed using Affymetrix Human Transcriptome Array 2.0 (≈ 70K gene-level probes, ≈ 900K isoform-level probes). All statistical tests were performed with age, gender, sample’s average expression and the first two eigenvectors (from principal components analysis to capture racial/ethnic differences) per chromosome as covariates. eQTL discovery was performed on CA8 utilizing 20 exon-level probes that recognize 4 unique CA8 gene transcripts with an expression intensity level above 3. Dataset for eQTL results can be found at GSE78150 (Geo Accession; NCBI).
An expression construct containing the naturally occurring CA8-204 transcript (NCBI accession: NM_001321837.1) including the G allele at nucleotide position 1416 (CA8-204G) was generated from in pCMV-Sport6 vector (Harvard Facility) and cloned using PCR with forward primer 5’-GGGGACAACTTTGTACAAAAAAGTTGGCATG GCGGACCTGAGCTTC-3’ and the reverse primer: 5’-GGGGACAACTTTGTACAAAAAAGTTGGCATGGCGGACCTGAGCTTC-3’ and cloned into pCMV-FLAG(DYKDDDDK)-N-terminal vector (Takara), with forward primer 5’-TTTGTCGACAACGCACGCCTGCTTGCAC-3’ and reverse primer 5’-TTTTGGTACCTTACAGTAATGCTGTCAAACACTTCAACAG-3’. The restriction enzymes SalI and KpnI (NEB) were used for restriction digestion on PCR products and for the pCMV-N-FLAG (Takara) vector having similar restriction sites. CA8-204C was produced using GENEART site-directed mutagenesis system (Invitrogen Life Technologies, Carlsbad, CA) and specific primers (forward: 5’-GTTGGATTCAGTCCACGTCTTGATGTTATTT-3’, reverse: 5’-AAATAACATCAAGACGTGGACTGAATCCAAC-3’) were employed to create one nucleotide substitution at position 1416 in the CA8-204 transcript. This substitution was in the CDS sequence using the forward primer: 5’-ATGCAGATAGAAGAATTTCGACACATGTCAAGGGGGCAGA-3’ and the reverse primer: 5’-TCTGCCCCCTTGACATGTGTCGAAATTCTTCTATCTGCAT-3’. The pCMV-N-FLAG constructs containing CA8-204G and CA8-204C sequence was confirmed using Sanger sequencing. Human wildtype CA8 cDNA (CA8-201) was purchased from Origene (NCBI accession: NM_004056). The plasmid was digested with enzymes BamH1 and Not1 (NEB). The purified cDNA was ligated with linearized pAAV-MCS vector using quickstep ligation (Invitrogen) and transformed. The sequence of the insert was confirmed through sequencing. The CA8-201 cDNA was mutagenized to produce the S100P mutation (CA8-201MT), as described previously [5].The transfer of CA8-204 to AAV-ITR vector through the PCR amplification of pCMV-N-FLAG–CA8-204G (CA8-204G) using primers specific for 1,697 bp long alternative variant of CA8 containing FLAG tag carrying BamHI and XhoI sites; forward primer: 5’-TTTGGATCCGCCACCATGGACTACAAGGACGACGATGA-3’ and reverse primer: 5’-TTTTCTCGAGTTACAGTAATGCTGTCAAACAC-3’. PCR fragments were restriction digested with BamHI and XhoI (NEB), purified and ligated with linearized pAAV-MCS (ITR) vector between BamHI and XhoI sites. The constructs were confirmed using Sanger sequencing. The recombinant AAV8-V5-CA8-201, AAV8-V5–CA8-201MT, AAV8-FLAG-CA8-204G viral particles were produced by the Miami Project Viral Vector Core at the University of Miami Miller School of Medicine, essentially as described previously [5]. The purified AAV particle titers were at least 1x 1013 genome copies per mL.
HEK293 and NBL (neuroblastoma, ATCC) cells, passed at least three times with 0.5% trypsin in DMEM/F12 (GIBCO) complemented with penicillin/streptomycin and 10% fetal bovine serum (FBS; GIBCO) were allowed to grow at the rate of 4 x 105 cells for 6-well plates and 2 x 105 cells for 12 well plates, for obtaining a 80–90% confluent layer after 24 hours. Briefly, all transfections were performed using LTX with plus reagent or Lipofectamine 2000 according to the manufacturer’s instructions (Invitrogen). All transfections were performed in Opti-MEM I reduced serum medium (Invitrogen) with 2 μg DNA and 6 μl Lipofectamine 2000 for 6-well plates or 0.5 μg-1 μg for 12-wells. Cells were maintained for 4–6 h (minimum of 2h) in the transfection media at 37o C/ 5% CO2 followed by replacing it with DMEM/F-12 (1:1) plus FBS and penicillin-streptomycin. A 48 h incubation of cells in these conditions is sufficient for their use in measurements of mRNA and protein expression using RNA extraction followed by RT (reverse transcriptase)- PCR and/or real-time PCR using SYBR-green PCR master mix on the Step One Plus real time PCR machine (Applied Biosystems), western blot and immunoprecipitation studies.
Total RNA was extracted from cultured HEK293 and NBL cells transfected with CA8-204G, CA8- 204C, CA8-201 and CA8-201MT expression constructs using the RNeasy RNA extraction kit (Qiagen) following the manufacturer’s protocol. Total RNA was quantified by an Epoch spectrophotometer (BIOTEK). Two-step RT-PCR was performed using the Access-quick RT-PCR system (Promega) and Faststart DNA Polymerase (Roche) according to the supplier’s protocols. RT-PCR was used to amplify exogenous CA8-204G and CA8-204C with the forward primer: 5’-ATGGACTACAAGGACGACGATG-3’ and the reverse primer: 5’-GGGCTATTTTCTGGGGTAAA-3’. For the RT-PCR, the PCR products were loaded on 1.2% agarose gel, and were visualized with ethidium bromide. No template was used as a negative control. Quantitative PCR was used to amplify exogenous CA8-204G and CA8-204C, flanking the 3’UTR region, with the forward primer: 5’-ATGGC GCTGGCCCCATGGGTT-3’ and the reverse primer: 5’-GGGC TATTTTCTGGGGTAAA-3’. ACTB (ß-actin) was used as internal control with primers forward: 5’-AAATCTGGCACCACACCTTC-3’ and reverse: 5’-CACCTTCACCGTTCCAGTTT-3’ (Sigma Aldrich, St. Louis, MO). qPCR and analysis was carried out on a Step One Plus system (Applied Biosystems, Invitrogen) using Power SYBR-green PCR master mix (Applied Biosystems). Primers (final concentration taken as 250nM) were designed across the beginning and end of the 3’UTR of the CA8-204 alternative variants with retained intron. No endogenous CA8-204G or CA8-204C were detected at baseline in either cell line. Quantitation of CA8-204G or CA8-204C was normalized to the expression of ß-actin (ACTB). Primer efficiency was determined using the melting curve analyses after qPCR.
Transfected and non-transfected HEK293 and NBL cell cultures were lysed in RIPA buffer complementing with appropriate concentrations (1μM) of proteinase and phosphatase inhibitors (Sigma). Cell lysates containing proteins were separated on 4–15% or 10% SDS polyacrylamide gels and transferred to a PVDF membrane (Biorad) using transfer buffer containing 20% methanol in 25mM Tris-HCl and 192mM glycine. For transfer of large proteins such as pITPR1 western blot, 5–6% Tris-HCl/SDS polyacrylamide gel and 10% methanol in transfer buffer were used to increase the transfer efficiency. Blocking of the membrane was done with 5% skimmed milk in Tris-buffered saline (TBS) with 0.1% tween-20 for 1 h at room temperature followed by wash with TBS containing tween-20 and incubation with primary antibodies at 4°C. The blots were then incubated for 1 h at HRP-conjugated secondary antibodies at room temperature (Santa Cruz laboratories). Pierce Super Signal substrate (Thermofisher Scientific, Rockford, IL), was used to visualize bands. Primary antibodies used were as follows: anti-CA8 (Santa Cruz, Santa Cruz, CA), anti-V5 (Invitrogen), anti-pITPR1 (Cell Signalling Tech.) and anti-ITPR1 (Cell Signaling Technology, Ser-1755), anti-FLAG (Sigma Aldrich), anti-FLAG (Aves), anti-Vinculin (Abcam) and anti-β-actin (Cell Signaling Tech.).
HEK293 and NBL (ATCC) cell lines, were passaged at least three times and were transfected with vectors vehicle (empty vector with FLAG tag), CA8-201 and CA8-204G 24 hours after the seeding cells on the 12 well plates. Cells (1x 105) were split onto 12 mm glass coverslips (Propper, Long Island, NY), previously coated with poly-lysine (12h) and laminin for 2 h. Transfections were done as described above. Cell media was replaced with Fura-2 dye from the Fluo-4 Calcium Assay Kit (Life Technologies) on the day of the assay followed by the incubation at 37°C for 30 minutes. Cells were moved to the room temperature, washed briefly twice with calcium containing buffer (buffer I; concentrations in mM, 130NaCl; 4.7 KCl; 2.3 MgSO4; 5 Glucose; 20 HEPES; 1.2 KH2PO4, Calcium Chloride; pH 7.4). Coverslips were loaded in the upright position on the microscope washed and perfused with Ca2+-free media containing EGTA (concentrations in mM: 130 NaCl; 4.7 KCl; 2.3 MgSO4; 5 Glucose; 20 HEPES; 10 EGTA; 1.2 KH2PO4, pH 7.4). Each coverslip was allow to equilibration for 5 minutes before imaging. Cells were visualized at every 2 seconds for 10 minutes while Ca2+ free buffer alone or media containing ATP were perfused onto the coverslips with cells. Individual cells were analyzed using the LAX system on Leica microscope (Leica Microsystems Inc, Buffalo, IL).
All experiments and procedures performed on animals were conducted in compliance with the guidelines from the Institutional Animal Care and Use Committee (IACUC) at University of Miami for the care and use of laboratory animals and the current guidelines for experimental pain in conscious animals, following an IACUC approved protocol [36, 37]. Male adult C57BL/6J mice, 8–12 weeks of age and weighing 25–35 g, were acquired from Jackson Laboratories (Bar Harbor, Maine, USA). Total of 5 mice were kept in home cage environment with access to food and water ad libitum. All animals were allowed to acclimatize for at least 10 days and were housed in a 12–12 h light–dark cycle in a sterile facility under controlled humidity and temperature.
Viral particles from AAV8-V5-CA8-201 (CA8-201) were used in this study as positive control group. As previously shown, analgesic and anti-hyperalgesic activity is abolished in the (S100P) null mutant AAV8-V5-CA8-201MT (CA8-201MT) [6], that was used as a negative control; and AAV8-FLAG-CA8-204G (CA8-204G) was used as the test group. SN (sciatic nerve) injections of viral particles was followed using a previously prescribed procedure [7]. We have used V5 and FLAG tags due to their small size as tags, and data that have shown these tags lack effects in neurobehavioral studies [5, 6, 38–40]. Preceding the SN exposure, 1.5 μl of 1.0E13 genome copies/ml of viral particles were injected into the sciatic nerve through a 35-gauge NanoFil needle using a NanoFil syringe (World Precision Instruments, Sarasota, FL). While the injection in all groups, the mice were anesthetized using intraperitoneal injection of a cocktail of ketamine, xylazine and acepromazine (VEDCO, Saint Joseph, Mo). The sciatic nerve injection was performed at a distance of 45 mm from the third toe. Needles were slowly removed at approximately 1 minute after sciatic nerve injection. A number of studies show that chronic and neuropathic pain is regulated by the sex differences [41, 42]. We did not observe significant sex-specific differences in CA8-204 transcript expression in our small DRG sample. In addition, GTEX data show no sex-specific differences in gene expression for CA8 across most tissues (GTEx Analysis Release V7 (dbGaP Accession phs000424.v7.p2; https://gtexportal.org/home/snp/rs6471859). Therefore, to minimize the number of animals used in these studies, we selected male C57BL/6J mice only. This represents a limitation of the current study.
Total volume of 25–30 μl of 0.1% λ-carrageenan (Sigma-Aldrich Corp., St. Louis, Missouri, USA) dissolved in saline (at 60o C) and cooled further, was subcutaneously injected in the mice through plantar surface of left hind paw.
Thermal pain behaviors (Hargreaves) were assessed as previously described [6]. Mice were distributed randomly into groups (N = 8, per group) and thermal behavioral tests were performed in a blinded manner. The baseline latencies were adjusted to 5–7 seconds with 15 seconds taken as cutoff threshold to prevent potential injury. The latencies were averaged over 5 trials, separated by a 10-min interval.
Analyses of data from immune staining of CA8-201MT, CA8-201 and CA8-204G in DRG and the percentage of overexpression of positive neurons and glial cells in L4 and L5 DRG sections from nonadjacent slices (N = 8) from each group of animals were determined as described previously [6]. Percentages were calculated from at least 4 sections from at least 8 mice in each group and were averaged to calculate marker overexpression and overlap. In order to quantify immune-reactive double-staining in DRG, markers for glial cells (S100) and neurons (advillin) were divided by the total number of positive FLAG (for CA8-204G) and V5 (for CA8-201MT and CA8-201) cells and for each occurrence. The investigator was blinded to the group and arrangements of ganglia on the slides while making a quantitative assessment. Data were summarized using mean ± SEM. Differences between groups were assessed using one-way / two-way ANOVA depending on the condition analyzed. The criterion for statistical significance was P<0.05 applying Bonferroni’s correction for multiple tests.
Quantitation of data in all the other experiments was performed using one-way ANOVA followed by Bonferroni correction / Tukey’s multiple group comparison as post-hoc test for experiments with triplicates or more, with experiments repeated 3 times or more. Quantitation of data on thermal pain behavior tests was performed using two-way ANOVA repeated measures for each time-point (day) the measurements were recorded, followed by Bonferroni’s post-hoc correction.
|
10.1371/journal.ppat.1001104 | Role of Acetyl-Phosphate in Activation of the Rrp2-RpoN-RpoS Pathway in Borrelia burgdorferi | Borrelia burgdorferi, the Lyme disease spirochete, dramatically alters its transcriptome and proteome as it cycles between the arthropod vector and mammalian host. During this enzootic cycle, a novel regulatory network, the Rrp2-RpoN-RpoS pathway (also known as the σ54–σS sigma factor cascade), plays a central role in modulating the differential expression of more than 10% of all B. burgdorferi genes, including the major virulence genes ospA and ospC. However, the mechanism(s) by which the upstream activator and response regulator Rrp2 is activated remains unclear. Here, we show that none of the histidine kinases present in the B. burgdorferi genome are required for the activation of Rrp2. Instead, we present biochemical and genetic evidence that supports the hypothesis that activation of the Rrp2-RpoN-RpoS pathway occurs via the small, high-energy, phosphoryl-donor acetyl phosphate (acetyl∼P), the intermediate of the Ack-Pta (acetate kinase-phosphate acetyltransferase) pathway that converts acetate to acetyl-CoA. Supplementation of the growth medium with acetate induced activation of the Rrp2-RpoN-RpoS pathway in a dose-dependent manner. Conversely, the overexpression of Pta virtually abolished acetate-induced activation of this pathway, suggesting that acetate works through acetyl∼P. Overexpression of Pta also greatly inhibited temperature and cell density-induced activation of RpoS and OspC, suggesting that these environmental cues affect the Rrp2-RpoN-RpoS pathway by influencing acetyl∼P. Finally, overexpression of Pta partially reduced infectivity of B. burgdorferi in mice. Taken together, these findings suggest that acetyl∼P is one of the key activating molecule for the activation of the Rrp2-RpoN-RpoS pathway and support the emerging concept that acetyl∼P can serve as a global signal in bacterial pathogenesis.
| Borrelia burgdorferi, the causative agent of Lyme disease, is maintained in nature in a complex enzootic cycle involving Ixodes ticks and mammals. A novel regulatory network, the Rrp2-RpoN-RpoS pathway, which governs differential expression of numerous genes of B. burgdorferi, is essential for this complex life cycle. In this study, we provide evidence showing that the activation of the Rrp2-RpoN-RpoS pathway is modulated, not by the predicted histidine kinase for Rrp2, but rather by acetyl phosphate (acetyl∼P), the intermediate of the Ack-Pta (acetate kinase-phosphate acetyltransferase) metabolic pathway. Based on our findings, we propose that during the enzootic cycle of B. burgdorferi, changes in environmental cues and nutrient conditions lead to an increase in the intracellular acetyl∼P pool in B. burgdorferi, which in turn modulates the activation of the Rrp2-RpoN-RpoS pathway.
| The enzootic life-cycle of Borrrelia burgdorferi is complex and typically involves transmission between an arthropod vector (Ixodes ticks) and a mammalian host (e.g., Peromyscus rodents) [1]. Accumulated evidence have shown that the alternative sigma factor RpoS plays a central role in this complex natural cycle of B. burgdorferi [2]–[8]. RpoS functions as a global regulator and governs differential expression of more than 10% of all B. burgdorferi genes, including the two major virulence genes ospA and ospC [9]–[13]. One unique feature about rpoS of B. burgdorferi is that its expression is directly controlled by the alternative second sigma factor RpoN (σ54) at a −24/−12 σ54-type promoter. Mutation within this promoter region or inactivation of rpoN that encodes the second alternative sigma factor RpoN (σ54) abolishes expression of rpoS and RpoS-dependent genes such as ospC [6], [8], [14]. This RpoN-dependent transcriptional activation appears to play a major role in modulating RpoS level in B. burgdorferi [3], [5]–[8], [14], [15]. In addition, a small RNA dsrA also has been shown to be involved in post-transcriptional regulation of RpoS [7].
RpoN(σ54)-dependent activation of transcription requires a highly conserved transcriptional activator, the so-called enhancer-binding proteins (EBPs) [16]. B. burgdorferi has a single EBP, Rrp2, a homolog of NtrC family [17], [18]. Members of NtrC family contain three putative functional domains: an N-terminal response regulator receiver domain, a central RpoN-activation domain, and a C-terminal helix-turn-helix (HTH) DNA-binding domain [19]. The central domain becomes activated upon phosphorylation at a conserved aspartic acid residue (corresponding to D52 in Rrp2) within the N-terminal receiver domain. The activated central domain then contacts the Eσ54-holoenzyme through DNA looping, hydrolyzes ATP, and promotes open promoter complex formation for transcriptional initiation. Although direct biochemical evidence remains lacking, genetic data indicates that Rrp2 is the activator for the σ54–σS cascade of B. burgdorferi. First, a single point mutation of glycine (G) residue 239 to cysteine (C) within one of the ATP-binding motifs in the central activation domain of Rrp2 abolishes expression of rpoS and RpoS-dependent genes [4], [18], [20]. Second, when a rpoS promoter-cat reporter and an inducible rrp2 gene were cloned into a surrogate E. coli system, the reporter was activated only upon induction of rrp2 [6]. Thus, Rrp2, RpoN, and RpoS appear to constitute a Rrp2-RpoN-RpoS pathway. Consistent with this notion, recent microarray analyses reveal that genes influenced by Rrp2, RpoN, or RpoS largely overlap [2]–[4], [20].
Given the importance of the Rrp2-RpoN-RpoS pathway to the infectious cycle of B. burgdorferi [3]–[5], [20], it is striking how little we know about the upstream event(s) that lead to its activation. Since Rrp2 is the upstream activator for the pathway, an understanding of the activation of Rrp2 is key to understand the mechanism of activation of this pathway. It is postulated that activation of Rrp2 is through a phosphorylation event by a cognate histidine kinase [21]–[23]. Because of the co-localization of rrp2 and hk2 in the genome (15) and because of the ability of Hk2 to phosphorylate Rrp2 in vitro [6], Hk2 is predicted to be the cognate histidine kinase for Rrp2. A recent study by Burtnick et al. [6], however, showed that an hk2 mutant remains capable of activating Rrp2 under in vitro cultivation conditions, indicating that the molecular mechanism activating the Rrp2-RpoN-RpoS pathway is more complex than previously envisioned. In addition, the contribution of Hk2 during the infectious cycle of B. burgdorferi remains unknown because the previous hk2 mutant lost an important endogenous plasmid (lp36) for mammalian infection [6].
Response regulators can be activated by factors other than their cognate histidine kinases. The best studied mechanisms are phosphorylation by non-cognate histidine kinases (a phenomenon called “cross-talk”) [24]–[28] and phosphorylation by small molecular weight high-energy donors, such as acetyl phosphate (acetyl∼P) or carbamoyl phosphate (carbamoyl∼P) [29]–[31]. While cross-talk appears to be quite rare (48), emerging evidence indicates that acetyl∼P can function in vivo as a global signal by donating its phosphoryl group to certain response regulators [32], [33]. B. burgdorferi possesses four predicted histidine kinases (Hk1, Hk2, CheA1, and CheA2) [17], [34] as well as pathways for the synthesis and degradation of both acetyl∼P and carbamoyl∼P [17]. Burtnick et al. [6] proposed that Hk2-independent activation of Rrp2 could be activated by receiving a phosphoryl group from a non-cognate histidine kinase or a small phosphorylated compound. However, this hypothesis has not been tested experimentally. In this study, we generated an hk2 mutant suitable for in vivo study and showed that Hk2 was not required for the activation of the Rrp2-RpoN-RpoS pathway under in vitro growth conditions or during murine infection. We further showed that cross-talk among two-component systems is not likely to account for Rrp2 activation. Rather, the results obtained support the hypothesis that acetyl∼P functions as an important phosphoryl donor for Rrp2, making this small molecule a key modulator of the activation of the Rrp2-RpoN-RpoS pathway in B. burgdorferi.
To study the mechanism of activation of the Rrp2-RpoN-RpoS pathway, we focused on the upstream activator Rrp2, a putative response regulator. Burtnick et al. [6] recently reported that inactivation of hk2, which encodes the putative cognate histidine kinase for Rrp2, did not affect activation of the Rrp2-RpoN-RpoS pathway when spirochetes were cultivated in vitro. However, this hk2 mutant was not phenotypically characterized in vivo [6]. Thus, we sought to generate an hk2 mutant suitable for in vivo study. A suicide vector harboring a disrupted hk2 region was transformed into the infectious B. burgdorferi strain B31-A3 (Fig. 1A) [35]. Disruption of hk2 in the transformants was confirmed by PCR (Fig. 1B) and the absence of Hk2 expression was verified by immunoblot analyses (Fig. 1C). Of note, inactivation of hk2 by the KanR cassette did not substantially affect expression of the protein encoded by the downstream gene, rrp2 (Fig. 1C). Three transformed clones were further subjected to plasmid profile analyses (data not shown). Two clones had a plasmid profile identical to that of parental wild-type B31-A3; one of these was designated hk2 and chosen for further study (Table 1).
Under in vitro growth conditions, a combination of elevated temperature and increased cell density activates the Rrp2-RpoN-RpoS pathway, leading to the production of RpoS and RpoS-controlled proteins such as OspC [2], [5], [6], [8], [18], [36]–[39]. To determine if Hk2 affects temperature and cell density-dependent activation of the Rrp2-RpoN-RpoS pathway, wild-type B. burgdorferi and isogenic hk2 mutant spirochetes were cultivated at elevated temperature (35°C) and harvested at the late-exponential stage of growth (5×107 spirochetes per ml), conditions under which the Rrp2-RpoN-RpoS pathway is known to be activated. The hk2 mutant and its parental strain expressed similar levels of RpoS and OspC (Fig. 1C). Under “non-inducing” conditions (i.e., low cell density or lower culture temperature), neither the hk2 mutant nor the parent strain expressed OspC (data not shown). Thus, consistent with studies by Burtnick et al. [6], the Rrp2-RpoN-RpoS pathway can be activated in vitro in an Hk2-independent manner.
In vitro growth conditions only partially mimic the B. burgdorferi gene expression patterns observed during tick feeding and mammalian infection. For example, spirochetes grown under elevated temperature and high cell density conditions upregulate ospC but do not downregulate ospA [2], [40]–[42]. Therefore, we next examined the phenotype of the hk2 mutant grown in mammalian host-adapted conditions by cultivating spirochetes in dialysis membrane chambers (DMCs) implanted in the peritoneal cavities of rats [2], [40]–[42]. As shown in Fig. 2, wild-type spirochetes cultivated in DMCs produced large amounts of OspC and undetectable amounts of OspA. An rpoS mutant exhibited the opposite phenotype, as previously reported [41]. In contrast, the DMC-cultivated hk2 mutant behaved much like its wild-type parent, indicating that Hk2 was not required for Rrp2 activation within this mammalian host environment.
To further determine whether Hk2 is required for murine infection, groups of C3H/HeN mice were inoculated intradermally with various doses of either wild-type B. burgdorferi B31-A3 or its isogenic hk2 mutant. As shown in Table 2, the infectivity of the hk2 mutant was similar to that of the parental strain. This result suggests that unlike Rrp2, RpoN and RpoS [3]–[5], [20], Hk2 was not required for infection of mice by B. burgdorferi.
The results described above indicate that Rrp2 could be activated by an Hk2-independent mechanism. To test the possibility that cross-talk may contribute to Rrp2 activation, we assessed the involvement of the other three B. burgdorferi histidine kinases identified to date [17]. We first constructed an hk1 mutant (hk1) in B. burgdorferi 297 using a strategy similar to that described for generating the hk2 mutant (Fig. 3A). The resulting mutant was verified using RT-PCR to test for the absence of hk1 expression and the lack of polarity on the downstream gene rrp1 (Fig. 3B). Spirochetes were cultivated at elevated temperature and harvested at the late-exponential stage of growth. Unlike the rrp2(G239C) mutant, which failed to express OspC, the hk1 mutant produced levels of OspC that were comparable to those of its wild-type parent, indicating that Hk1 is dispensable for Rrp2 activation (Fig. 3C).
It remained possible that Hk1 and Hk2 are involved in Rrp2 activation but that they may compensate for each other in a single knockout mutant. To rule out this possibility, we generated an hk1 hk2 double mutant in B. burgdorferi 297 by transforming the hk1 mutant with the suicide vector used for generating the hk2 mutant. Immunoblot analysis of the double mutant confirmed the absence of Hk2 in the hk1 hk2 mutant, and, more importantly, demonstrated that temperature and cell density-induced expression of OspC was unaffected despite the loss of both histidine kinases (Fig. 4A). These results indicate that during in vitro growth, Hk1 is not responsible for Rrp2 activation in the absence of Hk2.
In addition to Hk1 and Hk2, B. burgdorferi expresses two other histidine kinases, CheA1 and CheA2, both of which are involved in chemotaxis [43], [44]. To determine whether CheA1 or CheA2 participate in Rrp2 activation, we examined the ability of cheA1 and cheA2 mutants to produce OspC. As shown in Fig. 4B, both cheA mutants expressed normal levels of OspC, indicating that neither CheA1 nor CheA2 is required for Rrp2 activation under in vitro growth conditions.
As a putative two-component response regulator, it is predicted that Rrp2 becomes activated upon phosphorylation of a conserved aspartate residue (D52) located within its N-terminal receiver domain [6], [18] (Fig. 5A). Since deletion of each histidine kinase gene exerted no effect on the activation of the Rrp2-RpoN-RpoS pathway, we asked whether Rrp2 activation actually requires phosphorylation. Repeated attempts to replace the wild-type rrp2 with a mutated allele containing a D52A mutation were unsuccessful. As an alternative strategy, we reasoned that, if phosphorylation is important for Rrp2 activation, overexpression of a wild-type N-terminal Rrp2 fragment (Rrp2-N) (phosphorylatable but not active) would interfere with phosphorylation of endogenous full-length Rrp2 and therefore affect activation of the Rrp2-RpoN-RpoS pathway. Conversely, overexpression of a non-phosphorylatable mutant version of the Rrp2 N-terminus should have no effect. Accordingly, we constructed a series of shuttle vectors that carried the wild-type allele rrp2-N or the mutant alleles rrp2-N(D52A) or rrp2-N(D52E) under control of the constitutive flaB promoter (Fig. 5A). Each constructed vector then was transformed into a non-infectious but highly transformable strain, B31 13A. The resulting transformants were verified by immunoblot analysis showing that each produced native full-length Rrp2 and the overexpressed Rrp2-N fragment (Fig. 5B). We then evaluated the ability of these transformants to express OspC. Overexpression of wild-type Rrp2-N almost completely abolished expression of ospC (Fig. 5B and 5C). These results were consistent with the expectation that the Rrp2-N fragment can successfully compete with native full-length Rrp2 for phosphorylation and, thus, interfere with Rrp2 and RpoN (σ54)-dependent transcription of rpoS [14], [15]. In contrast, cells expressing non-phosphorylatable Rrp2-N(D52A) or Rrp2-N(D52E) behaved like the vector control (Fig. 5B and 5C), as would be expected if Rrp2 activation requires phosphorylation of D52.
Given that the Rrp2-RpoN-RpoS pathway is essential for mammalian infection, we hypothesized that overexpression of Rrp2-N, but not Rrp2-N(D52A) would affect the spirochete's ability to infect mice. To test this hypothesis, we re-transformed the corresponding shuttle vectors into the infectious strain B31-A3. Positive transformants that had endogenous plasmid profiles identical to that of B31-A3 were then needle-inoculated into groups of C3H/HeN mice. As shown in Table 3, although the strain overexpressing wild-type Rrp2-N was capable of infecting mice with a high dose of inoculation (1×105 spirochetes per mouse), its infectivity was greatly reduced; only 1 out of 5 mice was infected at the dose of 1×103 spirochetes (Table 3). In contrast, overexpression of Rrp2-N(D52A) exerted no such effect. Thus, overexpression of Rrp2-N impaired the activation of the Rrp2-RpoN-RpoS pathway both in vitro and in vivo, further supporting the hypothesis that phosphorylation of Rrp2 is likely required for the activation of the Rrp2-RpoN-RpoS pathway.
Since Rrp2 activation appears to require D52, but not the B. burgdorferi histidine kinases, we reasoned that small metabolic intermediates (e.g., carbamoyl∼P or acetyl∼P) might be responsible for phosphorylation of D52. The B. burgdorferi genome is predicted to encode a single pathway that can produce carbamoyl-P, the so-called arginine fermentation or ArcA-ArcB pathway, in which the enzyme arginine deaminase (ArcA) converts arginine to citrulline, which is then converted to ornithine and carbamoyl∼P by the enzyme ornithine carbamoyltransferase (ArcB) (Fig. 6A). To assess the ability of carbamoyl∼P to influence Rrp2 activation, we used transposon mutagenesis to construct an arcA (bb0841) mutant (see Materials and Methods). The arcA mutant had no growth defect in vitro (data not shown) and produced levels of OspC similar to those of the wild-type parent strain (Fig. 6B). Moreover, wild-type spirochetes cultivated in growth medium supplemented with an excess of arginine or ornithine showed no change in OspC expression (data not shown). Collectively, these results argue that carbamoyl∼P does not donate its phosphoryl group to activate Rrp2, at least under in vitro cultivation conditions.
Acetyl∼P is the intermediate in the acetate kinase (Ack) – phosphate acetyltransferase (Pta) pathway. B. burgdorferi possesses genes predicted to encode both Ack (BB0622) and Pta (BB0589) [17] (Fig. 7A). However, the B. burgdorferi genome encodes neither an AMP-ACS pathway that converts acetate to acetyl-coA nor other known pathways that produce acetyl-CoA. It also lacks the TCA cycle which utilizes acetyl-CoA for energy production. The genome does have a mevalonate pathway (BB0683-BB0688) that requires acetyl-CoA for cell wall synthesis. Therefore, the Ack-Pta pathway appears to be the sole pathway for biosynthesis of acetyl-CoA required for cell wall synthesis
As a short-chain fatty acid, acetate can diffuse into cells under neutral or acidic conditions [32]. Then the enzyme Ack can convert acetate to acetyl∼P, which in turn is converted to acetyl-CoA by the enzyme Pta. Thus, increasing concentrations of exogenous acetate can elevate intracellular levels of acetyl∼P [32]. To assess whether acetyl∼P plays a role in Rrp2 activation, wild-type B. burgdorferi B31-A3 were cultivated in BSK-H medium supplemented with increasing concentrations of sodium acetate (NaOAc) with the final medium pH adjusted to 7.0. In order to detect the effect of acetate, cells were harvested at low density (5×106 spirochetes/ml) when activation of the Rrp2-RpoN-RpoS pathway (monitored by RpoS and OspC expression) is low [37], [38]. As shown in Fig. 7B, supplementation of NaOAc to the growth media dramatically increased the expression of OspC and RpoS in a dose-dependent fashion. This increase was not due to an elevated salt concentration (or to osmotic shock) since supplementation of the medium with as much as 150 mM NaCl did not reproduce this effect (data not shown).
To determine whether acetate-induced RpoS and OspC expression occurs via the Ack-Pta pathway, we attempted to generate ack and pta mutants but were unsuccessful. We reasoned that the Ack-Pta pathway may be indispensable for borrelial growth (see discussion). As an alternative approach, we overexpressed Pta in wild-type spirochetes. We reasoned that if acetate-induced Rrp2 activation results from accumulation of acetyl∼P, then overexpression of Pta would reduce the level of acetyl∼P and abolish the acetate effect. A shuttle vector carrying the pta gene under the control of the flaB promoter was introduced into strain B31 13A. The resulting transformants were cultivated in the presence of 15 mM NaOAc at pH 7.0 and harvested at low cell density (5×106 spirochetes/ml). As shown in Fig. 7C, overexpression of Pta dramatically reduced acetate-induced Rrp2 activation as assessed by expression of OspC. These results are consistent with the hypothesis that acetate activates Rrp2 via accumulation of acetyl∼P.
A combination of elevated culture temperature and increased cell density or lowered pH (pH 6.8–7.0) induces RpoS and OspC expression [5], [37], [38], [45], yet the underlying mechanism remains unclear. Since temperature, cell density, and pH are capable of influencing intracellular level of acetyl∼P in other organisms, such as E. coli [32], we sought to determine if overexpression of Pta also affects temperature and cell density-induced Rrp2 activation. Thus, spirochetes were cultivated at 23 or 35°C in standard BSK-H and harvested during late exponential growth (∼5×107 spirochetes/ml). Consistent with previous observation, elevated temperature and cell density induced OspC expression in wild-type spirochetes (Fig. 7D, the left panel). However, overexpression of Pta dramatically inhibited such effect (Fig. 7D, the right panel). These results suggest that the effect of environmental cues such as temperature- and cell density on RpoS and OspC expression might be through the small molecule acetyl∼P.
To determine whether overexpression of Pta would affect mammalian infection by B. burgdorferi, we re-constructed a Pta-overexpressing strain in the infectious strain B31-A3. One of the transformed clones harboring flaBp-pta had an endogenous plasmid profile identical to that of B31-A3, and was chosen for subsequent infection study. As shown in Table 3, overexpression of Pta resulted in a moderate reduction of infectivity; half of the mice (4 out of 8) were infected at the dose of 1×103 spirochetes. This result suggests that the AckA-Pta pathway contributes to mammalian infection, likely by synthesizing acetyl∼P, which can donate its phorphoryl group to Rrp2.
To determine whether Rrp2 can be directly phosphorylated by acetyl∼P, we performed an in vitro phosphorylation assay. Different amounts of purified recombinant Rrp2, Rrp2-N, Rrp2-N(D52A), or Rrp2-N(D52E) were incubated with 32P-labeled acetyl∼P in the reaction buffer at 37°C for 15 or 30 min. As shown in Fig. 7E, phosphorylated Rrp2 was readily detected in a time- and dose-dependent manner. Furthermore, phosphorylation of Rrp2 requires D52, since wild-type Rrp2-N, but not Rrp2-N(D52A) or Rrp2-N(D52E) could be phosphorylated by acetyl∼P (Fig. 7E). These results indicate that acetyl∼P can directly donate its phosphoryl group to Rrp2 in a histidine kinase-independent manner.
The discovery of the central regulatory network, the Rrp2-RpoN-RpoS pathway, was a significant advance in B. burgdorferi gene regulation. However, the dearth of knowledge regarding the mechanism underlying the activation of this pathway has been a major gap in our understanding of Borrelia host adaptation. In this study, we showed that temperature- and cell density-induced Rrp2-RpoN-RpoS activation occurs via a histidine kinase-independent mechanism. We further provided evidence suggesting the hypothesis that the high-energy metabolic intermediate acetyl∼P plays a key role in Rrp2 phosphorylation and, consequently, the activation of the Rrp2-RpoN-RpoS pathway.
In this study we first extended the recent finding by Burtnick et al. [6] that Hk2 was not essential for Rrp2 activation under in vitro cultivation conditions, by further showing that the hk2 mutant was capable of activating the Rrp2-RpoN-RpoS pathway in a mammalian host-adapted model and establishing infection in mice. The fact that the hk2 mutant remained capable of upregulation of OspC and downregulation of OspA in the DMC model (Fig. 2) indicates that this sensor kinase and its PAS sensing domain does not play a major in sensing mammalian host-specific signals for RpoS activation. We next tested the hypothesis that Hk1, the only other B. burgdorferi histidine kinase with no assigned function, could be responsible for activation of the Rrp2 pathway. We found that the hk1 and hk1 hk2 mutants exhibited normal levels of temperature-induced Rrp2-dependent OspC expression. We further found that spirochetes lacking other histidine kinases identified in the B. burgdorferi genome, the chemotaxis histidine kinases CheA1 or CheA2, also exhibited normal OspC expression. One caveat is that we have not tested cheA1 hk2 and cheA2 hk2 double mutants and thus cannot formally rule out a possible compensatory effect between Hk2 and CheA1 or CheA2.
Several groups have reported the existence of atypical response regulators in other bacteria whose activities do not require phosphorylation of their receiver domains [46]–[48]. These atypical response regulators either do not possess the conserved aspartate residue shown to function as the phosphorylation site (e.g., HP1021 and HP1043 in Helicobacter pylori) [46], or lack conserved residues for Mg++ chelation, which is essential for phosphorylation (e.g., FrzS in Myxococcus or NblR in Synechococcus) [47], [48]. However, Rrp2 retains all the conserved residues for phosphorylation (D52), Mg++ binding (D8, D9), and signal transduction (T80, F99, K102). Thus, it is unlikely that Rrp2 is an atypical response regulator. Indeed, in this study, we showed that Rrp2 can autophosphorylate using acetyl∼P as its sole phosphoryl donor. Furthermore, overexpression of the phosphorylatable receiver domain of Rrp2 (Rrp2-N), but not variants of Rrp2-N that carry the D52A or D52E mutations, interfered with endogenous Rrp2 activity. This result is consistent with the assumption that Rrp2 activation requires phosphorylation of D52. Another evidence supporting phosphorylation-dependent Rrp2 activation is our previous observation that the ATPase activity of Rrp2, an activity that is essential for its transcriptional activation function, also is dependent on phosphorylation of Rrp2 [15]. Of note, overproduction of a protein from a strong constitutive promoter (e.g., flaB) could have pleiotropic effects. An ideal approach to study the function of Rrp2 phosphorylation would be to replace the endogenous copy of rrp2 with the D52A mutant allele. Despite multiple efforts, however, we failed to generate the desired strain. This lack of success is consistent with previous reports that inactivation of rrp2 may be lethal [6], [18]. We hypothesize that phosphorylated Rrp2 may be important for cell growth. Consistent with this hypothesis, overexpression of Rrp2 exhibited a moderate growth defect (data not shown).
The finding that activation of RpoS and OspC requires phosphorylation of Rrp2 but does not require any of the four histidine kinases led us to hypothesize that the phosphoryl donor might be a high-energy central metabolic intermediate [29], [31], [32]. Indeed, bioinformatic analysis of the B. burgdorferi genome revealed one pathway capable of producing carbamoyl-P (ArcA-ArcB) and one pathway that can synthesize acetyl∼P (Ack-Pta). Loss of ArcA, which should result in the inability to synthesize carbamoyl-P, had no effect upon Rrp2-dependent expression, suggesting that carbamoyl-P does not serve as the phosphoryl donor to Rrp2.
Acetyl∼P is the intermediate of the Ack-Pta pathway. The Ack-Pta pathway functions in acetogenesis through the conversion of acetyl-CoA obtained from pyruvate into acetate; operation of this pathway in the opposite direction enables other bacteria to use acetate as a carbon source by activating acetate to acetyl-CoA, which subsequently enters the tricarboxylic acid (TCA) cycle. In some organisms, such as E. coli, the pathway is reversible and thus can function in both acetogenesis and acetate activation [32]. The relatively small genome of B. burgdorferi, an obligate parasite, does not encode any enzyme known to convert pyruvate to acetyl-CoA, nor does it encode the enzymes of the TCA cycle. Instead, B. burgdorferi performs lactogenesis, converting pyruvate to lactate [17] (Xu H. and Yang, X.F., unpublished result). As such, the main function of the Ack-Pta pathway of B. burgdorferi is likely not for converting acetyl-CoA to acetate, but for generating acetyl-CoA from acetate. This acetyl-CoA could then be used for cell wall synthesis (via the mevalonate pathway [BB0683-BB0688]) and possibly for other metabolic pathways (Fig. 7A). Furthermore, B. burgdorferi seems to lack other acetyl-CoA synthetic pathways (e.g., the AMP-ACS pathway, β-oxidation of fatty acids, and several amino acid degradation pathways). Thus, the Ack-Pta pathway appears to be the sole pathway for biosynthesis of acetyl-CoA. If so, one would predict that the Ack-Pta pathway is essential for spirochetal growth. This notion is consistent with the fact that we failed to generate either an ack or a pta mutant by either targeted mutagenesis or random transposon mutagenesis (data not shown). What's the source of acetate for B. burgdorferi? Our measurement showed that acetate concentration in mouse blood and the midgut of fed ticks is ∼1.0 M and ∼1.8 mM, respectively (Xu H. and Yang, XF, unpublished data). One of the ingredients of the BSK-H medium, CMRL, also contains 0.61 mM acetate (other ingredients of this complex medium, such as rabbit serum, also may contribute to the overall levels of acetate). Through diffusion or an unknown transport system, B. burgdorferi may obtain sufficient acetate from these environments for acetyl-CoA production.
Acetyl∼P has drawn attention as a global regulator of gene expression via its ability to donate its phosphoryl group to a subset of response regulators under certain environmental conditions [32]. In E. coli, the intracellular acetyl∼P concentration can reach levels sufficient to phosphorylate a subset of response regulators [49] and thus influence the biological processes controlled by those proteins [32]. Although we have not yet measured the intracellular acetyl∼P levels to determine if this is also the case in B. burgdorferi, we were able to provide three lines of evidence to support the conclusion that acetyl∼P plays an important role in Rrp2 activation: (i) the activation of the Rrp2-RpoN-RpoS pathway can be induced by increasing concentration of exogenous acetate (Fig. 7B); (ii) overexpression of Pta reduced acetate-induced activation of the Rrp2-RpoN-RpoS pathway (Fig. 7C); and (iii) acetyl∼P served as a phosphoryl donor to Rrp2 in vitro (Fig. 7E). Note that overexpression of Pta did not completely abolish OspC production, suggesting that a low level of Rrp2 activation still occurs. This might be due to the presence of low levels of acetyl∼P, as overexpression of Pta does not abolish the production of acetyl∼P. Alternatively, Hk2 may contribute to Rrp2 activation. We are currently in the process of testing this possibility by overexpressing Pta in the hk2 mutant. Nevertheless, this partial inhibition of RpoS and OspC expression by overexpression of Pta is consistent with the in vivo phenotype that overexpression of Pta resulted in a moderate reduction of spirochetal infectivity in mice (Table 3).
It is well established that the Rrp2 pathway can be activated by many environmental cues such as temperature, pH, cell density, oxygen, and CO2 levels [37]–[39], [45], [50], [51]. However, the underlying mechanism for these phenomena has not been elucidated. In this regard, it is striking that virtually all the environmental cues that activate the Rrp2 pathway also have been shown to influence the acetyl∼P pool in E. coli [32]. This observation is consistent with our hypothesis that acetyl∼P serves as a signaling molecule that responds to environmental cues and in response activates the Rrp2 pathway. Indeed, we showed that overexpression of pta greatly inhibited both temperature- and cell density-induced activation of Rrp2 (Fig. 7D), suggesting that elevated temperature and increased cell density activate the Rrp2-RpoN-RpoS pathway in an acetyl∼P-dependent manner. Elevated temperature may increase acetyl∼P levels by enhancing diffusion of acetate into the cells and/or from increased transport efficiency via an unidentified transport system for acetate. Elevated temperature also increases cell growth rates that likely lead to increased levels of acetyl∼P [32], [52]. The effect of increased cell density on acetyl∼P levels, on the other hand, can result simply by a change in extracellular pH. As cell density increases, the culture pH diminishes from 7.5 to 7.0 or lower [38], which favors the passive diffusion of acetate into the cells [32].
One caveat of this study is that we used expression of RpoS and OspC as the readout for Rrp2 phosphorylation. An ideal approach for such study would be directly to detect the phosphorylated form of Rrp2. Unfortunately this approach is not technically feasible since most forms of the Asp-phosphorylation are unstable and there is no antibody available for detecting Asp-phosphorylation. Thus, a common approach for studying phosphorylation of response regulators is to monitor the output product as a result of phosphorylation of a response regulator. In the case of Rrp2, the only direct target gene identified thus far is rpoS and therefore, expression of rpoS faithfully reflects the activation of Rrp2 modulated by phosphorylation. One concern for this approach is whether the effect on RpoS expression observed in this study is through another transcriptional activator, BB647 (BosR). BB647 is a fur homologue and was recently shown that inactivation of this gene significantly reduced rpoS and ospC expression [53]–[56]. Although it remains unclear how BosR fits into the Rrp2-RpoN-RpoS pathway, we found that neither overexpression of Rrp2-N nor overexpression of Pta affected the level of BosR (data not shown), suggesting that the effects of Rrp2-N or Pta overexpression on RpoS and OspC was not through BosR, rather through Rrp2.
In summary, we have shown that temperature- and cell density-induced the activation of the Rrp2-RpoN-RpoS pathway proceeds independently of histidine kinases and carbamoyl-P. In contrast, biochemical and genetic manipulation of the acetyl∼P-producing Ack-Pta pathway dramatically impacts activation of the Rrp2-RpoN-RpoS pathway, providing strong evidence that acetyl∼P plays an important role in Rrp2 activation under in vitro growth conditions. We also provide evidence showing that, during mammalian infection, the Rrp2-RpoN-RpoS pathway is also activated via an Hk2-independent mechanism and that acetyl∼P plays an important role in this process. Then, what is the function of Hk2? One possibility is that Hk2 may play a role in sensing host signals and activating Rrp2 during the process of tick feeding. In this regard, we have examined the phenotype of the hk2 mutant in ticks and found that the hk2 mutant indeed has reduced infectivity via the route of tick infestation. Unfortunately, we have not been able to construct an infectious complemented strain and, thus, have been unable to show restoration of this defect, which prevents us from drawing a definitive conclusion on Hk2 function in the enzootic cycle of B. burgdorferi. Nevertheless, this preliminary finding suggests that Hk2 may contribute to Rrp2 activation during the process of tick feeding. In addition, spirochetes likely have increased levels of intracellular acetyl∼P in feeding ticks, as they encounter increased temperature [39], as well as a massive influx of nutrients that leads to a dramatic increase of growth rates during this process [57], [58]. Thus, we postulate that while acetyl∼P plays an important in activating the Rrp2-RpoN-RpoS pathway during mammalian infection, both acetyl∼P and Hk2 are likely involved in integrating complex environmental and host signals to modulate the Rrp2-RpoN-RpoS pathway during the process of spirochetal transmission from ticks to mammals.
All animal experimentation was conducted following the NIH guidelines for housing and care of laboratory animals and performed in accordance with Indiana University Institutional regulation after review and approval by the institutional Animal Care and Use Committee at Indiana University.
Low–passage, virulent B. burgdorferi strain B31-A3 was kindly provided by Dr. P. Rosa (Rocky Mountain Laboratories, National Institute of Allergy and Infectious Diseases, National Institutes of Health) [35]. Strain B31 13A that lacks lp25 was kindly provided by Dr. F. T. Liang (Louisiana State University) [59]. The rrp2 mutant was described previously [9] [20]. The cheA1 and cheA2 mutants were kindly provided by Dr. Li (New York medical college, NY) [44]. Borreliae were cultivated in Barbour-Stoenner-Kelly (BSK-H) medium (Sigma, St. Louis, MO) supplemented with 6% normal rabbit serum (Pel Freez Biologicals, Rogers, AR) at 35°C unless indicated otherwise. A shuttle vector pBSV2 (a gift from Dr. P. Rosa) was maintained in E. coli strain TOP10. Relevant antibiotics were added to the cultures in the following final concentrations: 300 µg/ml for kanamycin and 50 ng/ml for erythromycin.
To generate an hk2 mutant in strain B31-A3, a 2.5 kb fragment containing hk2 and its surrounding region was amplified with primers hk2-delF and hk2-delR (Supplemental Table S1) and cloned into the cloning vector pCR-XL-TOPO (Invitrogen). The plasmid was digested with Hind III (19 bp downstream of the 5' end of hk2) and ClaI (637 bp upstream of the 3' end of hk2), and a kanamycin-resistance cassette driven by the flaB promoter was then inserted into the disrupted hk2 gene (Fig. 1A). The suicide vector was confirmed by sequencing, and the plasmid DNA was transformed into B. burgdorferi strain B31-A3 as previously described [9], [60]. Whole cell lysates from positive clones were analyzed by PCR and Western immunoblot analysis using a monoclonal antibody against Hk2 to confirm marker insertion and inactivation of hk2. The plasmid profiles of the hk2 mutant clones were determined by PCR analyses with twenty-one pairs of primers specific for each of the endogenous plasmids [61]–[63]. Two of the three randomly picked clones had plasmid profiles that were identical to the parental strain B31-A3 [35], and one of these was chosen for further study.
Dialysis membrane chambers (DMCs) containing 1×103 organisms diluted from a mid-logarithmic growth culture at 33°C in vitro, were implanted into the peritoneal cavities of female Sprague-Dawley rats as previously described [40], [42]. The DMCs were explanted 192 h after implantation; the spirochetes then were harvested, washed with 1x PBS buffer, and then examined by SDS-PAGE and silver staining.
To construct a suicide vector for inactivation of hk1, regions of DNA corresponding to 1.3 kb upstream and 1.3 kb downstream of hk1 regions were PCR amplified from B31-A3 genomic DNA. The resulting DNA fragments were then cloned upstream and downstream of an erythromycin-resistant marker (ermR) within the pCR-XL-TOPO cloning vector, resulting in suicide vector pXY245. The inserts of pXY245 were confirmed by sequencing. The plasmid DNA was transformed into B. burgdorferi 297 strain BbAH130 as previously described [9], [60], resulting in a mutant with 3.4 kb deletion within hk1 (except the 460 bp to the 5' end and 385 bp to the 3' end of hk1) and an insertion of the ermR marker. Loss of hk1 expression was confirmed by RT-PCR analysis.
To construct the hk1 hk2 double mutant, the suicide vector pHX-hk2-kan DNA was transformed into the hk1 mutant. Kanamycin and erythromycin-resistant clones were selected and the loss of hk2 was confirmed by Western immunoblot analysis using an anti-Hk2 monoclonal antibody.
To constitutively express the wild-type Rrp2 N-terminal domain, the DNA fragment corresponding to the Rrp2-N terminal region was PCR-amplified from B. burgdorferi B31-A3 genomic DNA using primers rrp2-N-F and rrp2-N-R (Supplemental Table S1). Two restriction sites, NdeI and PstI, were incorporated into the designated primers and used for insertion of the digested PCR fragment into the pBSV2-derived shuttle vector pJD55 [4] harboring a flaB promoter. Thus, expression of Rrp2-N was placed under the control of the flaB promoter, flaBp-Rrp2-N. The resulting shuttle vector, pJD55/rrp2-N, was verified by sequencing and then transformed into B31 13A and B31-A3.
To introduce a single amino acid substitution (D52A or D52E) into the Rrp2-N terminal domain on pJD55/rrp2-N, site-directed mutagenesis was carried out by using the QuikChange II XL Site-Directed Mutagenesis Kit (Stratagene, La Jolla, CA) with the mutagenic PAGE-purified primers D52A-F/D52A-R and D52E-F/D52E-R (Supplemental Table S1) as described by the manufacturer. Briefly, PCR was carried out as follows: 95°C for 50 seconds, 60°C for 50 seconds, 68°C for 10 minutes and 18 cycles. The resulting shuttle vectors with point mutations in Rrp2-N were verified by sequencing and designated pJD55-Rrp2-N(D52A) and pJD55-Rrp2-N(D52E), respectively.
To overexpress Pta, the DNA fragment corresponding to pta (bb0589) was PCR amplified from B. burgdorferi B31-A3 genomic DNA using primers Bb589F and Bb589R (Supplemental Table S1) and then subsequently cloned into pJD55, which places pta under the control of the flaB promoter. The resulting shuttle vector was verified by sequencing and then transformed into B31 13A and B31-A3.
The arcA mutant was generated by transposon-mediated mutagenesis as part of an on-going transposon signature tagged mutagenesis (STM) study. Briefly, twelve independent mutant libraries, each having a unique 7 bp sequence tag, were created using modified versions of the suicide plasmid pMarGentKan derived from pMarGent [64] (kindly provided by Dr. P. E. Stewart, Rocky Mountain Laboratories, National Institutes of Health, Hamilton, MN). The resulting plasmids were transformed into B. burgdorferi B31 5A18; transformants were selected on solid BSK-II media containing 200 µg/ml of kanamycin and 40 µg/ml of gentamicin as described previously [65]. Transposon insertion sites were determined by restriction digestion of the Borrelia genomic DNA, plasmid rescue in E. coli, and sequencing, as described previously [1].
SDS-PAGE and immunoblot analyses were performed as previously described [66]. Monoclonal antibodies against OspC, RpoS, and FlaB were described previously [20], [38]. Monoclonal antibodies against Rrp2 and HK2 were produced using a previously described method [66]. Rrp2-N fragments were detected using a previously reported polyclonal rat antiserum specific against full length Rrp2 [18].
Three or four week-old C3H/HeN mice (Harlan, Indianapolis, IN) were subcutaneously inoculated with spirochetes at a dose of 105 spirochetes per mouse. Ear punch biopsy and tissue samples (skin, heart, spleen and joint) were collected at the time points indicated for each experiment and cultured in BSK-H medium supplemented with 1× Borrelia antibiotic mixture (Sigma, Saint Louis, MO). A single growth-positive culture was used as the criterion for infection of each mouse. All animal protocols were approved by the Institutional Animal Care and Use Committee at Indiana University.
RNA samples were extracted from B. burgdorferi cultures using the RNeasy® mini kit (Qiagen, Valencia, CA) according to the manufacturer's protocols. Three independent culture samples were used for each strain. Digestion of contaminating genomic DNA in the RNA samples was performed using RNase-free DNase I (Promega, Madison, WI), and removal of DNA was confirmed by PCR amplification using primers specific for the B. burgdorferi flaB gene [67]. The cDNA was synthesized using the SuperScript III reverse transcriptase with random primers (Invitrogen, Carlsbad, CA). To quantify the transcript levels of ospC, an absolute quantitation method was used by creating a standard curve in qPCR assay by following the manufacture's protocol (Strategene, La Jolla, CA). Briefly, a cloning vector containing the ospC gene serves as standard template. A series of ten-fold dilution (100 to 107 copies/µl) of the standard template was prepared and qPCR was performed to generate a standard curve by plotting the initial template quantity against the Ct values for the standards. The quantity of the ospC and flaB in cDNA samples were calculated by comparing their Ct values of the Standard Curve plot. Both standards and samples were performed in triplicate on an ABI 7000 Sequence Detection System using GREEN PCR Master Mix (ABI, Pleasanton, CA). Levels of ospC transcript were reported as per 1000 copies of flaB transcripts.
Purification of recombinant Rrp2 protein was described previously [15]. The PCR fragments encoding Rrp2-N, Rrp2-N/D52A and Rrp2-N/D52E were cloned into the expression vector pGEX4t-2 with a glutathione S-transferase (GST) at the N-terminus. Fusion proteins GST-Rrp2, GST/Rrp2-N, GST/Rrp2-N/D52A and GST/Rrp2-N/D52E were expressed in E. coli under inducible condition of 1 mM IPTG at 37°C for 6 hours. Proteins were purified from cell lysates using GST SpinTrap (GE Healthcare, Piscataway, NJ) according to the manufacturer's manual.
[32P]acetyl phosphate was synthesized as described by Quon et al. [68]. Briefly, the reaction mixture includes 0.3 U E. coli acetate kinase (Sigma), 10 µCi of [32P]ATP (6000 Ci/mmol, PerkinElmer) in AKP buffer (25 mM Tris-HCl [pH 7.4], 60 mM KOAc, 10 mM MgCl2; final pH 7.6) and was incubated at room temperature for 20 min. [32P]acetyl phosphate was used either without further treatment or with further purification by filtering through a 30 kDa cut-off membrane to remove acetate kinase (Amicon ultra with 30 kDa cut-off, Millipore). [32P]acetyl phosphate was mixed with recombinant Rrp2 (2.5 µl, 0.7 or 1.4 µg), Rrp2-N (2 µg), Rrp2-N/D52A (2 µg), Rrp2-N/D52E (2 µg) for 15 min or 30 min at 37°C. The reaction was terminated by addition of SDS-PAGE loading buffer and then loaded to 12% SDS-PAGE without boiling. The gel was then exposed to a Kodak X-ray film.
|
10.1371/journal.pntd.0001727 | In-silico Investigation of Antitrypanosomal Phytochemicals from Nigerian Medicinal Plants | Human African trypanosomiasis (HAT), a parasitic protozoal disease, is caused primarily by two subspecies of Trypanosoma brucei. HAT is a re-emerging disease and currently threatens millions of people in sub-Saharan Africa. Many affected people live in remote areas with limited access to health services and, therefore, rely on traditional herbal medicines for treatment.
A molecular docking study has been carried out on phytochemical agents that have been previously isolated and characterized from Nigerian medicinal plants, either known to be used ethnopharmacologically to treat parasitic infections or known to have in-vitro antitrypanosomal activity. A total of 386 compounds from 19 species of medicinal plants were investigated using in-silico molecular docking with validated Trypanosoma brucei protein targets that were available from the Protein Data Bank (PDB): Adenosine kinase (TbAK), pteridine reductase 1 (TbPTR1), dihydrofolate reductase (TbDHFR), trypanothione reductase (TbTR), cathepsin B (TbCatB), heat shock protein 90 (TbHSP90), sterol 14α-demethylase (TbCYP51), nucleoside hydrolase (TbNH), triose phosphate isomerase (TbTIM), nucleoside 2-deoxyribosyltransferase (TbNDRT), UDP-galactose 4′ epimerase (TbUDPGE), and ornithine decarboxylase (TbODC).
This study revealed that triterpenoid and steroid ligands were largely selective for sterol 14α-demethylase; anthraquinones, xanthones, and berberine alkaloids docked strongly to pteridine reductase 1 (TbPTR1); chromenes, pyrazole and pyridine alkaloids preferred docking to triose phosphate isomerase (TbTIM); and numerous indole alkaloids showed notable docking energies with UDP-galactose 4′ epimerase (TbUDPGE). Polyphenolic compounds such as flavonoid gallates or flavonoid glycosides tended to be promiscuous docking agents, giving strong docking energies with most proteins.
This in-silico molecular docking study has identified potential biomolecular targets of phytochemical components of antitrypanosomal plants and has determined which phytochemical classes and structural manifolds likely target trypanosomal enzymes. The results could provide the framework for synthetic modification of bioactive phytochemicals, de novo synthesis of structural motifs, and lead to further phytochemical investigations.
| Traditional herbal medicine continues to play a key role in health, particularly in remote areas with limited access to “modern medicines”. Many plants are used in traditional Nigerian medicine to treat parasitic diseases. While many of these plants have shown notable activity against parasitic protozoa, in most cases the mode of activity is not known. That is, it is not known what biochemical entities are being targeted by the plant chemical constituents. In this work, we have carried out molecular docking studies of known phytochemicals from Nigerian medicinal plants used to treat human African trypanosomiasis (sleeping sickness) with known biochemical targets in the Trypanosoma brucei parasite. The goals of this study were to identify the protein targets that the medicinal plants are affecting and to discern general trends in protein target selectivity for phytochemical classes. In doing so, we have theoretically identified strongly interacting plant chemicals and their biomolecular targets. These results should lead to further research to verify the efficacy of the phytochemical agents as well as delineate possible modifications of the active compounds to increase potency or selectivity.
| Human African trypanosomiasis (HAT), also known as sleeping sickness, is caused by the single-celled kinetoplastid parasites, Trypanosoma brucei, which are transmitted to humans by infected tsetse flies (Glossina spp.). Two sub-species of T. brucei (rhodesiense and gambiense) cause the two different forms of the disease. T. b. rhodesiense is found in southern and eastern Africa while T. b. gambiense is found in the western, central and some parts of eastern Africa. T. b. gambiense now accounts for about 90% of all reported cases of sleeping sickness. A third subspecies, T. b. brucei, does not cause HAT because of its susceptibility to lysis by human apolipoprotein L1 [1].
Current chemotherapies of HAT are directed either to the early or late stages of the disease. All the clinically available HAT chemotherapeutic drugs have been noted to be ineffective, and they also have severe side-effects. The only drug candidate in clinical trials for the treatment of HAT is the nitroimidazole fexinidazole. Fexinidazole is currently in clinical study for the treatment of the late stage form of HAT [2], [3]. It is worth noting that the number of reported cases of HAT fell in the past decade, and it has also been suggested that a possible elimination of the disease might be in sight [4]. This is a very delightful development for this “neglected” tropical disease, and it is our hope that continued research into new and effective chemotherapy against HAT remains an integral part of public health initiatives in endemic communities.
Medicinal plants from Nigeria's lush rainforest, as well as her very diverse montane and savanna vegetation, continue to play a vital role in her healthcare system. For tens of millions of Nigerians, indigenous traditional medicine is the major – and sometimes the only – access to pharmacological agents [5]. There have been several published reports on the biological activity of Nigerian plants, but most of the bioactive components of those plants have not been characterized. However, the country's big and loosely-regulated traditional medicine industry continues to promote the efficacy of extracts and concoctions made from most of the plants. A number of Nigerian plants have been used traditionally in West Africa to treat protozoal infections and many of these have shown in-vitro antiprotozoal activity (Table S1).
Several T. brucei protein targets have been identified and experimentally validated [6]. In addition to validated targets, several potential targets have been predicted in silico [7]. For a recent review of phytochemical agents that show activities against parasitic protozoans and protozoan biochemical targets, see [8], [9]. Some of the potential T. brucei drug targets that we considered in this work include adenosine kinase [10], pteridine reductase 1 [11], dihydrofolate reductase [12], trypanothione reductase [13], cathepsin B [14], heat shock protein 90 [15], as well as sterol 14α-demethylase (CYP51) [16], nucleoside hydrolase [17], triose phosphate isomerase [18], nucleoside 2-deoxyribosyltransferase [19], UDP-galactose 4′ epimerase [20] and ornithine decarboxylase [21]. In this computational study, we have evaluated the interaction of compounds that were isolated from some antitrypanosomal Nigerian medicinal plants (Table S1) against potential protein drug targets in Trypanosoma brucei for which X-ray crystal structures were available from the Protein Data Bank (PDB). We strove to address the questions of which phytochemical agents might be responsible for the observed antitrypanosomal activity and what are the likely targets of those phytochemicals. In doing so, we hope to identify particular classes of phytochemical agents that can be exploited for antiparasitic chemotherapy.
Protein-ligand docking studies were carried out based on the crystal structures of rhodesain (PDB 2p7u, [22] and PDB 2p86 [23]), T. brucei adenosine kinase, TbAK (PDB 2xtb and PDB 3otx [24]), T. brucei pteridine reductase 1, TbPTR1 (PDB 3jq7 [25]), T. brucei dihydrofolate reductase, TbDHFR (PDB 3rg9 and PDB 3qfx [26]), T. brucei trypanothione reductase, TbTR (PDB 2wow, [27]), T. brucei cathepsin B, TbCatB (PDB 3hhi [28]), T. brucei heat shock protein 90, TbHSP90 (PDB 3omu [29] and PDB 3opd [30]), T. brucei sterol 14α-demethylase, TbCYP51 (PDB 3gw9 [16]), T. brucei nucleoside hydrolase, TbNH (PDB 3fz0 [31]), T. brucei triosephosphate isomerase, TbTIM (PDB 1iih, PDB 6tim [32], and PDB 4tim [33]), T. brucei nucleoside 2-deoxyribosyltransferase, TbNDRT (PDB 2a0k, PDB 2f64, and PDB 2f67 [19]), T. brucei UDP-galactose 4′-epimerase, TbUDPGE (PDB 1gy8 [20]), and T. brucei ornithine decarboxylase, TbODC (PDB 1f3t [34], PDB 1njj [35], and PDB 1qu4 [21]). All solvent molecules and the co-crystallized ligands were removed from the structures. Molecular docking calculations for all compounds with each of the proteins were undertaken using Molegro Virtual Docker v. 4.3 [36], [37], with a sphere large enough to accommodate the cavity centered on the binding sites of each protein structure in order to allow each ligand to search. If a co-crystallized inhibitor or substrate was present in the structure, then that site was chosen as the binding site. If no co-crystallized ligand was present, then suitably sized cavities were used as potential binding sites. Standard protonation states of the proteins based on neutral pH were used in the docking studies. The protein was used as a rigid model structure; no relaxation of the protein was performed. Assignments of charges on each protein were based on standard templates as part of the Molegro Virtual Docker program; no other charges were set. Each ligand structure was built using Spartan '08 for Windows [38]. The structures were geometry optimized using the MMFF force field [39]. Flexible ligand models were used in the docking and subsequent optimization scheme. As a test of docking accuracy and for docking energy comparison, co-crystallized ligands were re-docked into the protein structures. Different orientations of the ligands were searched and ranked based on their energy scores. The RMSD threshold for multiple cluster poses was set at <1.00 Å. The docking algorithm was set at maximum iterations of 1500 with a simplex evolution population size of 50 and a minimum of 30 runs for each ligand. Each binding site of oligomeric structures was searched with each ligand. The lowest-energy (strongest-docking) poses for each ligand in each protein target are summarized in Tables S2–S20.
Phytochemical studies of Acacia nilotica [40]–[45] have shown an abundance of polyphenolic compounds (Table S2), including hydrolyzable tannins, flavonoid gallates, and flavonoid glycosides. Although these polyphenolics are notorious for being promiscuous protein complexing agents and they do show relatively strong docking to all proteins investigated in this study, some selectivity can be seen. Thus, for example, 1,3-digalloylglucose showed docking selectivity for TbUDPGE, 3′,5-digalloylcatechin was selective for TbAK, and 3′,7-digalloylcatechin selectively docked with TbNH and was the strongest binding ligand for that protein (−44.2 kcal/mol). 5,7-Digalloylcatechin was the strongest binding ligand for TbPTR1 (−42.7 kcal/mol) and 4′,7-digalloylcatechin was the strongest binding ligand for TbODC (−41.4 kcal/mol). A number of these polyphenolic ligands showed strong docking interactions with TbAK, TbPTR1, TbCYP51, TbNH, and TbUDPGE, and interactions with these protein targets may be responsible for the antitrypanosomal activity of A. nilotica [46]. The docking study suggests that rhodesain, TbDHFR, TbTR, TbCatB, and TbHSP90 are not targets for A. nilotica phytochemicals.
Ageratum conyzoides extracts have been dominated by flavonoids and chromanes (Table S3) [40], [47]–[49]. 5,6-Dimethoxy-2-isopropylbenzofuran, 6,7-dimethoxy-2-methyl-2-(2-methyl-1-propanone)-3-chromene, 6-acetyl-2,2-dimethylchroman, and O-methylenececalinol exhibited selectivity for TbTIM with docking energies comparable to the co-crystallized ligand, 3-phosphoglyceric acid (−21.6 kcal/mol). The flavonoid 3′,4′,5,5′,6,8-hexamethoxyflavone, on the other hand, showed selective docking to TbPTR1 and TbUDPGE. Nour and co-workers [49] have examined the antitrypanosomal activities of several methylated flavonoids and a chromene from A. conyzoides. The flavonoids all have similar antitrypanosomal activities with IC50 values ranging from 3.0 to 6.7 µg/mL. The chromene, O-methylencedalinol, on the other hand, was much less active (IC50 = 78.4 µg/mL). The docking energies for many of the protein targets was much more negative (stronger docking) for the flavonoids than for the chromene. Thus, for example, there is good correlation between log(IC50) and docking energies of the ligands with TbPTR1 or with TbUDPGE (R2 = 0.712 and 0.751, respectively).
Compounds isolated from Annona senegalensis include annonaceous acetogenins, diterpenoids, and sesquiterpenoids, and aporphine alkaloids (Table S4) [40], [50]–[53]. The acetogenins (annogalene, annonacin, annonacin A, annosenegalin, and senegalene) are probably responsible for the antitrypanosomal activity of the plant [54], [55]. These compounds show a propensity for docking with TbAK, TbCYP51, and TbUDPGE. The acetogenins are very flexible with a great deal of conformational mobility. Nevertheless, docking with these protein targets is largely hydrophobic. Key interactions of the acetogenins with TbAK include Phe337, Gly298, Asn295, Asn67, and Gly296. Additionally, the acetogenin annogalene is one of the best binding ligands for TbDPGE (−42.9 kcal/mol).
Bridelia ferruginea has been phytochemically characterized with polyphenolic and triterpenoid constituents (Table S5) [40], [56]. The flavonoids delphinidin and ferrugin showed docking selectivity for TbPTR1. The tannin epigallocatechin(7→4′)gallocatechin showed notably strong docking with TbCYP51. Although they are relatively weak docking ligands, the triterpenoids friedelin and taraxerol docked selectively with TbUDPGE.
Limonoids are characteristic phytochemicals of the Meliaceae, including Carapa procera (Table S6) [40], [57], and numerous limonoids have exhibited antiprotozoal activities [58]–[62]. Six of the eleven C. procera limonoids showed notably strong docking with TbCYP51 (docking energies<−26 kcal/mol). A similar trend was noted for docking of Khaya limonoids (see below). Carapolides A, B, and C showed particularly strong docking with docking energies of −31.8, −29.3, and −28.5 kcal/mol, respectively; comparable to the docking energy of the co-crystallized ligand, N-[(1R)-1-(2,4-dichlorophenyl)-2-(1H-imidazol-1-yl)ethyl]-4-(5-phenyl-1,3,4-oxadiazol-2-yl)benzamide [16] (−28.6 kcal/mol), for this protein. The limonoids all dock with TbCYP51 near the heme cofactor (Fig. 1). In addition, preferential docking of individual limonoids with other protein targets include: mexicanolide with TbAK, 3β-isobutyroloxy-1-oxomeliac-8(30)-enate with TbPTR1, and evodulone with TbCatB. We conclude, therefore, that T. brucei sterol 14α-demethylase, TbCYP51, is a protein target of C. procera limonoids.
Enantia chlorantha is dominated by aporphine and berberine alkaloids (Table S7) [40], [63], [64]. E. chlorantha aporphine alkaloids seem to show a propensity for docking with TbPTR1 or with TbUDPGE while the berberine alkaloids showed selectivity for TbPTR1. Both pseudocolumbamine and pseudopalmatine docked with TbPTR1 with docking energies (−27.5 kcal/mol) comparable to the co-crystallized ligand, 6-phenylpteridine-2,4,7-triamine [25] (−27.6 kcal/mol). These planar alkaloids dock into the active site by way of hydrophobic interactions with the NADP+ cofactor and a hydrophobic pocket formed by Phe97, Met163, Cys168, Pro210, Trp221, and Leu209 (Fig. 2). Liriodenine and columbamine docked selectively to TbTIM with docking energies lower (−24.0 and −24.7 kcal/mol) than the co-crystallized ligand, 3-phosphoglyceric acid [32] (−21.6 kcal/mol). These nearly planar alkaloids are known also to be DNA intercalators and topoisomerase inhibitors [65].
Polyphenolic compounds, flavonoids, biflavonoids, etc., have been isolated and identified from Garcinia kola (Table S8) [40], [66]. G. kola biflavonoids docked favorably with TbAK and TbODC. The biflavonoids do not dock at the adenosine binding sites of TbAK, but rather in a pocket between the two sites bounded by residues Asn222, Gly298, Ala297, Thr264, Asp266, Glu225, Arg132, and Asn195 (see Fig. 3). Likewise, biflavonoid docking with TbODC does not occur at the ornithine/putrescine binding site or the geneticin binding site, but rather in a pocket bounded by Asp243, Asp385, Val335, Asp332, Ala334, Ala244, and Arg277 (Fig. 4). This would suggest that if G. kola biflavonoids inhibit either TbAK or TbODC, they act as allosteric inhibitors of these proteins. The two tocotrienols garcinal and garcinoic acid, on the other hand, docked more favorably with TbUDPGE. Key interactions of the tocotrienols with the protein are hydrogen-bonding of the phenolic –OH of the ligands with Pro253 and Phe255, hydrogen-bonding of the carbonyl group of the ligand side chains with Arg268, hydrogen-bonding of the pyran ring oxygen atom with Arg235, face-to-face π – π interactions of the ligand aromatic rings with Phe255, and hydrophobic interactions of the tocotrienol ligands with Leu222, His221 and the NAD cofactor (Fig. 5 top). The prenylated benzophenone kolanone docked very strongly with TbNH (docking energy = −37.1 kcal/mol) in the nucleoside binding site (Fig. 6), a hydrophobic pocket bounded by Trp80, Phe178, Asn171, Trp205, and Val277, with additional hydrogen-bonding with Asn171.
The phytochemical compositions of Khaya ivorensis [67] and K. senegalensis [58], [68]–[72], like other members of the Meliaceae, are characterized by limonoids [40]. Many of the Khaya limonoids showed markedly strong docking to TbAK as well as TbCYP51 (see Table S9). Of particular note, 3-O-acetylkhayalactone strongly docked with TbAK, TbDHFR, and TbUDPGE (−31.6, −32.2, and −34.2 kcal/mol, respectively). This ligand docked in the same site in TbAK as the Garcinia biflavonoids (above), but in a different position in TbUDPGE (Fig. 5 bottom). Important hydrogen-bonding interactions of 3-O-acetylkhayalactone with TbUDPGE are with residues Glu214, Ser219, Leu102, Thr220, and His221. 3-O-Acetylkhayalactone docked in the active site of TbDHFR in the same general location as the co-crystallized ligand (Fig. 7). In addition, the docking energy of 3-O-acetylkhayalactone (−32.2 kcal/mol) was lower than either of the co-crystallized ligands, 5-(4-chlorophenyl)-6-ethylpyrimidine-2,4-diamine (pyrimethamine) and 6,6-dimethyl-1-[3-(2,4,5-trichlorophenoxy)propoxy]-1,6-dihydro-1,3,5-triazine-2,4-diamine [26] (−22.7 and −30.1 kcal/mol, respectively). In general, the Khaya limonoids showed weak or no docking with TbNH, TbTIM, or TbNDRT.
The sterols and triterpenoids [40] from Lawsonia inermis showed preferential docking to TbCYP51 (T. brucei sterol 14α-demethylase) (Table S10). This is perhaps not surprising since the normal substrates for this enzyme are sterols. The laxanthones from L. inermis showed preferential docking to TbPTR1 with docking energies comparable to the co-crystallized ligand. In addition, they docked in the same positions and orientations as pseudocolumbamine and pseudopalmatine from Enantia chlorantha (see above and Fig. 2).
Anthraquinones dominate the phytochemistry of Morinda lucida [40], [73], [74], along with triterpenoid acids [75] (see Table S11). Anthraquinones, as a class, demonstrated significant docking affinity for TbPRT1 and TbTIM (Fig. 8). Anthraquinones are also known to be DNA intercalators and topoisomerase inhibitors [76]. The triterpenoid acids oleanolic acid and ursolic acid showed notable docking energies with TbCYP51 (see above). The phenylpropanoid oruwacin docked very strongly to TbPTR1 (docking energy = −32.4 kcal/mol) and TbNH (docking energy = −31.6 kcal/mol, Fig. 6).
The phytochemistry of Morinda morindoides [40] is dominated by flavonoid glycosides [77] and phenylpropanoid-conjugated iridoid glycosides [78] (Table S12). Of these, epoxygaertneroside and morindaoside were selectively strongly binding ligands for TbAK, and morindaoside also docked strongly to TbPTR1. A number of gaertneroside derivatives showed docking selectivity for TbDHFR (see Table S12), while dehydroepoxymethoxygaertneroside docked very strongly with TbNH, occupying the nucleoside binding site (Fig. 6) with the same hydrophobic interactions as kolanone and oruwacin (above). It is unlikely that these glycosides will remain intact in vivo, and hydrolysis may be necessary for absorption and general bioavailability [79]. Of the flavonoid aglycones from M. morindoides, apigenin, chrysoeriol, kaempferol, quercetin, and ombuin selectively docked with TbPTR1.
Phytochemical investigations of Nauclea latifolia have revealed numerous indole alkaloids [40], [80], [81] (Table S13). The alkaloid glycosides 10-hydroxystrictosamide, cadambine, dihydrocadambine, and tetrahydrodesoxycordifoline showed notably strong docking to TbUDPGE, whereas non-glycosylated alkaloids showed preferential docking with TbPTR1and/or TbAK. 10-Hydroxyangustine, angustine, naucleamide B, and naucletine, in particular, docked more strongly with TbPTR1 than the co-crystallized ligand, 6-phenylpteridine-2,4,7-triamine [25] (docking energy = −27.6 kcal/mol).
T. brucei triosephosphate isomerase, TbTIM, is the likely protein target for the phytochemical agents of Newbouldia laevis. Both furanonaphthoquinones [82], [83] and pyrazole alkaloids [84], [85] from this plant showed remarkable selective affinity for this protein (Table S14). The monomeric furanonaphthoquinone ligands all occupy the same site with hydrogen bonding of the furan oxygen and C(9) carbonyl oxygen to Lys313; C(4) carbonyl oxygen with Ser513 and Val514; a van der Waals surface provided by Val533, Gly534, Gly535; and a hydrophobic pocket to accommodate the isopropenyl moiety provided by Ile472, Gly512, and Leu532 (see Fig. 9). Similarly, the pyrazole alkaloid 4′-hydroxywithasomnine has key hydrogen-bonding interactions between the pyrazole ring nitrogens and Ser513 and Val514. The aromatic ring lies in the hydrophobic pocket made up of Ile472 and Leu532, and there is an additional hydrogen-bonding interaction between the phenolic –OH group and His395 (Fig. 9).
Withanolide triterpenoids, abundant components of Physalis angulata [40], [86]–[88], generally showed preferential docking to T. brucei sterol 14α-demethylase, TbCYP51 (Table S15). This is consistent with the docking of L. inermis triterpenoids and steroids, C. procera and Khaya limonoids (see above). Five of the withanolides, 14-hydroxyixocarpanolide, physagulin J, physagulin L, withangulatin H, and withangulatin I, docked more strongly to TbCYP51 than the co-crystallized ligand [16]. Three withanolides, physagulin A, physagulin L′, and withangulatin A, docked more strongly into TbODC than the co-crystallized ligand (pyridoxal 5′-phosphate) for that protein [34]. The pyrrolidine alkaloid, phygrine, docked with TbTIM preferentially.
Picralima nitida glycosylated coumestans [89] showed strong binding to most of the protein targets, except for TbNDRT (Table S16). They are, for example, along with Acacia nilotica flavonoid gallates, the only ligands that dock to rhodesain with docking energies comparable to the co-crystallized ligand. Of the ligands examined in this work, coumestan 2 is the strongest-binding ligand for TbAK (−44.1 kcal/mol) and TbUDPGE (−43.7 kcal/mol). The corresponding aglycones, 4, 5, and 6, however, were selective for TbPTR1 as well as TbUDPGE.
The phytochemistry of Prosopis africana is characterized by piperidine alkaloids (Table S17) [40], [90]. These ligands exhibited similar docking energies with all protein targets, owing presumably to the small, flexible nature of the compounds. They did, however, show slightly better affinity for TbAK.
Phytochemical investigations of Rauwolfia vomitoria have revealed this plant to be replete with indole alkaloids (Table S18) [40], [91], [92]. The structural diversity of these indole alkaloids seems to defy targeting any one particular protein. There are some notable docking results, however. 3-Epirescinnamine docked with TbPRT1 and with TbODC more strongly than the co-crystallized ligands, 6-phenylpteridine-2,4,7-triamine [25] and putrescine [34], respectively. Isoreserpiline, raumitorine, and rauvanine also docked strongly to TbPTR1. Ajmalimine, isoreserpiline, rauvomitine, and serpenticine had remarkable docking energies with TbNH. The trimethoxybenzoyl and trimethoxycinnamyl esters, renoxydine, rescidine, rescinnamine, reserpine, along with methyl 3,4-dimethoxybenzoylreserpate, were all excellent ligands for TbUDPGE, with docking energies comparable to uridine-5′-diphosphate, the co-crystallized ligand [20]. Of these, renoxydine, rescidine, and reserpine, along with neonorreserpine, docked strongly to TbCYP51.
Cinnamate esters from Securidaca longipedunculata [40] showed selective docking to TbTIM while S. longipedunculata xanthones [40], [93] had a docking preference for TbPTR1. Both of these protein targets have relatively small binding sites, which are more suitable for the small ligands (Fig. 10). The xanthones also docked relatively strongly with TbUDPGE. The S. longipedunculata indole alkaloids dehydroelymoclavine and alkaloid A [94] docked strongly to TbPTR1 and TbAK, respectively, as well as with TbUDPGE (see Table S19).
Phytochemicals isolated from Strychnos spinosa include secoiridoids [95], indole, pyridine, and naphthyridine alkaloids [40], sterols and triterpenoids [96] (Table S20). Relatively small pyridine and naphthyridine alkaloids from Strychnos spinosa showed preferential docking to TbPTR1 and/or TbTIM. The indole alkaloids akagerine, 10-hydroxyakagerine, and kribine also docked preferentially to TbPTR1, while iridoid glucosides preferred TbUDPGE. Both sterols and triterpenoids from S. spinosa selectively docked with TbCYP51, similar to what was observed with L. inermis sterols and triterpenoids (see above), but these compounds also showed notably strong docking with TbODC. Interestingly, a comparison of docking energies of triterpenoid and steroid ligands with their antitrypanosomal activities [96] shows no correlation, even comparing TbCYP51 docking or TbODC docking. Plots of log(IC50) vs. docking energies gives R2 values of 0.043 and 0.007 for TbCYP51 and TbODC, respectively. It may be that inhibition of some other protein target [8], [9] is the biochemical mechanism of activity for these compounds.
In terms of natural products drug discovery, it is useful to examine whether different phytochemical classes show selectivity for particular protein targets. Simple flavonoid ligands showed docking preferences for TbPTR1 and TbUDPGE. Flavonoid gallates, on the other hand, were shown to be promiscuous docking ligands to all protein targets, but were particularly strongly docking with TbAK, TbPRT1, TbCYP51, and TbNH. Likewise, flavonoid glycosides tended to be promiscuous docking agents, but with preference for TbAK, TbPRT1, and TbNH. Oligomeric flavonoids (tannin-like polyphenolics) showed strong docking to TbAK. The diversity of flavonoid structures has led to diverse biological activities, including antiprotozoal activity, but the modes of antiprotozoal activity have not been well elucidated [97].
As previously noted (see above), triterpenoid ligands were largely selective for TbCYP51. Withanolide triterpenoids also showed a docking preference for TbCYP51, while limonoids preferentially docked with TbAK as well as TbCYP51. Not surprisingly, sterols showed a propensity to dock with TbCYP51, but also docked strongly with TbUDPGE.
All of the anthraquinone ligands examined in this docking study, docked with strong binding energies to TbPRT1. Likewise, xanthone ligands exhibited docking selectivity for TbPTR1. Naphthoquinones, on the other hand, docked preferentially with TbTIM. Most chromene ligands also showed notable docking energies to TbTIM. The phenylpropanoids examined showed preferences for TbTIM as well as TbUDPGE, while glycoside derivatives of phenylpropanoids showed selectivity for TbDHFR.
Berberine alkaloids docked preferentially to TbPTR1 while aporphine alkaloids showed some selectivity for TbPTR1 and TbUDPGE. Piperidine alkaloids were also selective for TbUDPGE. Pyrazole and pyridine alkaloids, on the other hand, preferred docking to TbTIM. A total of 93 indole alkaloids were examined in this docking study and many of them showed notable docking energies with TbUDPGE and some with TbAK and TbPTR1. Glycoside derivatives of alkaloids also preferentially docked with TbUDPGE.
Overall, the protein objects most targeted by the phytochemical ligands in this study were TbUDPGE, targeted by many alkaloids; TbPTR1, preferred by planar-like ligands; TbCYP51, which docked terpenoid ligands well; and TbAK, which docked many different classes of phytochemicals. Those proteins least preferred in terms of docking energies were rhodesain, TbCatB, and TbNDRT.
Rhodesain and TbCatB are both cysteine proteases with relatively small binding sites. It may be that the docking energies reflect the fact that only relatively small ligands, with inherently small docking energies, can fit well into the binding sites of these two proteins. The docking energies do not, however, reflect the potential for covalent bonding to the active sites of these proteins. It is useful, therefore, to examine small electrophilic ligands for energetically favorable docking orientations that would allow for reaction of nucleophilic amino acid side chains to the electrophilic sites of the ligands.
Although umbelliferone does not dock with particularly strong energies to rhodesain or TbCatB, it does dock in poses such that the nucleophilic Cys25 of rhodesain or Cys122 of TbCatB are poised to undergo conjugate addition to the pyrone ring (Fig. 11). The S atom of Cys25 is 3.16 Å from C(4) of docked umbelliferone in rhodesain, while in TbCatB, Cys122 is 3.65 Å from C(4) of umbelliferone. Coumarins have been shown to be trypanocidal agents [98] and it has been suggested that umbelliferone undergoes conjugate addition with available cysteine thiol groups [99].
Many naphthoquinones have been shown to be antitrypanosomal [100], and are suspected to interfere with redox thiol metabolism by inhibition of TbTR [101], [102]. There are docking poses, albeit not the lowest-energy poses, of isoplumbagin (docking pose energy = −9.9 kcal/mol) and lawsone (docking pose energy = −8.6 kcal/mol) with TbTR such that these quinone ligands are in the proximity of reduced trypanothione (Fig. 12). Similarly, both isoplumbagin and lawsone dock with the cysteine proteases rhodesain and TbCatB with the electrophilic carbons near the active-site cysteine residues (Fig. 13). N. laevis furanonaphthoquinones (α-lapachone derivatives) also dock with rhodesain in poses such that the nucleophilic Cys25 can undergo Michael addition to the quinone ring (Fig. 14). None of the furanonaphthoquinones docked near the trypanothione thiol groups in TbTR, however.
This in-silico investigation suggests that trypanosomal phytochemicals may target different protein targets. There are several caveats to these docking results: (a) many of the phytochemical agents may not be bioavailable due to limited solubility, membrane permeability, hydrolysis, or other metabolic decomposition; (b) tannins and other polyphenolics are promiscuous protein binding agents and are likely, therefore, not selective antitrypanosomal ligands; (c) the docking studies do not account for synergism in bioactivity of phytochemicals; (d) this current study does not address the binding of ligand to human homologous isozymes, which may also be targeted; (e) there are likely additional phytochemicals in each of the medicinal plants that have not been isolated or identified; and (f) there are likely additional trypanosomal proteins or other biochemical targets that have not yet been identified. Nevertheless, this in-silico molecular docking study has provided evidence for what phytochemical classes and structural manifolds are targeting particular trypanosomal protein targets and could provide the framework for synthetic modification of bioactive phytochemicals, de novo synthesis of structural motifs, and further phytochemical investigations.
|
10.1371/journal.ppat.1003190 | Intracellular Bacillary Burden Reflects a Burst Size for Mycobacterium tuberculosis In Vivo | We previously reported that Mycobacterium tuberculosis triggers macrophage necrosis in vitro at a threshold intracellular load of ∼25 bacilli. This suggests a model for tuberculosis where bacilli invading lung macrophages at low multiplicity of infection proliferate to burst size and spread to naïve phagocytes for repeated cycles of replication and cytolysis. The current study evaluated that model in vivo, an environment significantly more complex than in vitro culture. In the lungs of mice infected with M. tuberculosis by aerosol we observed three distinct mononuclear leukocyte populations (CD11b− CD11c+/hi, CD11b+/lo CD11clo/−, CD11b+/hi CD11c+/hi) and neutrophils hosting bacilli. Four weeks after aerosol challenge, CD11b+/hi CD11c+/hi mononuclear cells and neutrophils were the predominant hosts for M. tuberculosis while CD11b+/lo CD11clo/− cells assumed that role by ten weeks. Alveolar macrophages (CD11b− CD11c+/hi) were a minority infected cell type at both time points. The burst size model predicts that individual lung phagocytes would harbor a range of bacillary loads with most containing few bacilli, a smaller proportion containing many bacilli, and few or none exceeding a burst size load. Bacterial load per cell was enumerated in lung monocytic cells and neutrophils at time points after aerosol challenge of wild type and interferon-γ null mice. The resulting data fulfilled those predictions, suggesting a median in vivo burst size in the range of 20 to 40 bacilli for monocytic cells. Most heavily burdened monocytic cells were nonviable, with morphological features similar to those observed after high multiplicity challenge in vitro: nuclear condensation without fragmentation and disintegration of cell membranes without apoptotic vesicle formation. Neutrophils had a narrow range and lower peak bacillary burden than monocytic cells and some exhibited cell death with release of extracellular neutrophil traps. Our studies suggest that burst size cytolysis is a major cause of infection-induced mononuclear cell death in tuberculosis.
| Macrophages patrol the lung to ingest and destroy inhaled microbes. Mycobacterium tuberculosis, the bacteria causing tuberculosis, can survive within macrophages and use them as a protected environment for growth. Macrophages by themselves are poorly equipped to kill M. tuberculosis but may undergo programmed cell death (apoptosis) to limit bacterial replication. Virulent M. tuberculosis has evolved the capacity to inhibit macrophage apoptosis, thereby protecting the replication niche. In previous studies we showed that upon reaching a threshold intracellular number (burst size), virulent M. tuberculosis kills macrophages by necrosis and escapes for spreading infection. The present study was designed to test whether this mechanism seen in vitro operates during pulmonary tuberculosis in vivo. The distribution of M. tuberculosis numbers inside lung phagocytes of mice with tuberculosis conformed to predictions based on the burst size hypothesis, as did the appearance of dying cells. We identified four different types of phagocytes hosting intracellular M. tuberculosis. The distribution of M. tuberculosis load within individual phagocytes and between different types of phagocyte changed over the course of tuberculosis disease. These studies reveal the complexity of host defense in tuberculosis that must be considered as new therapies are sought.
| Natural infection with Mycobacterium tuberculosis (Mtb) occurs by inhalation, followed by invasion of resident alveolar macrophages that provide the major initial replication niche for the pathogen. Macrophages infected with Mtb in vitro may die with primarily apoptotic or necrotic features [1]; the cell death mode most relevant to TB disease in vivo remains undefined. A widely held paradigm is that macrophage apoptosis promotes host defense in TB while necrosis favors spreading infection. We previously reported that the cytolytic activity of Mtb correlates with intracellular bacillary burden in macrophages, increasing dramatically at a threshold load of ∼25 bacilli per macrophage [2]. At high intracellular burden, Mtb triggers a primarily necrotic death dependent on bacterial genes regulated by the PhoPR 2-component system [3]. Our in vitro studies and data from other groups suggest that virulent Mtb strains suppress apoptosis of host macrophages [4]–[8] and grow to a threshold burden [2], [9] whereupon necrosis is triggered as an exit mechanism analogous to the burst size of lytic viruses.
In the present study, we investigated whether the necrotic death described for Mtb-infected macrophages in vitro is relevant to the fate of monocytic cells in the lung that become infected during the course of TB disease in vivo. Inhalation of Mtb is followed by the invasion of a small number of resident alveolar macrophages. We posit that within each infected macrophage, bacterial replication expands an initial low multiplicity of infection (MOI) to a burst size value. Once this threshold is exceeded, the liberated bacilli spread to naïve phagocytes. Successive rounds of invasion, replication and escape will result in a distribution of bacillary loads across the population of infected phagocytes. This model predicts that at any given time point after low dose aerosol challenge, phagocytes harboring 1–10 bacilli will outnumber those with higher bacillary loads, and that host cells containing ≥25 bacilli will be a distinct minority of infected cells. The model also predicts that with the induction of adaptive immunity (∼3 weeks after aerosol challenge), inhibition of Mtb replication will rescue many infected cells with a low bacillary burden from progressing to burst size. This will increase the proportion of cells containing 1–10 bacilli while heavily infected cells will die and be replaced at a low rate thereby reducing the proportion of cells containing ≥25 bacilli.
To test those predictions we enumerated acid fast bacilli (AFB) per cell in whole lung leukocytes and bronchoalveolar lavage (BAL) cells harvested from mice after low dose aerosol infection with Mtb Edrman. The distribution of AFB burden in monocytic cells harvested from wild type (WT) C57BL/6 mice followed the predicted pattern. Analysis of interferon-γ knockout (GKO) mice with TB also conformed to the predicted effects of unrestricted Mtb replication on the distribution of bacillary loads per cell. The morphology of heavily infected monocytic cells isolated from the lungs of mice with TB exhibited features similar to those seen after high MOI challenge in vitro: nuclear condensation without fragmentation and loss of cell membrane integrity without cell swelling or apoptotic vesicle formation. Taken together, these results support the burst size hypothesis for TB in vivo.
During the course of these experiments it became apparent that the diversity of cell types hosting Mtb in vivo adds an additional layer of complexity to TB pathogenesis. Neutrophils were major Mtb host cells 2–3 weeks post infection (p.i.) but rarely contained >10 AFB per cell. We also observed differences in the distribution of Mtb between three subpopulations of mononuclear leukocytes classified by CD11b and CD11c expression. The proportion of infected host cells differed between these subpopulations, and the proportions changed dynamically between 4 and 10 weeks p.i. The relative permissiveness or restriction of intracellular Mtb replication as well as the regulation of host cell fate may differ for each of these phagocytic cell types with implications for host defense, immune pathology and latent TB infection.
Alveolar macrophages are the overwhelming majority of leukocytes in the normal alveolar space and they are considered the primary leukocyte type initially infected by inhaled Mtb in vivo. Any investigation of the burst size hypothesis for TB must consider the complexity of the in vivo environment and the potential for Mtb to invade a diverse spectrum of phagocytic cells in the lung [10], [11]. To study the distribution of Mtb within subpopulations of monocytic cells, we harvested lungs of WT mice 4 and 10 weeks after aerosol challenge with Mtb Erdman. Whole lung CD45+ leukocytes were sorted on the basis of CD11b and CD11c expression. Previous reports have extensively characterized subpopulations of leukocytes in lung tissues of mice with TB [11]–[18]. In accordance with those studies, we classified alveolar macrophages (AM) as CD11b− CD11c+/hi, recruited monocyte/macrophages (RM) as CD11b+/lo CD11clo/− and myeloid dendritic cells (mDC) as CD11b+/hi CD11c+/hi (Fig. S1A). For clarity we refer to AM, RM and mDC populations in this report based on CD11b and CD11c expression, recognizing that this might not invariably correspond to functional identity. We evaluated additional cell surface markers including Ly-6G (Fig. S1B), CD115, F4/80 and MHC class II (not shown) but these provided no additional useful discriminatory information. Uninfected mice were sampled for comparison to the TB group at both time points (Table S1).
Mononuclear cells from uninfected lungs comprised a roughly equal proportion of AM and RM, with mDC a distinct minority at 1.6% (Fig. 1A). By 4 weeks p.i., the total number of mDC increased >40-fold and they expanded proportionately from 1.6% to 18.5% of monocytic cells. The distribution of monocytic cells changed slightly at 10 weeks p.i., with a moderate expansion of RM and modest contraction of the AM and mDC populations. Mtb-infected monocytic cells were enumerated by microscopy on cytospin slides with Ziehl-Neelsen staining (Table S2). After 4 weeks of TB disease, mDC were the predominant Mtb-infected cell type, representing 78.2% of total AFB+ monocytic cells (Fig. 1B). By 10 weeks p.i., the distribution of infected cells shifted, with RM becoming the primary Mtb host cells (60.4%), followed by mDC (31.2%). At both time points, AM represented approximately 10% of AFB+ monocytic cells. The relative propensity for different mononuclear leukocyte cell types to harbor Mtb was estimated based their representation in the total and AFB+ cells within the total population. On that basis, mDC were 13 and 18 times more likely to be infected with Mtb compared to AM or RM, respectively, at 4 weeks p.i.
Resident AM, by far the predominant leukocytes in the airspace of normal lungs, are the first phagocytes to become infected by Mtb following inhalation of droplet nuclei. In the absence of infection or inflammation they are extremely long-lived, non-replicating cells with negligible replacement by bone marrow-derived leukocytes [19]. To investigate the impact of pulmonary TB on resident AM, we labeled these cells by tracheal instillation of a replication incompetent, VSVG-pseudotyped, GFP-expressing lentivirus (CMV-GFP-W) as previously described [20]. Labeling efficiency is typically 30–40% of BAL cells, with no transduction of epithelial cells or parenchymal leukocytes. After an early loss of transduced cells in the first several weeks, the number of GFP+ cells stabilizes at 20–30% of BAL cells for up to 2 years. To track the fate of resident AM in TB, mice were transduced with intratracheal CMV-GFP-W and rested for 8 weeks to allow the GFP+ cell population to stabilize. One group of transduced mice was then infected with Mtb Erdman by aerosol while control transduced mice remained uninfected. Sets of Mtb-infected and control mice were sacrificed 4 weeks and 10 weeks after infection of the TB group and whole lung leukocytes were isolated for analysis by flow cytometry (Fig. 1C).
In uninfected mice, GFP+ leukocytes comprised 97% AM, 2.8% mDC and <0.3% RM or other cells (Table S3) reflecting the typical composition of leukocytes in airspace under basal conditions. After 4 weeks of TB disease, the total number of GFP+ cells was little changed from baseline but the proportion AM among all GFP+ cells fell to 68% while mDC increased to 27% of GFP+ cells. Thus, ∼22% of AM shifted their surface phenotype to one resembling mDC. Whether this represents induction of CD11b surface expression on cells that retain AM properties, or if these cells convert into functional mDC remains to be determined. The phenotypic shift of resident AM contributed only ∼3% of the total increase in lung mDC at 4 weeks p.i., indicating increased mDC population mostly resulted from differentiation of monocytic cells newly recruited to the lung. Comparison of total GFP+ cells counts between groups was limited by variation in labeling efficiency but the data suggest a trend after 10 weeks of TB for a modest loss of AM that were resident in the lung prior to infection. We found no GFP+ leukocytes containing AFB at 4 or 10 weeks p.i., suggesting little or no horizontal spread of infection through the resident AM population present in the lung prior to infection. Most if not all of the AM accounting for ∼10% of Mtb-infected monocytic cells at 4 and 10 weeks p.i. (Fig. 1B) were recruited to the lung after infection. We conclude that resident AM may be critical host cells for intracellular infection by inhaled Mtb, but only for the first round of bacillary replication to burst size. Subsequently, the rapidly expanding number of bacilli shifts to phagocytes newly recruited to the lung, most of which do not differentiate into AM.
To investigate the in vivo relevance of a burst size model for macrophage cell death in TB, we challenged C57BL/6 mice with virulent Mtb Erdman by aerosol set to deliver ∼300 CFU to the lung. Lung cells were subsequently harvested in different sets of infected mice exclusively by BAL or by enzymatic digestion to isolate whole lung leukocytes. Cells were immobilized onto slides by cytocentrifugation and intracellular bacilli were visualized with Ziehl-Neelsen staining. The percent and total number of AFB+ cells was enumerated at time points between 1 and 8 weeks p.i. In parallel with the characteristic kinetics of lung bacillary load in the aerosol TB model, the total number of AFB+ cells increased logarithmically until reaching a plateau value between 2 and 3 weeks p.i. (Fig. 2A). When compared to whole lung leukocytes, the total number of AFB+ BAL cells declined by 8 weeks p.i., likely reflecting a loss of airspace available to lavage. In contrast, the total number of AFB+ lung leukocytes remained stable between 2 and 8 weeks p.i., reflecting the major site of TB disease more accurately than BAL at later time points. The proportion of BAL cells and lung leukocytes infected with Mtb peaked around 1.5% at 2 weeks p.i. and then declined, owing to the recruitment of naïve leukocytes to the infected lung (Fig. 2B). Light microscopy allowed reliable identification of neutrophils that were therefore counted separately from monocytic cells. Mononuclear leukocyte subsets, comprising AM, RM and mDC, are not visually distinguishable. AFB+ neutrophils were not seen in BAL or lung leukocytes at 1 week p.i. but equaled monocytic cells as hosts for Mtb at weeks 2 and 3, the period of greatest bacillary expansion (Fig. 2C).
At each time point after aerosol challenge, AFB per neutrophil or monocytic cell was counted in individual cells. AFB counts were grouped into five bins (1–5, 6–10, 11–15, 16–20 and ≥21) with the top bin reflecting the highest burden that we could reliably count to ±2 AFB (Fig. 3). Consistent with the distribution of intracellular bacillary loads predicted for a burst size model, the number of AFB+ monocytic cells containing 1–5 AFB was greater than cells with higher bacillary loads at all time points (Fig. 4A). Also consistent with the burst size hypothesis, the proportion of AFB+ monocytic cells harboring ≥21 bacilli peaked during the period of logarithmic Mtb replication at weeks 1 to 3 p.i. and then declined in parallel with induction of adaptive immunity. We previously reported an in vitro burst size of ∼25 bacilli for Mtb-infected bone marrow derived macrophages (BMDM) [2]. In the present study, the bin of whole lung monocytic cells containing ≥21 AFB peaked at 2.8% of all AFB+ monocytic cells by 2 weeks p.i. and then fell to 0.1% by 8 weeks p.i. (Fig. S2). We posit that cells with high burden reflect Mtb replication towards burst size, which dropped by a factor of 10 following the induction of adaptive immunity. The absolute number of heavily infected cells declined between 3 and 8 weeks p.i., indicating that cells dying after reaching burst size were being replaced at a reduced rate as host immunity limited Mtb replication. Monocytic cells estimated to contain up to 50 or more AFB were very rarely seen, representing 0.017% of AFB+ cells at 3 weeks p.i. Most such cells appeared nonviable, with faintly stained cytoplasm and bacilli breaching the plasma membrane (Fig. S3A).
We counted AFB per cell in cytospin preparations of flow-sorted monocytic cell types but found few cells containing >10 AFB. This was inconsistent with results obtained with BAL cells or lung leukocytes that were cytocentrifuged directly onto the slides without further processing. Speculating that heavily burdened phagocytes progressing to necrosis were fragile and unable to withstand the stress of flow sorting, we examined the sorted population of non-viable cells defined by forward-scatter and side-scatter characteristics. Ziehl-Neelsen stained cytospins of those samples revealed much higher numbers of AFB in these dead and dying cells than in the sorted populations of viable cells (Fig. S3B). Collectively, these results support the concept of a mononuclear leukocyte burst size for virulent Mtb in vivo with a median value in the range of 20–40 bacilli. Of interest, while lung leukocytes with ≥16 AFB declined by 84.5% between 3 and 8 weeks p.i., some high burden cells were still seen at the later time point, well after the induction of adaptive immunity when total lung bacillary load is held stable. Our findings in TB contrast with a report that viable macrophages isolated from the footpads of M. leprae-infected athymic nu/nu mice contained an average of 120 AFB per cell [21]. Unlike Mtb, M. leprae is not cytolytic for macrophages in vitro even at MOI 100 [22].
In contrast to the distribution of Mtb in monocytic cells, neutrophils containing >15 AFB were not seen at any time point, and neutrophils with 11–15 AFB were identified only at 2 and 3 weeks p.i. (Fig. 4B). This corresponds to the period of logarithmic Mtb expansion in the lung and the peak number of the most heavily infected monocytic cells. These results suggest that neutrophils are recruited to the vicinity of necrotic monocytic cells and acquire bacilli at low to moderate MOI. The data also imply that neutrophils may be poor hosts for Mtb replication, presumably owing to their limited lifespan. Alternatively, neutrophils could be subject to burst size cytolysis with a lower threshold value or some other cell death mode in the context of TB.
Interferon (IFN)- γ plays a critical role in protective immunity against Mtb by activating macrophages to limit bacterial replication [23]. After aerosol challenge of GKO mice, lung Mtb burden increases logarithmically until death by 4–6 weeks p.i. We delivered 100 CFU of Mtb Erdman to WT and GKO mice and then harvested BAL cells for cytospin and Ziehl-Neelsen staining at five time points from 7 to 21 days. As expected, the total number of AFB+ BAL cells increased progressively in GKO mice while in WT mice it was held to a plateau value after day 18 p.i. (Fig. 5). It was recently proposed that IFN-γ limits neutrophil recruitment to the lung in the transition from innate to adaptive immunity in TB [24]. Consistent with that report, neutrophils represented a higher proportion of BAL cells in GKO compared to WT mice with TB (Fig. 6A). While neutrophils represented only 16% of BAL leukocytes in WT mice at 18 days p.i., they accounted for ∼50% of AFB+ cells at that time point (Fig. 6B). This suggests that despite the influence of IFN-γ on neutrophil trafficking, these cells are recruited to the immediate vicinity of Mtb infection where they may exert comparatively high phagocytic activity.
Based on the burst size hypothesis, unrestricted intracellular Mtb replication in GKO mice is predicted to result in a persistently high proportion of heavily infected monocytic cells past the time point when WT mice start to restrict Mtb replication. Bacterial load per cell in BAL monocytic cells and neutrophils from GKO and WT mice was counted at each time point and tabulated in bins (Fig. 7). GKO mice had a higher proportion and total number of heavily infected cells on day 7 p.i. suggesting an innate IFN-γ response [25], [26] that limits Mtb replication before adaptive immunity is expressed. Prior to the full induction of adaptive immunity, Mtb replication in WT mice also proceeded at a rapid rate such that on day 14 p.i. the distribution of AFB loads was similar to that of GKO mice. As an effective IFN-γ-dominated adaptive immune response was expressed in WT mice, the proportion of heavily infected BAL cells declined. At day 21 p.i., the proportion AFB+ cells in the top three bins were significantly higher in GKO compared to WT mice. AFB+ cells containing 1–5 bacilli remained the most abundantly populated bin in GKO mice at all time points and cells containing >50 AFB were very rarely seen, as was the case in WT mice. We interpret the distribution of AFB loads in GKO compared to WT mice as supporting the burst size hypothesis and also indicating that IFN-γ has little if any direct influence on the burst size value.
Macrophages challenged in vitro with Mtb at MOI ≥25 rapidly undergo an atypical, caspase-independent cell death dominated by lipolytic attack on lipid membranes [3]. This cell death mode has unique morphological features including nuclear condensation without fragmentation, and disintegration of lipid bilayers throughout the cell. In the present study, we compared the morphology of BAL cells harvested 3–4 weeks after low dose aerosol Mtb Erdman challenge to that of BMDM infected for 3 h in vitro with Mtb at MOI 25. Similar to the characteristic changes seen in vitro, monocytic cells from the lungs of mice with TB exhibited nuclear condensation and this was restricted to those cells with a high AFB burden (Fig. 8A). Nuclear fragmentation, a characteristic of caspase-mediated apoptosis, was not seen in >5×105 DAPI-stained lung cells from mice with TB. The morphology of lung leukocytes having low numbers of intracellular bacilli was uniformly similar to the normal appearance of uninfected cells (Fig. S4).
Mtb-induced macrophage cytolysis in vitro is characterized by disintegration of mitochondrial, nuclear and plasma membranes without cell swelling or formation of apoptotic vesicles. To examine the ultrastructural features of infected lung leukocytes in the context of pulmonary TB, BAL cell cytospin preparations were visualized by scanning electron microscopy (EM) and compared to BMDM challenged with Mtb Erdman in vitro at MOI 25 (Fig. 8B). BAL cells isolated from mice with TB showed a similar pattern of injury to BMDM infected in vitro: plasma membrane damage and no evidence of budding vesicles or osmotic lysis. We also observed extrusion of chromatin through damaged nuclear membranes in dying BAL cells (Fig. S5), akin to results we previously reported with in vitro Mtb infection [3]. Taken together, these observations demonstrate consistent similarities between Mtb-induced cytolysis in vitro and in vivo.
A high proportion of lung neutrophils were infected with Mtb after aerosol challenge and they were a significant host cell compartment for bacilli even after 8 weeks of TB disease. Compared to monocytic cells, neutrophils have a short lifespan in the lung in the absence of inflammation, which would seem to make them unproductive hosts for Mtb replication. Examining cytospin preparations of BAL cells from GKO mice, we observed large masses of amorphous extracellular material with numerous associated AFB (Fig. S6A). A combination of carbolfuchsin and DAPI stains indicated that these structures had a high content of extracellular DNA. Scanning EM revealed a network of extracellular fibrillar structures with adherent bacilli similar to the morphological features first described by Brinkmann and Zychlinsky [27]. Furthermore, the thread-like structures were abundantly studded with globular domains. The composition of NETs includes characteristic components including myeloperoxidase (MPO), neutrophil elastase and cleaved histones [28], [29]. Confocal scanning laser microscopy and immunostaining with antibodies against neutrophil elastase, MPO and histones showed co-localization of these molecules with extracellular DNA (Fig. S6B). Together, these data indicate the presence of neutrophil extracellular traps (NETs). We did not see NETs in cytospins from WT mice with TB. Release of NETs could be limited to conditions of uncontrolled Mtb replication in GKO mice but more likely occurs at a low frequency in WT mice that is undetectable by the methods we used. Furthermore, we observed Mtb-induced NETs in BAL from hypercholesterolemic ApoE null mice with TB (Fig. S6C). At the time of sampling p.i., these mice express comparable levels of IFN-γ with WT, but are unable to control bacterial replication and they develop severe neutrophilic lung inflammation [30]. Ramos-Kichik et al. [31] reported that Mtb induces NET release in vitro and that NETs trap the bacilli but are unable to kill them, in contrast to microbicidal activity of NETs against Listeria monocytogenes. We believe that our data from GKO and ApoE null mice are the first evidence for NET release in the context of TB disease in vivo and indicate that an extracellular population of bacilli may be adherent to NETs in the lung.
We examined leukocytes from the lungs of mice infected with Mtb by aerosol to test a model of burst size cytolysis suggested by prior in vitro studies. The distribution of Mtb load in monocytic cells was skewed such that most AFB+ cells contained few bacilli while a minority had a high bacillary load. The morphology of heavily infected cells mirrored that seen with Mtb-induced necrosis in vitro; they appeared nonviable, with condensed nuclei and disrupted plasma and nuclear membranes. We interpret these findings as consistent with burst size cytolysis at median threshold in the range of 20–40 AFB. That value is close to the burst size reported for in vitro infection of BMDM [2]. The comparison of WT and GKO mice demonstrated that by limiting Mtb replication to burst size, IFN-γ promotes the survival of monocytic host cells with a sublethal bacillary burden. A similar phenomenon was reported for BMDM in vitro, where virulent Mtb strains introduced at low MOI grew rapidly and caused necrosis but cytolysis was prevented when Mtb replication was inhibited by exogenous IFN-γ [9]. Despite an effective immune response in WT mice, some monocytic cells with >15 AFB were present at 8 weeks p.i., accounting for 0.6% of all AFB+ cells at that time. This implies ongoing Mtb replication in a limited population of monocytic cells balanced by microbicidal activity in as yet unknown compartments during the period of “stationary persistence” that is not reflected by an increase in total lung CFU. A similar conclusion was reached by Gill et al. [32] in a study that employed in silico modeling and in vivo experiments using Mtb transformed with an unstable plasmid replication clock.
Virulent Mtb strains inhibit host-protective apoptotic death of infected macrophages [4]–[8], permitting optimal replication before spreading to other cells. Transit between replication niches requires time and incurs risk for bacilli that may be trapped in the extracellular environment, subjected to antimicrobial activities, or taken up by phagocytes that do not support replication. Delaying host cell death for 3–5 Mtb doublings should accelerate the increase in total lung bacillary load in the critical period prior to the induction of adaptive immunity. Lacking any means to manipulate burst size in biological experiments, we took advantage of an existing computational model to test the effects of different burst size values on Mtb accumulation in the lung. This in silico agent-based model replicates the interplay between host and pathogen, taking multiple variables into account over three biological scales: molecular, cellular and tissue in a 2 mm×2 mm section of lung. The model captures burst size by setting a value to the maximum carrying capacity of Mtb per macrophage, above which the macrophage bursts, releasing viable bacilli to infect naïve cells. We varied burst size values in the computational model to 10, 20, 30, 40, and 50, keeping all other parameter values fixed in order to analyze the effects of this isolated change. To account for stochastic variability, each experiment was run 20 times (equivalent to using 20 mice for each time point).
Figure S7 shows five time courses of total bacterial counts for each burst size assumption. By day 20, a burst size >20 resulted in a significantly greater increase in total bacteria as compared to smaller burst sizes. For burst sizes 10 and 20, a peak is reached at day 20 and a lower steady state is achieved as adaptive immunity is expressed in the lung. For burst sizes >20, the Mtb count either stabilizes at the peak (burst size 30) or trends up at a slower rate (burst sizes 40 and 50). Overall, increasing the burst size resulted in higher bacterial loads, consistent with advantage for the pathogen with burst size >20. While providing independent support for the burst size hypothesis, the model has several limitations. In its current iteration, only macrophages are considered as hosts for Mtb and it assumes that every bacillus liberated from a dying macrophage invades a different new host cell at MOI 1. Our animal data show that AM, RM, mDC and neutrophils all harbor Mtb and these cell types likely differ in their capacity to support or inhibit bacillary replication. The biological data also suggest that the efficiency of Mtb escape and reinfection of new host cells is lower than the model assumes since we saw neutrophils (cells unlikely to support multiple rounds of Mtb replication) harboring up to 15 AFB. We also frequently observed Mtb in clumps that would deliver multiple bacilli if ingested by a single phagocyte (Fig. S3C). Insights from the in vivo TB study presented here will be applied to future refinements of the agent-based model.
We identified three discrete monocytic cell populations hosting Mtb, in general agreement with prior reports [10]–[12]. We recognize that monocytic cells hosting Mtb in the lung may be even further subclassified [14], and that functional heterogeneity between individual cells of the same surface phenotype is likely. Cells classified as mDC were increased in number and proportion by 4 weeks p.i. and were favored hosts for Mtb at that time. This increase was due mostly to recruitment, with a minor contribution from phenotypic shift of AM resident in the lung prior to aerosol Mtb challenge as demonstrated by lentiviral GFP labeling. By 10 weeks p.i., the predominant AFB+ monocytic cells were RM. The basis for that switch is presently unknown but might relate to the reported increase of GM-CSF and decrease of M-CSF in the lung over time after aerosol Mtb infection, which correlates with reduced DC-like cell surface markers and increased foamy macrophages [13]. The extent to which mDC and RM defined by surface phenotype differ at a functional level is unknown. They might differentially restrict or permit Mtb replication, differ in their response to IFN-γ, or differ in susceptibility to cytolysis. Ryan et al. [33] reported that Mtb induces non-apoptotic death of human peripheral blood-derived DC, with some features similar to those of murine BMDM challenged at high MOI in vitro. Our in vivo data imply that tissue mDC are subject to burst size cytolysis, but this has not been directly tested in vitro. IFN-γ-activated BMDM restrict Mtb replication in vitro more effectively than activated bone marrow-derived DC [34]. If that difference holds in vivo, it would favor RM survival with a sub-lethal Mtb load and the preferential accumulation of these cells in the lung over time p.i. Loss of heavily infected monocytic cells during the process of flow sorting prevented us from comparing the distribution of AFB loads within purified populations of mDC, RM and AM. We are exploring alternative approaches to generate such data.
We confirmed that neutrophils are also major Mtb host cells in vivo, albeit with a narrower range of AFB load than monocytic cells. As a proportion of all AFB+ cells, neutrophils were major Mtb hosts in the period of logarithmic increase of total lung bacillary load. The proportion of neutrophils with relatively high burden (6–15 AFB) also peaked at 2–3 weeks p.i. in WT mice. In GKO mice, total lung Mtb burden and infected neutrophils increased logarithmically until death. Cytokines play a major role in the positive and negative regulation of neutrophil trafficking to the lung in TB. Nandi and Behar reported that coincident with the induction of adaptive immunity, IFN-γ inhibits neutrophil accumulation in part by reducing Th17 differentiation [24]. We found that neutrophils accounted for half of all AFB+ leukocytes on day 18 p.i. in WT mice, despite being a distinct minority of total lung leukocytes at that time. This indicates that neutrophils are recruited to the proximity of necrotic cells at foci of Mtb infection in the lung. We propose that death-associated molecular patterns (DAMPs) released from monocytic cells undergoing Mtb-induced necrosis contribute to neutrophil recruitment and activation in TB. In that regard, we previously described neutrophil-rich inflammation and a high frequency of cell death in pulmonary TB lesions of diabetic and hypercholesterolemic mice that express IFN-γ to an equal or greater extent than mice without metabolic disorders [30], [35], [36], and we showed that HMGB1 is released in the course of Mtb burst size cytolysis [37]. Despite their recruitment to TB lesions and phagocytosis of bacilli, neutrophils are unlikely to host multiple rounds of Mtb replication before dying by spontaneous apoptosis or by NETosis.
In summary, our data support a burst size model for Mtb cytolysis in vivo. The features of this atypical necrotic death that we characterized in vitro are favorable for an exit mechanism in vivo. Mtb-induced cell death occurs at a threshold intracellular burden, it liberates the bacilli free of apoptotic vesicles, and it has little impact on Mtb viability. We did not find any DAPI-stained cells with fragmented nuclei or signs of apoptotic vesicle formation by scanning EM of BAL or lung leukocytes in the present study. While classical apoptosis has clearly been demonstrated in TB, our data suggest that burst size necrosis is a common fate for Mtb-infected monocytic cells in vivo. The burst size model logically fits into the pathogenesis of TB but our results highlight complex host-pathogen interactions. Resident AM are a transient niche for Mtb immediately after inhalation, but bacilli rapidly move into cells with surface phenotypes of mDC or RM, preferentially infecting the former early in disease and then shifting to the latter during stationary persistence. Neutrophils avidly acquire bacilli in the 3 week interval of logarithmic increase in total lung bacterial load. Their trafficking may be regulated in part by DAMPs released through burst size cytolysis of Mtb-infected cells. Neutrophils may promote host defense in the transition from innate to adaptive immunity [38], [39], but play a detrimental role if they accumulate in excess, as occurs with poorly controlled TB in mice and in humans [36], [40], [41]. NETs lack antimicrobial activity against Mtb in vitro [31]. Their potential to reduce Mtb viability in vivo is presently unknown. In neutrophilic TB lesions, NETs might promote lung injury as they were shown to do in a mouse influenza model [42]. It is interesting to consider what role NETs could play in forming a milieu that supports extracellular persistence of Mtb in necrotic lung lesions. A refined understanding of host-pathogen interactions in TB will require analysis of unique Mtb interactions with each of these phagocyte types in vivo, using cells isolated from the tuberculous lung.
Experiments with animals were conducted according to the National Institutes of Health guidelines for housing and care of laboratory animals and performed under protocols approved by the Institutional Animal Care and Use Committee and the Institutional Biosafety Committee at The University of Massachusetts Medical School (UMMS).
C57BL/6 WT, IFN-γ−/− (B6.129S7-Ifngtm1Ts/J) knockout mice (#2287), and ApoE−/− were purchased from The Jackson Laboratory. Mice were housed in specific pathogen-free environment at Animal Medicine facility of UMMS.
Mtb Erdman was used for in vitro and aerosol infections. Bacterial stocks for experiments were prepared as described previously [3]. For in vitro infections, BMDM were generated as previously described [2] and plated in Lab-Tek tissue culture chamber slides (Nalge Nunc International) at a density of 2×105 cells per well, or in 24-well cell culture plates at 5×105 cells per well in complete DMEM. Cells were infected with Mtb Erdman (MOI 25, 3 h, 37°C), washed with PBS and then overlaid with fresh complete DMEM. For aerosol infections, mice were exposed to Mtb in a Glas-Col Inhalation Exposure System set to deliver ∼100 CFU or ∼300 CFU to the lung. For each experiment, 2 mice were sacrificed 24 hours p.i. to verify the delivered dose as described.
Lung leukocytes were isolated as previously described [35]. Briefly, mice were sacrificed and lungs were perfused through the heart with PBS. Excised lungs were minced and digested with 150 U/ml collagenase IV and 60 U/ml DNase (Sigma-Aldrich; 45 min, 37°C). Processed tissues were filtered using a 40 µm cell strainer and treated with Gey's Solution (Sigma-Aldrich). BAL cells were collected by flushing lungs three times with 0.75 ml PBS containing 0.2% BSA and 0.2 mM EGTA, and added to1.0 ml of 20% FBS in PBS and placed immediately on ice. BAL fluid was washed in PBS and treated with Gey's Solution. Whole lung leukocytes and BAL cells prepared in this manner were fixed in 1.5% paraformaldehyde for overnight at 4°C. Fixed cell suspensions were washed, re-suspended in PBS and stored in 4°C. Cell counts were determined using a hemocytometer.
BAL cells and lung leukocytes were harvested from Mtb infected mice at predetermined time points. Slides were prepared using cytocentrifugation to immobilize 1×105 cells per slide (Thermo Electron Corporation). Cytospin slides were heat-fixed for Ziehl-Neelsen staining kit (TB Stain Kit ZN, BD Diagnostic Systems) following manufacturer's protocol. Stained slides were visualized using a Nikon Eclipse E400 Microscope and photomicrographs were obtained with a Nikon DS-Ri1 camera using NIS-Elements Microscope Imaging Software. Individual cells were interrogated for intracellular bacteria by counting AFB encased or surrounded by cellular membrane. Accurately counting intracellular AFB was reliable at low bacillary burden but became progressive more difficult in high burden cells with clumped bacilli. AFB counts were grouped into five bins: 1–5, 6–10, 11–15, 16–20, and ≥21. Cells were identified as monocytic cells (comprising AM, RM, mDC) or neutrophils based on nuclear morphology. AFB counts were tallied separately for these two categories.
To examine the nuclear morphology, cytospin slides were heat fixed and submerged in TB Carbolfuchsin ZN (BD Diagnostic Systems). Slides were heated in microwave oven for two consecutive intervals of 15 sec separated by 2 min at room temperature and then gently rinsed under running distilled water and decolorized with TB Decolorizer (BD Diagnostic Systems). Slides were rinsed again and then stained with 0.5 g/ml of 4′,6′-diamidino-2-phenylindole, dihydrochloride (DAPI) staining for 2 min. After a final rinse, slides were dried and cover slips mounted with ProLong Gold Antifade reagent (Invitrogen).
For immunostaining, BAL cells were affixed to Cell-Tak (BD Biosciences) treated cytospin slides and blocked with 3% BSA and 10% goat serum in PBS. Cells were stained with primary antibodies, 1∶50 myeloperoxidase (LS Bio) and 1∶50 histone H2B (Santa Cruz Biotechnology) or 1∶50 neutrophil elastase (Calbiochem). Fluorescent anti-rabbit antibodies conjugated to Alexa Fluor 488, 568, 594 or 647 (Invitrogen) were used as secondary antibodies. Cells were mounted and stained with DAPI with Prolong Gold Antifade Reagent with DAPI (Invitrogen). Analysis of immunostained cells were performed with confocal scanning laser microscopy (SP2 AOBS Leica) and images were captured using LCS software.
BAL cells and lung leukocytes were washed and incubated with CD16/CD32 mAb (BD Biosciences) to block Fc binding. Cells were then stained with the following mAb purchased from eBioscience (San Diego, CA): eFluor450–anti-CD11b (M1/70); phycoerythrin–anti-CD11c (N418); allophycocyanin–anti-CD45 (30-F11); APC-eFluor780-anti-Ly-6G (RB6-8C5); and Live/Dead Fixable Dead Cell Stain Kit by Invitrogen. An LSRII flow cytometer (BD Biosciences) was used for acquisition and data were analyzed with FlowJo software (TreeStar). Unless otherwise stated, gating was set to exclude dead cells and lymphocyte populations in forward/side scatter graph and to include singlet cells in a dot plot of pulse area against pulse height. Gating on viable cells, we defined resident AM as CD11b− CD11c+/hi cells, RM as CD11b+/lo CD11clo/− and mDC as CD11b+/hi CD11c+/hi (Fig. S1). Cells were sorted utilizing BD FACSAria Cell Sorter (BD Biosciences) with the same gating strategies used for flow cytometry. Subsets of sorted cell populations were collected and affixed onto cytospin slides for Ziehl-Neelsen staining and enumeration of intracellular AFB by light microscopy.
A replication incompetent, VSVG-pseudotyped, lentivirus expressing GFP under the control of a CMV promoter (CMV-GFP-W) was used to transduce resident lung leukocytes. The vector was created using a 5-plasmid transfection method previously described [20], [43]. Briefly, 293T cells were transfected with the pHAGE backbone lentiviral vector together with 4 expression vectors encoding the packaging proteins Gag-Pol, Rev, Tat, and the G protein of the vesicular stomatitis virus (VSV-G). To transduce lung cells, the viral titer was adjusted to 5×109/ml in DMEM with 10% FBS and mixed with lipofectamine 2000 (Invitrogen) at a ratio of 100∶5 (v∶v) on ice for 15–30 min. Mice were then infected by tracheal instillation of 5×107 virions in a volume of 50 ul.
Samples of non-adherent cells infected with Mtb were processed by first preparing microscope slides with Cell-Tak (BD Biosciences). Cell suspensions were added to treated and dried slides by cytocentrifugation and allowed to bond to the Cell-Tak. The cells on the slides were fixed by immersion in 2% paraformaldehyde (v/v)/2.5% glutaraldehyde (v/v) in 0.1 M Na cacodylate-HCl buffer (pH 7.2) overnight at 4°C. The next day the fixed samples were washed three times in 0.5 M Na cacodylate-HCl buffer (pH 7.0) and then post-fixed for 1 hr in 1% osmium tetroxide (w/v) in the same buffer. Following post-fixation, samples were dehydrated through a graded series of ethanol to two changes of 100% ethanol and critical point dried in liquid CO2. The microscope slides were cut to remove the excess glass, mounted onto aluminum stubs with silver conductive paste and then coated with carbon (1 nm) and then sputter coated with gold/palladium (4 nm). Specimens were then examined using an FEI Quanta 200 FEG MK II scanning electron microscope.
A 2-dimensional (2D) agent-based model (ABM) framework developed [44]–[46] for spatially characterizing the mechanisms of immunity in the lung during TB infection was used to test the burst size concept. The virtual environment reflects a 2 mm×2 mm section of lung parenchyma represented as a 100×100 2D grid with micro-compartments scaled to the approximate size of a macrophage (∼20 µm). A virtual low dose infection is triggered by one infected macrophage (MI), with one intracellular Mtb. The ABM describes interactions between intracellular and extracellular Mtb, various states of macrophages (resting, infected, chronically infected and activated), T cell populations including CD4+, CD8+ and regulatory T cells along with major cytokines tumor necrosis factor-α and IFN-γ and chemokine effector molecules (e.g., CCL2, CCL5, CXCL9/10/11). Each immune cell's behavior adapts based on its environment and its interactions with other immune cells and Mtb. As infection progresses, Mtb is tracked continuously. Extracellular Mtb proliferation follows a logistic growth function (48 h doubling time) within a single micro-compartment with a given carrying capacity while intracellular Mtb follows an exponential growth curve with a doubling time of 24 h. Intracellular Mtb doubling time is set to 72 h after adaptive immunity appears at the infection site (i.e., 20 days p.i.).With the current mechanisms already present in our model that can affect Mtb levels, we captured the process of burst size cytolysis by setting a maximum carrying capacity for a chronically infected macrophage. If the intracellular bacterial load exceeds this threshold, the macrophage bursts releasing viable bacteria into the extracellular space. The model allows a user defined parameter value for this threshold, labeled as burst size. For this study, we varied the burst sizes from 10 to 50 while maintaining fixed values for the remaining parameters to analyze affects of different burst sizes on the total lung bacterial burden.
Uncertainty and sensitivity analysis (U/SA) [47] has been used in this model to ensure that the selected parameter values influencing outcomes of infection (e.g. clearance, containment or dissemination) are in accordance with known dynamics. The results of U/SA analysis provide constructive evaluation of the critical processes and mechanisms suggesting strategies for model reduction, questions requiring additional in vivo experimentation, and to generate alternative hypotheses if burst size is not supported by model results [48].
Unless otherwise stated, data from independent experiments are shown as mean ± SD or SEM. Comparisons between groups were evaluated with Student t-test using GraphPad Prism. Differences in the distribution of AFB load in frequency bins obtained from experiments with GKO and WT mice evaluated using analysis of variance for mixed model [49] with Restricted Maximum Likelihood (REML) algorithm [50] for fitting the model. Load data were transformed using natural logarithms to better approximate normally distributed errors, an assumption of the mixed model ANOVA. The distributional characteristics of the data were evaluated using the Kolmogorov-Smirnov goodness of fit test [51] upon model residuals. A p value<0.05 was regarded as statistically significant.
In the computational model, standard unidirectional t-test, with heteroscedasticity assumption (i.e., different variability between groups) was used to test statistically significant differences (p<0.05) between time course predictions with different burst sizes at different time points.
|
10.1371/journal.pntd.0005837 | Community effectiveness of indoor spraying as a dengue vector control method: A systematic review | The prevention and control of dengue rely mainly on vector control methods, including indoor residual spraying (IRS) and indoor space spraying (ISS). This study aimed to systematically review the available evidence on community effectiveness of indoor spraying.
A systematic review was conducted using seven databases (PubMed, EMBASE, LILACS, Web of Science, WHOLIS, Cochrane, and Google Scholar) and a manual search of the reference lists of the identified studies. Data from included studies were extracted, analysed and reported.
The review generated seven studies only, three IRS and four ISS (two/three controlled studies respectively). Two IRS studies measuring human transmission showed a decline. One IRS and all four ISS studies measuring adult mosquitoes showed a very good effect, up to 100%, but not sustained. Two IRS studies and one ISS measuring immature mosquitoes, showed mixed results.
It is evident that IRS and also ISS are effective adulticidal interventions against Aedes mosquitoes. However, evidence to suggest effectiveness of IRS as a larvicidal intervention and to reduce human dengue cases is limited–and even more so for ISS. Overall, there is a paucity of studies available on these two interventions that may be promising for dengue vector control, particularly for IRS with its residual effect.
| The effectiveness of indoor residual spraying (IRS) and indoor space spraying (ISS) as dengue vector control methods depends on many factors. This study aims to systematically review the evidence on the community effectiveness of indoor spraying of insecticides to reduce Aedes mosquito populations and thereby to control dengue transmission. A systematic literature review was performed in PubMed, EMBASE, LILACS, Web of Science, WHO library database (WHOLIS), Cochrane, and Google Scholar, including a manual search of the reference lists of the identified studies since its inceptions until 15.02.2017. A total of 39 articles were retrieved for full assessment. Seven studies were included and analysed after final application of inclusion and exclusion criteria: two IRS studies with control, one without, three ISS studies and one, respectively. One IRS study and four ISS studies showed good evidence of effectiveness on adult Aedes mosquitoes. Evidence of effectiveness of IRS as a larvicidal intervention exists but is still inadequate, and is weak for ISS. Evidence of effectiveness of IRS on human dengue cases as a single intervention exists, but was limited and not available for ISS. It is recommended to scale up the research regarding the community effectiveness of IRS and ISS, including measuring dengue transmission, particularly, for IRS with its residual effect. It is also suggested to study in depth the factors that could affect the community effectiveness of IRS and ISS on Aedes populations and on human dengue cases.
| Dengue is the most prevalent arthropod-borne viral disease, infecting 300 to 500 million individuals each year. Approximately 100 million infections are symptomatic, which can range from mild to severe disease [1,2,3]. An estimated 500 000 people suffer from the severe forms, nearly 90% of whom are children, with a resulting 22 000 dengue-related deaths annually [4]. Global climate change, urbanisation, travel, poor sanitation, and inadequate public health services, all have the potential to increase the intensity of dengue transmission [5,6].
The four serotypes of the dengue virus (DENV 1–4) are transmitted principally by female Aedes aegypti and to a lesser extent by Aedes albopictus mosquitos [7]. Aedes species are anthropophilic, feed in the dark, the early morning and twilight hours and show an indoor-resting behaviour preferentially in secluded stationary locations e.g. under furniture, lower walls, under sinks, in curtain folds, or in wardrobes [2,8,9]. Dzul-Manzanilla [10] determined that Aedes aegypti rested mostly below 1.5 meters of height, and mostly in bedrooms (44%), living rooms (25%) and bathrooms (20%).
At present, there is no effective vaccine available, for public health use, to prevent or treat dengue infections, efficacy of the existing vaccine is variable and not high [11,12]. Therefore, vector control is the primary method of dengue prevention and control. Since the turn of the 19th century, chemical insecticides applied to the environment in a variety of methods have served as one of the mainstays of dengue vector control programmes, basically outdoors against immature and indoors-outdoors against adult vectors.
Indoor application of insecticides (IAI) includes indoor space spraying (ISS) or indoor residual spraying (IRS). Both target the endophilic adult Aedes mosquitoes that bite and rest indoors [10,13]. IRS entails the coating of walls and surfaces of the entire house with a residual insecticide [14]. ISS is done to treat indoor spaces to control flying insects with less residual effect. IRS can potentially target Aedes aegypti as it was used for the first time in Malaysia in 1952 [15]. IRS, however, is not generally recommended for dengue vector control, as it is thought that adult Aedes aegypti often rest on non-sprayable surfaces in houses [16]. Despite this, reductions in Aedes aegypti populations have been observed in areas where IRS is utilised for malaria control. A recent meta-analysis [17] concluded that there is a need of more empirical evidence supporting the potential utility of IRS for dengue prevention, since it was based on only two studies. For a meta-analysis comparability of studies precludes inclusion of many articles, thus providing a justification with an update and further inclusion and analysis of studies using IRS/ISS with a further systematic review.
This study systematically reviews the available evidence on community effectiveness of IRS and ISS for reducing Aedes populations and thereby for controlling dengue transmission.
This review follows the guidelines set forth in the PRISMA criteria for the reporting of systematic reviews and meta-analyses [18]. The literature search was conducted in parallel by two data extractors until 15.02.2017, with an update until 28.02.17.
A wide range of search terms was used in combinations to identify all relevant studies. The search terms included (a) disease specific terms: Dengue, Dengue hemorrhagic fever, Dengue haemorrhagic fever, Dengue shock syndrome, DHF, and DF, (b) vector specific terms: Aedes, Aedes aegypti, Aedes albopictus, Ae. aeygpti, and Ae. albopictus, and (c) intervention specific terms: Indoor space spray, ISS, Indoor residual spray, IRS, Residual house spray, and Intra domiciliary residual spray.
For the purposes of this review, IRS was defined as the application of chemical insecticides on walls and other surfaces with the aim to control Aedes mosquitoes inside houses, using substances which remain effective for 1 month or more. ISS was defined as any indoor spray using ultra-low volume spray (ULV), low-volume spray (LV), thermal fogging and other devices such as insecticide fumigant canisters. This review is limited to public health application of IRS/ISS, not commercial (household) use. Community effectiveness studies were defined as those studies conducted to evaluate the impact of IRS/ISS under normal field conditions, while efficacy studies were defined as those studies conducted under laboratory conditions.
The above search strategy was applied to the following databases: PubMed, EMBASE, LILACS, Web of Science, WHO library database (WHOLIS), Cochrane, and Google Scholar.
Eligible studies met the following inclusion criteria: 1) peer-reviewed publications presenting original data evaluating the community effectiveness of IRS/ISS 2) studies with control group(s) during intervention, studies with pre- and post-intervention assessments, and cross-sectional studies, 3) no language restrictions were applied, and 4) the target was vector and human populations. The exclusion criteria were limited to the following: 1) abstracts, conference posters, short communications, and letters to the editor, 2) studies with not enough information on community effectiveness of IRS/ISS, 3) efficacy studies and 4) surveillance data or reviews.
All identified studies were screened by title and abstract. Relevant studies were sent to EndNote X7 reference manager software. The numbers of relevant, irrelevant, and duplicated articles were identified and recorded for each database.
Full texts of selected studies were retrieved either through online databases or through Heidelberg University libraries. All reference lists of retrieved studies were screened for additional relevant studies. The full eligibility criteria were applied to all retrieved articles to identify the final list of included studies. The systematic literature search and the review followed the assessment of multiple system reviews, AMSTAR, for assuring the methodological quality [19].
By using a pre-designed extraction sheet, the following data were extracted from each included study: author name, year of publication, source database, study title, geographical location, objective(s) of the study, study design, relevant outcomes, main results, and key conclusions of the authors (Table 1).
To assess for quality, included studies were categorised into studies with and without a control arm. They were further classified by study design and number of interventions. Outcome measures were extracted, classified and summarised across studies. Different measures were used to record the frequency of observations and the way of presentation varied according to the type of the presented data. Different insecticides and methods of application, together with varying statistical methods and outcome measures across the studies precluded any attempt at meta-analysis. Articles included through an update of the initial searches, until 28.02.17, are presented in the discussion section.
A comprehensive literature search of the seven databases identified 825 potentially relevant citations. After screening for title and abstract, 144 duplicates and 649 irrelevant articles were excluded. The reference lists of the 32 remaining articles added seven more studies. The 39 studies were retrieved for full text assessment. Upon meeting eligibility criteria, seven studies were included and 32 studies were excluded, most of the latter were efficacy studies only (Fig 1). Summaries of included studies were arranged chronologically in an evidence table (Table 1).
Seven studies met the pre-specified eligibility criteria (1) Three IRS studies: Parades-Esquivel 2015 [20], Vazquez-Prokopec 2010/1 [21], Lien 1994 [22]; 2) Four ISS studies: Mani 2005 [23], Perich 2003 [24], Perich 2001 [25] and Koenraadt 2007 26]). Most dengue risk areas were represented except Africa, with three studies from Asia [22,23,26], one study from Australia [21] and three from Latin America and the Caribbean [20,24,25]. All articles were reported in English. The time period of publication ranged from 1994 to 2015.
The seven studies were broadly classified into five controlled studies, two for IRS and three for ISS, and two non-controlled studies (one each IRS/ISS) (Table 1). Controlled studies were subsequently classified into four intervention control studies all testing IRS [20] and ISS [23,24,25]with multiple study arms. One cross-sectional time series compared data from sprayed and non-sprayed areas [21]. For the two non-controlled studies [22,26], Lien [22] had one study arm only, Koenraadt [26] had multiple study arms.
Reporting on sample size varied across included studies either for the diversity of methods or for the unavailability of data in some studies. For controlled studies, the smallest sample size for intervention was 36 houses [20] and the biggest three residential colonies with 216–260 houses each. The non-controlled studies covered 36977 houses [22] and four houses in two areas [26].
All included studies reported on geographical locations. Parades-Esquivel [20], Vazquez-Prokopec [21], Mani [23] reported on meteorological conditions. All studies reported on the season/time period of the study, in relation to dry and rainy seasons. All studies discussed factors that might influence dengue transmission, and mosquito abundance, such as ecology and housing structures. The latter is described in detail by Parades-Esquivel [20], Vazquez-Prokopec [21] and Perich [24,25]. Parades-Esquivel [20] and Mani [23] present target populations and their socio-economic background. Pre-intervention dengue estimates were reported by Vazquez-Prokopec [21] and Lien [22]. Vazquez-Prokopec [21] reported on previous dengue outbreaks.
Method(s) of intervention: All studies used IRS as a single intervention. Although Vazquez-Prokopec [21] compared data from areas sprayed and not sprayed with IRS, in addition to the ongoing local control programme, including control of breeding places. Mani [23] compared ISS and peridomestic spraying, Perich [24] compared ISS with ULV, LV and thermal fogging and Perich [25] compared ISS with ULV and thermal fogging. Koenraadt [26] compared ISS and peridomestic spraying, with different insecticide concentrations.
Forms of application and formulations: Forms of application varied considerably, but including either ultra-low volume spray (ULV), thermal fog spray, or low-volume spray (LV). Formulations varied as well, including deltamethrin [20], lambda-cyhalothrin [21,24,25], alphacypermethrin [22], pyrethrin [26] and deltacide, a mixture of Deltamethrin 0.5%, S-Bioallethrin 0.75% and Piperonyl Butoxide 10% [23].
Duration of residual effect: Paredes-Esquivel [20] estimates a good residual effect of IRS up to 16 weeks, for ISS Perich [24,25] demonstrated three weeks and four weeks’ residual effect, respectively. Mani [23] and Koenraadt [26] showed a residual effect of one week.
Five of seven studies incorporated a control group into the study. They were assigned in different ways according to the methods used in each study. 1) IRS studies: Parades-Esquivel [20] used three single houses with similar structures to the 3 clusters of intervention houses (12 each). Vazquez-Prokopec [21] compared 97 sprayed houses to 151 non-sprayed houses, as the data were retrospectively available. 2) For ISS studies: Mani [23] used one cluster of houses (216–260) of three clusters for control. Perich [24] used two residential blocks with 12 untreated houses as control, Perich [25] used one residential bock with 6 untreated houses.
A variety of entomological and disease specific outcome measures were used to assess the impact of IRS: 1) Measures for adult Aedes: Adult mosquito mortality and knock down (KD) rates[20,23,24,25]; Adult mosquito density [20, 23,24,25,26] and spatial and temporal patterns [26]; 2) Measures for immature Aedes: Breteau Index (BI) [20,22,23]; House Index (HI) [20,22]; Percentages of breeding site [23]; Number of parous females [26]; 3) Disease specific measures: Age adjusted dengue incidence [21]; Odds of secondary dengue infection [21]; reported number of cases [22].
The effect of indoor spraying of insecticides on adult mosquitoes is strong immediately after application in all studies measuring these parameters. For IRS studies, in Peru [20] the Adult Index fell from 18.5 to 3.1 four weeks’ after intervention (p < 0.05). For ISS studies, adult mortality percentage reduction was 100% post indoor spraying, 77.8% on day 5, 6.25 on day 7[23]. Similarly, adult density dropped to 0 after spraying with thermal fog and ULV, increasing after day 7 and continued to increase until 7 weeks post spraying, with similar results in Costa Rica [24] and Honduras [25]. In an uncontrolled setting in Thailand [26], indoor spraying reduced the number of adult mosquitoes to around 10%, however gradually recovering after day 2. The latter study measured also that there was a relationship between mosquito density and distance to the centre of application with an area of protection extending to 85 m. Parity rates also dropped after spraying.
The effect on immature mosquitoes is less strong on all studies measuring larval indices. For IRS studies, deltamethrin in Peru reduced all immature indices in the first week and sustained throughout the period of studies [20]. Also, there was a noted reduction of BI from 35 to 5 in Taiwan [22]. However, for ISS, in India, with a BI of 50 at baseline, this reduced to 29.6 post 7 days, and recovered post 14 days to 37.5.
For human dengue infection parameters, there are only two IRS studies. Odds of dengue infection shown by Vazquez-Prokopec [21], in Australia, were significantly higher at unsprayed than at sprayed premises (OR = 2.8; 95%CI = 1.1–6.9; p = 0.03). When 60% of the premises were sprayed around the index case house the odds reduced significantly to zero. Also the number of dengue cases was strongly and positively correlated to the number of IRS applications (r >0.6). Also, in Taiwan [22], the number of cases reported over time, dropped with IRS applications from above 3000 to 1000 (no control).
The evidence presented here suggests that IRS and ISS can be an effective dengue control intervention. The majority of included studies demonstrated a significant post-intervention reduction in adult and some effect on immature Aedes populations. Notably, of the studies that measured dengue incidence, both showed decreases in new dengue cases after the application of IRS. These findings support the use of IRS as a component of integrated vector management (IVM) [27], and perhaps ISS as well.
While the differing methodologies and interventions precluded meta-analysis, the included studies consistently show effective killing of adult Aedes mosquitos almost immediately after application of IRS and ISS. Estimates of the duration of effect are limited by the relative short time-frames studied, but multiple studies reported residual efficacy up to two months post-intervention. The impact of IRS on the incidence of dengue may be of even longer duration. The impact of ISS on dengue transmission was not measured. Further confirmations of the effect of IRS and ISS arise by two further studies [28,29]–the studies focused however on other elements and were excluded in the analysis. Ritchie [28] noticed an effect that started late but continued, using a combination of containers treated with S-methoprene or lambda-cyhalothrin and adult control with IRS using lambda-cyhalothrin, “human cases subsequently dropped from a high of seven cases per day in mid-March to only sporadic cases in late April, with the final reported onset of 7 May”. Stoddard [29] analysed surveillance data of dengue for explanatory models of transmission, ISS delivered in three cycles, using deltamethrin, cypermethrin, or alpha-cypermethrin, resulted in a good reduction of dengue transmission in trimester III. An update of the searches generated a further article, published shortly after the initial searches [30]. The authors conducted a study using space-time statistical data modelling from Cairns, Australia (data from 2008 and 2009). Targeted IRS “in potential exposure locations reduced the probability of future DENV transmission by 86 to 96%, compared to unsprayed premises”. This study strongly confirms the potential of IRS for reducing dengue transmission.
While there is evidence for indoor spraying in the control of dengue, there are a number of challenges with scaling up such interventions. Since, indoor spraying can require high levels of coverage, which requires widespread community acceptance and participation. Few studies included in the review reported qualitative estimates of community acceptance, although IRS is often popular as it has the ancillary benefit of killing many nuisance insects [1,4]. However, Chang [31] emphasised how communities are still reluctant to take appropriate dengue control measures. Furthermore, Gürtler [32] suggested integrating sustained social participation into IVM activities like source reduction, biological control, and environmental management, in order to overcome such a challenge and to ensure long-term sustainability of dengue prevention and control.
In addition, none of the included studies examined the associated costs of indoor spraying. In Australia however, where IRS is used for dengue control, a cost-analysis shows that the total costs of preparedness through surveillance are far lower than the ones needed to respond to the introduction of vector-borne pathogens [33]. Universal application and re-application is likely beyond the resources of many dengue-affected countries. Therefore, effective use of indoor spraying will require timely surveillance and response mechanisms. Combination of effective early warning systems with vector control measures could reduce densities of Aedes and subsequently dengue transmission [34]. Response systems could include mapping technologies like GIS [35]. Using space-temporal units besides such technologies is essential in delivering the resources and in measuring the coverage [36]. Analysis of one of the included studies showed similar evidence on how early detection of dengue outbreak helped to implement rapid and effective control actions, including early use of residual pesticides [22].
Experiments emphasised an association between type of insecticide used and its residual effect on Aedes and showed how the susceptibility of mosquitoes differs from one insecticide to another [37,38]. Perich [24,25] reported on another factor, which was the droplet size and linked it to post-spray residual effect. Sulaiman [39] pointed out how applying IRS on wooden surfaces is potentially controlling dengue. Another efficacy study in Malaysia linked house construction to the residual activity of IRS, since its wall bioassays indicated that both Ae. aegypti and Ae. albopictus were more susceptible to IRS on wooden surfaces than on brick surfaces [40]. Other challenges that are not well addressed in the included studies are optimal application and insecticide resistance, the latter is of a growing concern. Resistance particularly may affect severely the effectiveness of IRS/ISS. For residual treatment for example a study in Brazil showed a mortality of only 10% for Aedes in some communities for Deltamethrin [41].
A further challenge is the application of IRS, and where IRS is targeted. Whereas for Malaria and transmitting vectors IRS is defined as “the application of insecticide to the inside of dwellings, on walls and other surfaces that serve as a resting place for malaria-infected mosquitoes” and conditions for the use of IRS are set as “1) Majority of vectors (i.e., organisms that transmit malaria) must feed and rest indoors 2) Vectors are susceptible to the insecticide in use, 3) Houses have “sprayable” surfaces and 4) A high proportion of the houses in target areas are sprayed (more than 80 percent)” [42], such conditions are not as clear set for dengue vectors.
In addition to routine control measures, the use of indoor spraying as an emergency response is also feasible. Perich [24] pointed out how ISS successfully fulfils the criteria to be used as an emergency operation, which were: 1) providing an initial kill of adult Aedes, and 2) allowing a significant level of residual activity. Although residual activity with ISS may be mixed up with a time lag in recovery of mosquito populations. Evidence from that study and other efficacy studies in Malaysia and Taiwan plus ineffectiveness of outdoor spraying to control indoor Aedes populations make indoor spraying a true effective alternative for emergency suppression of Aedes mosquitoes [22,23,24,25,39,43]. This may also include the use of household (commercial) insecticides, another field that warrants analysis.
The key limitation of this systematic review is the very limited number of studies that typically researched community effectiveness of IRS and ISS. This study reports therefore on the different forms of application in relation to the outcomes. Also, potential publication and selection bias are most concerning. It is well documented that studies with positive outcomes are more often reported in literature than negative outcomes. The diversified and extensive search strategy along with no restrictions in languages should minimise the publication and selection bias.
The findings must also be interpreted with regard to the quality of the included studies: 1) Different methodologies, 2) Different study settings, 3) Limited use of statistical methods to assess for significance/control for confounding, 4) Relatively short study periods and 5) Lack of randomisation in most studies, influence the results.
However, the review is the most comprehensive to date and highlights the need for future work in this area. Concluding, evidence obtained from this systematic review showed that the use of IRS and ISS can produce significant reductions of Aedes populations (adult and immature forms). IRS can also produce significant reductions in human dengue cases, with very limited available evidence, but no data are available for ISS. However, evidence to suggest the effectiveness of IRS/ISS either on immature and adult stages of Aedes or on human dengue cases as a single intervention is limited.
The community effectiveness of IRS is affected, directly and indirectly, by many factors. Examples for these factors are disease epidemiology, virus dynamics, human movements, effective surveillance systems, community participation in vector control, the insecticides used, particularly considering insecticide resistance, environmental factors, and house construction. When these factors work in harmony with IRS/ISS applications, they would maximise its community effectiveness. Moreover, they could maximise the applicability of IRS/ISS, also being used as an emergency control measure during epidemics instead of being just applied as a routine control measure.
|
10.1371/journal.pgen.1000789 | TGF-ß Sma/Mab Signaling Mutations Uncouple Reproductive Aging from Somatic Aging | Female reproductive cessation is one of the earliest age-related declines humans experience, occurring in mid-adulthood. Similarly, Caenorhabditis elegans' reproductive span is short relative to its total life span, with reproduction ceasing about a third into its 15–20 day adulthood. All of the known mutations and treatments that extend C. elegans' reproductive period also regulate longevity, suggesting that reproductive span is normally linked to life span. C. elegans has two canonical TGF-ß signaling pathways. We recently found that the TGF-ß Dauer pathway regulates longevity through the Insulin/IGF-1 Signaling (IIS) pathway; here we show that this pathway has a moderate effect on reproductive span. By contrast, TGF-ß Sma/Mab signaling mutants exhibit a substantially extended reproductive period, more than doubling reproductive span in some cases. Sma/Mab mutations extend reproductive span disproportionately to life span and act independently of known regulators of somatic aging, such as Insulin/IGF-1 Signaling and Dietary Restriction. This is the first discovery of a pathway that regulates reproductive span independently of longevity and the first identification of the TGF-ß Sma/Mab pathway as a regulator of reproductive aging. Our results suggest that longevity and reproductive span regulation can be uncoupled, although they appear to normally be linked through regulatory pathways.
| Female reproductive cessation is the earliest aging phenotype humans experience, and its importance as a clinical issue is growing as more women opt to have children later in life. While much work has been done to understand the general aging process, little is currently known about the regulation of reproductive aging. Like longevity, the ability to produce progeny with advanced age is likely to be genetically regulated. Thus, understanding the processes that regulate reproductive aging may allow us to address the problems of maternal age-related infertility and birth defects. C. elegans and humans both have long post-reproductive life spans, leaving open the possibility that their reproductive spans might be extendable. C. elegans has been used previously to discover conserved regulators of aging, and here we use worms to identify a new regulator of reproductive aging, a highly conserved TGF-ß signaling pathway. We find that TGF-ß signaling regulates reproductive aging independently of somatic aging. This is the first identification of a pathway that breaks the coupling that normally links the two processes. Our work will provide new insights into the improvement of human fertility and prevention of age-related birth defects, and it has implications for the evolutionary relationship between reproduction and longevity regulation.
| Among human age-related declines, female reproductive cessation is one of the earliest to occur, with infertility and maternal age-related birth defects arising during the fourth decade of life [1]. While artificial reproductive technologies have improved late conception success [2]–[5], the underlying molecular regulators of reproductive cessation remain largely unknown. Like longevity, the ability to produce progeny with advanced age is likely to be genetically regulated. Thus, understanding the processes that regulate reproductive aging may allow us to address the problems of maternal age-related infertility and birth defects.
Although C. elegans produces large broods of progeny and does not care for its young, its reproductive schedule is similar to human females' in that its fertility and reproduction sharply decline in early/mid-adulthood, followed by a long post-reproductive period [6],[7]. The similarities between C. elegans and human reproductive schedules suggest the intriguing possibility that studies in this model organism may reveal mechanisms regulating reproductive cessation across species.
While many C. elegans studies have focused on reproductive fitness, measuring total numbers of offspring produced and average generation time, we are instead interested in identifying regulators of late-life reproduction (i.e., the ability of individual mothers to continue to reproduce viable progeny as they age). The standard assessments of fitness and fertility do not ascertain the length of time that an individual is capable of successful reproduction. However, we and others [6]–[8] have recently become interested in the latter aspect of reproduction, because of the obvious possible parallels with human reproductive aging, in particular, Advanced Maternal Age (AMA) and its related clinical problems. In other words, these model organism studies of reproductive aging are focused on determining the capacity to reproduce successfully late in life, rather than on total progeny production or evolutionary fitness. Hughes, et al. recently showed that worms undergo reproductive aging, a process that is dependent neither on tissue wear (as manipulation of early progeny number had no influence) nor on sperm availability [7]. Thus, the reproductive system of C. elegans ages significantly during the first week of adulthood, which is also reflected in the degree of germ line degeneration and oocyte quality decline [8],[9]. This germ line aging results in reproductive cessation days to weeks prior to death and a relatively long post-reproductive life span, similar to human females' long post-reproductive life span.
The C. elegans mutants currently known to delay reproductive aging were originally identified through their longevity phenotypes [7]. These longevity mutants include the insulin/IGF-1 receptor mutant daf-2 [7],[10],[11], and a model of Dietary Restriction (DR), eat-2 [6],[12]. Insulin/IGF-1 signaling (IIS) and FOXO transcription factor activity have been implicated in the regulation of reproduction in several other organisms, including Drosophila [13], mice [14],[15], and humans [16]. Life span extension and slowing of reproductive activity are also hallmarks of Dietary Restriction. Dietary Restriction reduces progeny number and lengthens the reproductive period of C. elegans hermaphrodites [7], female Drosophila [16], and female rodents [18],[19]. C. elegans shifts its reproductive strategy after starvation, modifying its production of males and its outcrossing frequency [17], and many animals adjust their reproductive life span in response to predation levels [18]. Together, these data indicate that the reproductive schedule is flexible, poised to respond to environmental and molecular perturbations, and that the mechanisms regulating reproductive aging, like longevity, are likely to be evolutionarily conserved.
C. elegans has two highly conserved Transforming Growth Factor-ß (TGF-ß) signaling pathways, the Dauer (daf-7) and Sma/Mab (dbl-1) pathways. We recently found that TGF-ß Dauer signaling regulates lifespan through its interactions with the Insulin/IGF-1 Signaling (IIS) pathway [19]. Members of the TGF-ß Dauer pathway include the ligand DAF-7, receptor heterodimers DAF-1 and DAF-4, the R-Smads (receptor-regulated Smad signal transducer) DAF-8 and DAF-14, the Co-Smad (common-mediator Smad) DAF-3, and the transcription factor DAF-5 (Figure 1A; mammalian homologs are shown in Figure S1A) [20],[21]. DAF-4 is a type II receptor that is shared between the Dauer pathway and a second TGF-ß pathway, the Sma/Mab pathway (Small body/Male tail abnormal). The Sma/Mab pathway includes the ligand DBL-1, the type I receptor SMA-6, SMA-2 and SMA-3 R-Smads, the Co-Smad SMA-4, and the SMA-9 transcription co-factor (Figure 1E; mammalian homologs shown in Figure S2A) [22]–[25].
Here we show that the Dauer pathway has a moderate effect on reproductive span, mediated at least in part by insulin/FOXO activity. More importantly, we have found that its shared member with the TGF-ß Sma/Mab pathway, daf-4, and the entire TGF-ß Sma/Mab pathway, strongly influence reproductive aging. Reduced TGF-ß Sma/Mab signaling extends reproductive span disproportionately to life span, and is genetically independent of known longevity regulators. The TGF-ß Sma/Mab pathway is a novel regulator of reproductive aging, and the first regulator of reproductive aging to be identified independently of somatic aging regulation. Our results demonstrate that somatic aging and reproductive aging can be uncoupled, suggesting that different molecular mechanisms underlie the two processes, but may normally be linked.
In addition to its regulation of dauer formation [26], we recently found that the TGF-ß Dauer pathway (Figure 1A, Figure S1A) regulates longevity [19]. However, whether this pathway also plays a role in the regulation of reproductive aging is unknown. To analyze the effect of TGF-ß Dauer mutants on reproduction, we determined the proportion of adults capable of progeny production as a function of age. The “reproductive span” calculated from such assays (see Materials and Methods) reflects the period of time animals produce viable progeny, as described previously [6],[7]. We found that members of the TGF-ß Dauer pathway moderately extended reproduction (Figure 1B; Figure S1B, S1C): while wild type's mean reproductive span was ∼3.5 days, the means of daf-7, daf-1, daf-8, and daf-14 mutants were 4–5 days, extensions of 25–50% (Table S2; Figure S1D). In addition, their maximum reproductive spans were ∼1 day longer than wild type's. At least part of the moderate reproductive span extension is likely a result of delayed onset of reproduction due to an egg-laying (Egl) defect (Figure 1D) [19],[27],[28]; by the end of daf-7's reproductive span, many progeny hatched into L1 larvae immediately upon being laid, as opposed to the typical 12–16 hour hatching time of wild-type eggs. daf-7 progeny do not develop into adults faster than wild type, so the advanced developmental stage of the progeny is likely due to egg retention in the mother.
Unlike the moderate reproductive span extensions of other TGF-ß Dauer pathway mutants, daf-4's ∼8 day mean is more than double the reproductive span of wild-type animals (Figure 1C; Figure S1D; Table S2). daf-4 mutants continued to steadily produce progeny for several days after reproductive cessation in wild-type animals, and its maximum reproductive span was 4–5 days longer than wild type (Figure S1G). This dramatic difference cannot be explained by the egg-laying defect typical of the TGF-ß Dauer pathway mutants, which extends reproductive span a maximum of one day. daf-4 encodes C. elegans' sole ortholog of the type II TGF-ß co-receptor, and is utilized by both the Dauer pathway and a second TGF-ß pathway, the Sma/Mab pathway (Figure 1E, Figure S2A) [22]–[25]. The large reproductive span extension that we observed in daf-4 animals, but not other TGF-ß Dauer mutants, suggested the possibility that the Sma/Mab pathway might be important in the regulation of reproductive aging.
We measured the reproductive spans of seven alleles of TGF-ß Sma/Mab pathway mutants (Figure 1E, Figure S2A), and found that decreased TGF-ß Sma/Mab signaling indeed increased reproductive span significantly: similar to daf-4, the mean reproductive spans of sma-2 and dbl-1 were over 7 days, compared to ∼3.5 days in wild type; the rest of the mutants in this pathway (sma-3, sma-4, sma-6, and sma-9) also increased reproductive span substantially (Figure 1F and 1G; Table S1; Figure S2B, S2C). The hatching rates of Sma/Mab mutants were comparable to wild type (Figure 1D), the onset of progeny production was not delayed, and progeny were steadily produced beyond the age when wild-type reproduction ceased (Figure S3A, S3B, S3C). Similar to daf-2 and eat-2, mutants that also extend reproductive span [7], Sma/Mab mutants produce fewer total progeny over a longer period of time (Table S3; Figure S3). The reproductive span extensions and progeny production profiles of the Sma/Mab mutants contrast with the delayed onset and sharp decline in the number of progeny produced after peak reproduction by the TGF-ß Dauer mutants (Figure S1E, S1F), suggesting that the Dauer and Sma/Mab mutants are distinct in their reproductive aging phenotypes.
Sma/Mab mutants exhibited a highly penetrant late egg-laying defect and internal hatching (matricide, or “bagging”) at the end of their reproductive period, in contrast to the Dauer mutants' very early egg-laying (Egl) and bagging phenotypes (Figure 1H, Figure S2E). In fact, several assays were terminated when a large fraction of the worms were still reproductive, due to the Sma/Mab mutants' late matricide phenotype (see asterisked sma-4 in Figure 1F and sma-3 in Figure 1G). It is likely that the full late reproductive capacity of the Sma/Mab mutants is masked by this late matricide defect. Thus, while the TGF-ß Dauer mutants' delayed onset of reproduction and Egl phenotypes may account for part of their moderate reproductive span increases (a maximum of one day), these factors are not likely the cause of TGF-ß Sma/Mab mutants' dramatic reproductive span extensions.
C. elegans hermaphrodite sperm number limits wild-type self-fertilized reproduction, but mating with young (day 1 adult) males, whose sperm are not limited and outcompete those of the hermaphrodite, increases and prolongs progeny production [29]–[31]. To rule out the possibility that the reproductive span extensions we observed in TGF-ß Sma/Mab mutants are due to increased or extended sperm production or utilization, we mated Sma/Mab mutant hermaphrodites with young wild-type males. We found that Sma/Mab mutants significantly and consistently increased mated reproductive span, from wild type's mean mated reproductive span of ∼6.0 days to a mean of 10–11 days (Figure 2A–2D; Table S1). In fact, Sma/Mab mutants were usually still fertile through Day 12–13 of adulthood, compared to the complete cessation of reproduction by Day 8–9 in wild-type animals. (Interestingly, when mated with wild-type males at an older age (day 4), the Sma/Mab mutants still had significant reproductive span extensions (Figure S2D), further supporting the notion that sperm quality and number are not the limiting factor, as shown previously [7].)
To further eliminate the possibility of sperm contribution, we also tested feminized (fem-1) mutant hermaphrodites, which fail to make sperm when raised at restrictive temperatures. fem-1 mutants mated with wild-type males have a mean reproductive span of 6.3 days, while more than 90% of the mated sma-2;fem-1 double mutants were still fertile at day 12 (Figure 2E). Notably, the self-fertilized reproductive spans of Sma/Mab mutants are even longer than wild type's mated reproductive span, highlighting the extreme extensions shown by Sma/Mab mutants (compare Figure 1F and Figure 2; Table S1). Additionally, neither the self-fertilized nor the mated Sma/Mab mutants delay the onset of progeny production, and both continue to produce progeny steadily beyond the age of wild-type reproductive cessation (Figure S3).
In self-fertilized worms, sperm is only made prior to oogenesis [24], and in mated worms sperm is in excess, thus the extended reproductive span we observed cannot be due to extended spermatogenesis. Our results, together with the Hughes, et al. data, suggest that significant reproductive aging already occurs prior to the cessation of sperm availability in self-fertilized animals, and that Sma/Mab mutants, like IIS (daf-2) and DR (eat-2) mutants [7], slow the rate of aging of the reproductive system.
We noticed that the Sma/Mab mutants produced fewer progeny than wild type each day in the early phase of reproduction (Figure S3; Figure S4D, S4E, S4F), and have smaller broods (Table S3, Table S4). This reduction in progeny number reflects slower ovulation rate of the mutants in early reproduction (Figure S4C), likely due to their small body size (Figure S4A, S4B). In fact, it has been suggested that reduction of C. elegans progeny number is linked to small body size via physical constraint of the maternal gonad and/or body size [32]–[35]. The downstream transcription co-factor of the Sma/Mab pathway, sma-9, is required in early larval development for the regulation of body size before gametogenesis [25],[36]; however, we find that reduction of sma-9 only in adulthood is sufficient to extend late reproduction (Figure S5), suggesting that the growth and reproductive aging functions of the Sma/Mab pathway are independent.
In mated assays, sperm number is not limiting; therefore, one possible explanation for extended reproductive span of the Sma/Mab mutants is that oocyte number is limiting, and thus slower ovulation allows the mutants to use up their oocyte supply more slowly. To test this hypothesis as well as the body size effect on reproduction, we examined five small but non-TGF-ß mutants (dpy-6(e2762), sma-1(ru18), sma-1(e30), dpy-1(e1) and dpy-9(e12)) whose body sizes are similar to the TGF-ß Sma/Mab mutants (compare Figure S4A and S4B and Figure S6A and S6B). Importantly, none of these strains have been reported to have egg-laying or embryonic developmental abnormalities or effects on longevity, and therefore serve as a fair set of samples for comparison. As expected, the five small mutants have slower ovulation rates (Figure 3A and 3B), and as a result produce fewer early progeny and fewer total progeny (Figure 3C and 3D). However, unlike the TGF-ß mutants, none of the small, non-TGF-ß mutants extended mated reproductive span (Figure 3E and 3F). In fact, all of the mutants had shorter reproductive spans. These data suggest that small body size and reductions in ovulation rate and progeny number do not increase reproductive span, but rather are usually associated with shorter reproductive spans. Therefore the TGF-ß Sma/Mab mutants are special in their extension of reproductive span.
We also addressed whether oocyte number becomes the limiting factor when sperm is no longer limiting by mating animals with young wild-type males, which is one basis for the assumption that slow ovulation extends reproductive span. On day 3, all wild-type animals (n = 12) produced only eggs that were able to hatch and develop to viable progeny (Figure 3G), but on day 7, 59% (n = 17) of the animals laid oocytes that failed to be fertilized (Figure 3H) and/or eggs that were unable to hatch (Figure 3H and 3I), resulting in cessation of viable progeny production. Therefore, the limiting factor is not number of oocytes, which are clearly still in excess, but the quality of the oocytes. By contrast, the majority of sma-2 mutants still produced exclusively viable eggs on day 7 (Figure 3J), with only 19% (n = 16, p = 0.03 compared with wild type) of the animals starting to lay unfertilized oocytes or unhatchable eggs. Our data, together with the observation that late reproduction is independent of early reproduction [7], suggest that Sma/Mab mutants extend reproductive span independently of body size, ovulation rate, early progeny number, and brood size, but instead by improving oocyte quality.
While daf-2 and eat-2 regulate reproductive aging [6],[7],[11], they are known foremost for their roles in life span extension [10],[12]. Thus far, all of the known mutations and treatments that extend C. elegans reproductive period also regulate longevity [6],[7]. In addition, the link between longevity and reproduction has been suggested and/or reported in multiple organisms [37]–[43]. These data suggest the possibility that the regulation of life span and reproductive span are coupled, or even regulated by the same mechanisms. Our TGF-ß Dauer pathway data further support this notion, as the mutants increase both life span and reproductive span (Figure 4E) [19].
However, we found that Sma/Mab pathway mutants only mildly affect longevity, despite their dramatic effects on reproductive span. Some Sma/Mab mutants increased life span moderately (dbl-1, sma-6) or inconsistently (sma-2, sma-4), while others appeared to have no effect on longevity (sma-3, sma-9) (Figure 4A–4D; Table S6, S7, S8, S9, S10). Because these alleles are not nulls, the inconsistencies in longevity between mutants in the pathway could be due to varying hypomorphic effects. Therefore we compared each mutant allele's effect on reproductive span and life span (Figure 4E). While daf-4 mutants doubled both reproductive span and life span, and daf-7 mutations had a moderate effect on each, exclusive members of the Sma/Mab pathway disproportionately extended reproductive span compared to life span (Figure 4E). This effect was maintained when the mated reproductive spans were considered (compare sma-2 and daf-7 in Figures 4E, Figure 5G and 5H). daf-4's effect on the two processes is likely due to its dual roles in TGF-ß Dauer regulation of life span [19] and in TGF-ß Sma/Mab regulation of reproductive span.
Matricide is a common event in the late reproductive period, but we noticed that the TGF-ß mutants are different from longevity mutants in this regard. In mated wild-type animals, the matricide frequency increased with age within the reproductive period (Figure 5A), perhaps reflecting aging of the musculature required for egg-laying. After day 9, reproduction stopped completely, therefore no matricide was observed. The matricide frequency of sma-2 mutants also increased with age at a similar rate as wild type (Figure 5A). Because sma-2 mutants continued to reproduce, however, the matricide rate continued to rise further, and therefore more mutant animals suffered from matricide than wild type. The matricide frequency of daf-2 and eat-2 mutants, however, increased at a much slower rate (Figure 5B). For example, on day 8 about 70% of worms died from matricide in both wild type and sma-2 animals, whereas only 40% of daf-2 and 30% of eat-2 animals died of matricide (Figure 5A and 5B). (daf-7 mutants exhibited high matricide rate from very early age (Figure 5B), due to their Egl defects (Figure 1D), therefore are different from the other strains.) The matricide frequency data suggest that late-reproduction matricide may be a somatically-controlled event, separate from reproductive aging. Together with the life span data, the matricide data suggest that sma-2's soma ages at a rate that is similar to wild type, unlike the daf-2 and eat-2 longevity mutants.
To further investigate sma-2's role in somatic aging, we compared the effects of mutations in sma-2, daf-2, eat-2, and daf-7 on life span and reproductive span. The other mutants have longer life spans than sma-2 (Figure 5C and 5D; Table S7 and Table S8), but their reproductive spans are either shorter or comparable to sma-2 (Figure 5E and 5F; Table S7, S8). In fact, daf-2 animals increase life span to a greater degree than reproductive span (Figure 5G and 5H), while sma-2 and eat-2 have greater effects on reproductive span than life span. sma-2's effect on reproductive span is the most disproportionate among all the mutants. For example, when comparing the increases in mated reproductive spans and life spans of all the mutants (Figure 5H), sma-2's increase in reproductive span is 10-fold its increase in lifespan, whereas eat-2's effect on reproductive span is only 4-fold, and the daf-7's and daf-2's are both less than one fold. Together, our data show that TGF-ß Sma/Mab signaling affects reproductive aging disproportionately to its effect on longevity compared to other reproductive span and life span mutants.
The FOXO transcription factor DAF-16 is required for longevity of the IIS pathway mutant daf-2 [10] and is also required for daf-2's effects on reproduction [7] (W. Shaw & C.T. Murphy, unpublished data). Previously, we showed that TGF-ß Dauer signaling regulates longevity through its interactions with the IIS pathway, as the lifespan extension of daf-7 mutants is suppressed by loss of DAF-16/FOXO transcription factor activity [19] (Figure 6A). To test the role of daf-16 in TGF-ß Dauer pathway regulation of reproduction, we compared the reproductive spans of daf-7(e1372), daf-16(mu86), and daf-16(mu86);daf-7(e1372) double mutants. We found that daf-7's reproductive span extension was significantly suppressed by loss of daf-16 activity (Figure 6B and 6C; Table S9), suggesting that DAF-16/FOXO activity is required for both the life span and reproductive span extensions of daf-7 mutants. Interestingly, loss of daf-16 activity also suppressed sma-2's small life span extension (Figure 6D, Figure S7A; Table S9); sma-2's occasional moderate effect on longevity may be due to cross-talk between the TGF-ß and IIS pathways rather than a primary output of TGF-ß Sma/Mab signaling, reminiscent of the cross-talk between the two TGF-ß pathways in dauer regulation [44],[45], and TGF-ß Dauer/IIS cross-talk in longevity regulation [19]. By contrast with its effect on sma-2 life span, loss of daf-16 activity failed to suppress sma-2 and dbl-1 self-fertilized and mated reproductive span extensions (Figure 6E and 6F; Figure S7C and S7D; Table S9). However, the double mutant's peak matricide frequency shifted earlier (Figure S7B), consistent with daf-16;sma-2 and daf-16's shorter life spans. These results suggest that the TGF-ß Sma/Mab pathway's regulation of reproductive span is not mediated by DAF-16/FOXO activity. Together with its disproportionate effect on the two processes, the TGF-ß Sma/Mab pathway appears to have genetically uncoupled regulation of reproduction and longevity.
The FoxA transcription factor PHA-4 is required for the life span extension of the Dietary Restriction model eat-2 (Figure 7A) [46]. We found that eat-2's reproductive span extension was also significantly suppressed by loss of PHA-4 activity (Figure 7C and 7D; Table S10) when treated from L4 onward. (To check the efficacy of pha-4 RNAi, we determined the fraction of arrested L1 progeny from L4-onward-fed mothers, and found that sma-2 and eat-2 animals are similarly sensitive to pha-4 RNAi (Figure 7B).) To test whether TGF-ß Sma/Mab signaling utilizes the Dietary Restriction pathway, we tested sma-2's requirement for PHA-4 activity. In contrast to eat-2's requirement for pha-4, we found that pha-4 RNAi treatment did not suppress the reproductive span extension of sma-2 mutants (Figure 7E and 7F; Table S10). Thus, while pha-4 is required for the reproductive span changes associated with Dietary Restriction, it is not required for sma-2's reproductive span extension.
If the reproductive span extension of Dietary Restriction animals required TGF-ß Sma/Mab signaling, we might expect the DBL-1 ligand to act downstream of the eat-2-induced DR effect, which is caused by eat-2 mutants' inability to pump their pharynx to properly ingest food. However, we observed that over-expression of the DBL-1 ligand is not sufficient to suppress eat-2's reproductive span extension (Figure 7G; Table S11), consistent with the hypothesis that Dietary Restriction and TGF-ß Sma/Mab regulate reproductive span independently. (Interestingly, the eat-2;dbl-1 OE animals, like dbl-1 OE single mutants, are larger than wild type (Figure 7H) yet have a long reproductive span, further supporting the notion that there is no direct correlation between body size and reproductive span.)
Our results show that in addition to its role in body size regulation, the TGF-ß Sma/Mab pathway is a novel regulator of reproductive aging. Sma/Mab signaling regulates reproductive aging independently of at least two known regulators of somatic aging, Dietary Restriction and IIS/FOXO signaling, and uncouples reproductive aging from somatic aging.
Here we have shown that loss of function of a canonical TGF-ß pathway significantly delays C. elegans reproductive aging. Intriguingly, TGF-ß Sma/Mab mutants extend reproductive span without a proportional extension of life span. We have also found that TGF-ß Sma/Mab reproductive span extension is genetically independent of regulation by Dietary Restriction and Insulin/IGF-1 Signaling. The uncoupling of reproductive and somatic aging in TGF-ß Sma/Mab signaling mutants suggests that the molecular mechanisms underlying the maintenance of somatic [47],[48] and reproductive tissues are distinct. This is the first identification of a pathway that regulates reproductive aging independently of somatic aging, and may lead to insights into mechanisms that specifically govern age-related reproductive cessation.
Signaling from the reproductive system to the soma to regulate aging has already been shown through germ line and somatic gonad ablation experiments in worms and flies [49],[50]. Because germ line and somatic gonad ablation both result in sterility, but have opposite effects on life span, direct resource allocation from the germ line to the soma cannot be the cause of this longevity. Instead, signals from the germ line and somatic gonad normally communicate with the rest of the soma to regulate life span, acting through the insulin/FOXO and daf-12 nuclear hormone pathways [49]–[51]. Our results suggest that the reproductive system may also normally receive signals through these regulatory pathways and the TGF-ß Sma/Mab pathway to regulate its rate of aging, allowing the animal to adjust its reproductive rate to its environment.
The prevalence of matricide by the TGF-ß Sma/Mab mutants illustrates the importance of signaling that normally links reproductive and somatic aging. Many animals slow their reproductive rates in response to environmental factors, such as high predation and food shortages, in order to optimize their reproductive fitness [7],[16],[18],[52]. In some cases, such as under environmental stress, late reproduction can increase fitness, allowing increased genetic diversity through mating (“facultative outcrossing”) [17]. Like daf-2 and eat-2 mutants, reduced TGF-ß Sma/Mab signaling slows the rate of reproductive aging and thus extends reproductive span. However, TGF-ß Sma/Mab mutants do not concomitantly slow somatic aging, and as a result they often suffer from high reproductive-age mortality induced by the physical stresses of reproduction, whereas same-aged longevity mutants daf-2 and eat-2 experience less age-related matricide (see Figures 2D and 2E and Figure 5A and 5B). From these results we infer that the slowing of somatic aging is required for successful delayed reproduction, which may be necessary under certain environmental conditions. Signals through the TGF-ß Sma/Mab pathway may allow the animal to adjust its reproductive rate to its environment. In turn, signals from the germ line and somatic gonad coordinate somatic aging rate with reproductive aging rate to allow successful reproduction.
If reproductive and somatic aging are coupled, why do worms and humans live so long after reproduction has ceased? A popular but controversial theory specific to humans postulates that investment in grand-progeny by grandmothers increases fitness to a greater degree than would continuing reproduction, and thus downregulation of reproductive ability in mid-life is evolutionarily beneficial (the “Grandmother Hypothesis” [53],[54]). However, as C. elegans does not care for its young, such an investment in its grand-progeny cannot explain its similarly early decline in reproduction and long post-reproductive life span.
Instead, we propose that reproduction itself requires the soma to function at its highest level, but the body can survive well below this threshold level of function. That is, if one were to plot every parameter of function (reproduction, motility, pathogen resistance, survival, etc.) in both worms and humans, these functions would all peak during the reproductive period, but begin to decline post-reproductively, each at a different rate. Successful reproduction requires peak physical condition of the soma; the increased matricide rates in older, reproductive Sma/Mab mutants shows that increased oocyte quality in the absence of healthy somatic tissues can be catastrophic, at least to the mother and any unproduced late progeny. By contrast, “survival” is the lowest measurable function (assayed as live vs. dead) and thus persists much longer than reproductive activity. Humans survive well past the age of peak physical function, a period that overlaps with female reproduction. Improvements in medicine, nutrition, hygiene, and environment have extended human life span significantly [55],[56], extending the post-reproductive life span but having little effect on maximum reproductive span. Analogously, the worm's life span in the low-predation and low-pathogen conditions of the laboratory is likely longer than in the wild, but likely largely affects post-reproductive life span. Therefore, the long post-reproductive life span of both worms and human females could be attributed to the high level of somatic function required for successful reproduction, essentially a side effect of the requirements for successful reproduction earlier in life.
Longevity is regulated by insulin/IGF-1/FOXO signaling and by Dietary Restriction in worms through mammals [10], [12], [57]–[59], despite large differences in chronological life spans of these organisms. Additionally, Insulin/IGF-1 Signaling and Dietary Restriction have been implicated in regulation of mammalian reproductive aging [14],[52],[60],[61]. Intriguingly, TGF-ß levels are upregulated in aged mouse oocytes [62], and TGF-ß activity regulates mammalian follicle cell activity [63]. Thus, it is possible that regulation of reproductive aging, like the regulation of somatic aging by IIS and DR pathways, is evolutionarily conserved, and that TGF-ß signaling may regulate human reproductive cessation.
Despite the vast differences in their life histories and chronological time frames, our work suggests that the regulation of worms and humans' longevity and reproductive spans may be conserved. Future studies will determine whether Sma/Mab mutants use a conserved mechanism to slow reproductive aging in C. elegans. If so, modulation of TGF-ß signaling may offer new avenues to improve fertility and offspring health in mothers of advanced age.
All strains were cultured using standard methods [64]. In all experiments, N2 is the wild type. LG I: daf-16(mu86), daf-8(e1393). LG II: eat-2(ad465), rrf-3(pk1426), sma-6(wk7). LG III: daf-2(e1370), daf-7(e1372), daf-4(e1364), sma-2(e502), sma-3(wk28), sma-3(wk20), sma-4(e729). LG IV: daf-1(m40), daf-14(m77), fem-1(hc17). LG V: dbl-1(nk3), dbl-1(wk70). LG X: sma-9(qc3), sma-9(wk55).
Strains: BW1940: ctIs40 X [ZC421(+) containing dbl-1;sur-5::gfp]. CQ33: eat-2(ad465) II; ctIs40 X [ZC421(+) containing dbl-1;sur-5::gfp]. CQ17: daf-1(m40) IV outcrossed to N2 3×. CQ16: daf-7(e1372) III outcrossed to N2 3×. CQ14: daf-14(m77) IV outcrossed to N2 3×. CQ19: sma-2(e502) III outcrossed to N2 3×. CQ18: sma-9(wk55) X outcrossed to N2 3×. CQ53: sma-2(e502);fem-1(hc17).
CQ49: daf-16(mu86);sma-2(e502). CQ25: daf-16(mu86);daf-7(e1372). CF1041: daf-2(e1370) III.
Individual synchronized L4 hermaphrodites were moved to fresh plates daily until reproduction ceased for at least two days. The last day of viable progeny production was noted as the day of reproduction cessation for each individual. When matricide occurred, the animal was censored from the experiment on that day. All experiments were performed at 20°C, except that sma-2(e502);fem-1(hc17) and fem-1(hc17) worms were shifted to 25°C from L3 and back to 20°C after L4. All experiments were performed with at least 10 individuals per strain (most experiments included >25 individuals, as indicated in Supplementary Tables). The log-rank (Mantel-Cox) method was used to test the null hypothesis. In mating reproductive span assays, L4 hermaphrodites were mated to young wild-type males at a 1∶3 ratio for 24 hours before being separated onto individual plates. Successful mating was ascertained by the fraction of male progeny each day. For the pha-4 RNAi reproductive span experiments, mothers were moved onto RNAi bacteria starting at L4.
Individual synchronized L4 hermaphrodites were moved to fresh plates and the number of progeny produced by each individual was counted daily until reproduction ceased for at least two days. When matricide occurred, the animal was censored from the experiment on that day. All experiments were performed at 20°C with at least 6 individuals per strain (most experiments included 20–40 individuals).
The assay was performed as described in Reproductive Span analysis, except the cumulative percentage of hermaphrodites that underwent matricide was calculated daily. The matricide frequency was determined as the frequency of reproductive worms that die of matricide; as matricide is caused by internal progeny hatching, non-reproductive worms by definition never die of matricide, and thus are not included in calculation. This number reflects the likelihood of matricide.
Eggs were synchronized by hypochlorite treatment and allowed to develop at 20°C until day 3 of adulthood. ∼100 synchronized hermaphrodites were transferred to a new plate and allowed to lay progeny for 1 hour; eggs and L1 progeny were counted at 3-hour intervals.
The first day of adulthood was defined as t = 0, and the log-rank (Mantel-Cox) method was used to test the null hypothesis in Kaplan-Meier survival analysis, as previously described [65]. All experiments were carried out at 20°C with 50 µM FUdR starting at L4; n>60 in each experiment.
Bacterial feeding RNAi experiments were carried out as previously described [66] with IPTG at 1 mM. Each clone was verified by PCR and sequence analysis.
pha-4 RNAi efficacy was determined by counting the arrested L1s produced by mothers fed from L4 onward compared with control vector.
|
10.1371/journal.pgen.1006162 | Fragment Length of Circulating Tumor DNA | Malignant tumors shed DNA into the circulation. The transient half-life of circulating tumor DNA (ctDNA) may afford the opportunity to diagnose, monitor recurrence, and evaluate response to therapy solely through a non-invasive blood draw. However, detecting ctDNA against the normally occurring background of cell-free DNA derived from healthy cells has proven challenging, particularly in non-metastatic solid tumors. In this study, distinct differences in fragment length size between ctDNAs and normal cell-free DNA are defined. Human ctDNA in rat plasma derived from human glioblastoma multiforme stem-like cells in the rat brain and human hepatocellular carcinoma in the rat flank were found to have a shorter principal fragment length than the background rat cell-free DNA (134–144 bp vs. 167 bp, respectively). Subsequently, a similar shift in the fragment length of ctDNA in humans with melanoma and lung cancer was identified compared to healthy controls. Comparison of fragment lengths from cell-free DNA between a melanoma patient and healthy controls found that the BRAF V600E mutant allele occurred more commonly at a shorter fragment length than the fragment length of the wild-type allele (132–145 bp vs. 165 bp, respectively). Moreover, size-selecting for shorter cell-free DNA fragment lengths substantially increased the EGFR T790M mutant allele frequency in human lung cancer. These findings provide compelling evidence that experimental or bioinformatic isolation of a specific subset of fragment lengths from cell-free DNA may improve detection of ctDNA.
| During cell death, DNA that is not contained within a membrane (i.e., cell-free DNA) enters the circulation. Detecting cell-free DNA originating from solid tumors (i.e., circulating tumor DNA, ctDNA), particularly solid tumors that have not metastasized, has proven challenging due to the relatively abundant background of normally occurring cell-free DNA derived from healthy cells. Our study defines the subtle but distinct differences in fragment length between normal cell-free DNA and ctDNA from a variety of solid tumors. Specifically, ctDNA was overall consistently shorter than the fragment length of normal cell-free DNA. Subsequently, we showed that a size-selection for shorter cell-free DNA fragments increased the proportion of ctDNA within a sample. These results provide compelling evidence that development of techniques to isolate a subset of cell-free DNA consistent with the ctDNA fragment lengths described in our study may substantially improve detection of non-metastatic solid tumors. As such, our findings may have a direct impact on the clinical utility of ctDNA for the non-invasive detection and diagnosis of solid tumors (i.e., the “liquid biopsy”), monitoring tumor recurrence, and evaluating tumor response to therapy.
| Increased quantity of cell-free DNA in the circulation has been associated with malignant solid tumors [1]. Longitudinal studies have reported reductions in cell-free DNA quantity in response to therapy and elevations associated with recurrence suggesting quantification of cell-free DNA may be useful for monitoring disease status [2–4]. However, quantifying cell-free DNA as a marker of disease and its extent has been limited. The quantity of cell-free DNA has not correlated well with stage and histological subtype [5, 6]. In addition, large inter-subject variations of cell-free DNA quantification have been described leading to overlap between malignant disease, benign tumors, and healthy controls [7, 8]. Moreover, increased quantity of cell-free DNA is non-specific to cancer and has been associated with other conditions such as autoimmune disease and environmental exposures [9, 10]. Finally, except in patients with advanced metastatic disease, tumor-derived cell-free DNA (i.e., circulating tumor DNA, ctDNA) forms only a small minority of the cell-free DNA in circulation against a background of fragments mostly derived from normal cells. Therefore, the quantification of cell-free DNA alone is of little prognostic value.
As an alternative, detecting specific variants or mutational hotspots in ctDNA may have important clinical implications in the shift towards personalized medicine for diagnosing and/or monitoring malignancies. In lung cancer, EGFR mutations in ctDNA have been associated with prognosis and utilized for determining therapy (e.g., activating mutations that confer sensitivity to tyrosine kinase inhibitors) [11]. However, molecular ctDNA studies in a variety tumor types have largely focused on advanced or metastatic disease in which ctDNA is more readily detectable compared to localized disease [12]. Bettegowda et al. reported a substantial reduction in detectability of ctDNA in localized disease compared to metastatic tumors for breast, colon, pancreas, and gastroesophageal cancers [13]. Moreover, ctDNA from glioblastoma multiforme (GBM), a primary brain tumor associated with neovascularization and disruption of the blood-brain barrier, was undetectable [13]. This latter finding supports the general perception that detection of ctDNA from non-metastatic solid tumors is particularly challenging since GBM does not metastasize beyond the central nervous system.
Emerging approaches to improve detection of ctDNA include amplicon-based strategies in colorectal cancer [14] and integrated digital error suppression during deep sequencing in lung cancer [15]. While the latter methods seek to eliminate artifacts during sequencing to improve bioinformatic analytic sensitivity of mutant allele detection, the former techniques exploit apparent size differences between ctDNA and cell-free DNA. Specifically, previous amplicon-based studies have shown that ctDNA is highly fragmented and occurs most commonly at a size <100 bp, while normal cell-free DNA is proportionally more represented at a size >400 bp [16]. In this study, we initially sought to determine the feasibility of detecting ctDNA associated with GBM by utilizing a xenograft tumor model to exploit genomic species differences to separate ctDNA from the background host animal benign cell-free DNA. In so doing, we identified precise differences in fragment lengths between ctDNA and normal cell-free DNA, which were more narrow and more consistent than previously described [16, 17]. In addition, we found strong evidence of a 10 bp periodicity in ctDNA that was less prominent in normal cell-free DNA. These observations led us to explore if similar findings were present in tumors outside the brain and subsequently translated to cell-free DNA samples obtained from cancer patients. Collectively, the results described herein demonstrate that the fractional selection of cell-free DNA with a specific size range that is 20–50 bp shorter than the size of normal healthy cell-free DNA may substantially enrich for ctDNA in human cancer.
Established human GBM stem-like cell lines (GBM4 and GBM8) [18, 19] were implanted in the nude rat brain. Control animals underwent an identical surgical procedure and were inoculated with medium only. Quantitative magnetic resonance imaging techniques were implemented on a 3T whole-body clinical scanner (Philips Achieva) to phenotype the tumors. Fast bound-pool fraction imaging (FBFI), a method validated with histology to measure myelin density and identify tumor associated disruption of normal brain tissues [20], was used to produce bound-pool fraction maps (f maps) to detect and differentiate between bulky and infiltrative lesions. The variable flip angle method [21] was used to measure T1 relaxivity (R1 maps, where R1 = 1/T1) before and after administration of gadolinium (gadopentetate dimeglumine, Bayer HealthCare), an intravenous contrast agent that shortens T1 and identifies disruption of the blood-brain barrier (i.e., hyperintense signal on post-contrast R1 maps relative to pre-contrast R1 maps). Our initial experiments found that GBM4 yielded small, focal, non-enhancing lesions (Fig 1A and 1B, S1A and S1B Fig). In contrast, GBM8 produced heterogenous lesions that ranged from large, well-circumscribed tumors with strong contrast enhancement (i.e. disruption of the blood-brain barrier) to infiltrative lesions with absent or minimal contrast enhancement (Fig 1A and 1B, S1A and S1B Fig).
Detection of human ctDNA associated with GBM4 was not greater than the control animals (Fig 1C), which was attributable to the small tumor size and absence of blood-brain barrier disruption. Human ctDNA was detected in all animals implanted with GBM8 including infiltrative lesions with absent (e.g., GBM83, Fig 1A and 1B) to minimal (e.g., GBM82, S1A and S1B Fig) disruption of the blood-brain barrier (Fig 1C, S1C Fig). The percent human ctDNA in the buffy coat, where some residual plasma remains present, appeared correlated with the fraction in plasma, but much lower, indicating that neither intact tumor cells nor a high molecular weight fraction of ctDNA were present in the circulation (Fig 1C). Unexpectedly, there was a precise difference in fragment length between human ctDNA and rat cell-free DNA. The most common fragment lengths in human ctDNA were 134 bp and 144 bp (Fig 1D, S1D Fig), which was in contrast to the most common fragment length of 167 bp in rat cell-free DNA (Fig 1F). Human ctDNA fragment lengths also exhibited a strong ~10 bp periodicity that was not as evident in the rat cell-free DNA. This pattern was consistent across all animals where human ctDNA was detected (Fig 1E).
To determine if the fragment length and periodicity extended beyond the GBM8 cell line, nude rats were again implanted with either GBM4 or GBM8. Animals implanted with GBM4 were serially imaged until the presence of blood-brain barrier disruption was evident (i.e., contrast-enhancement on MRI) or animals lost more than 10% bodyweight. As before, GBM8 animals developed tumors that were large and exhibited a range of phenotypes (Fig 2A and S2 Fig). Fragment length and periodicity was consistent with that present in the initial experiment (Fig 2C). After a post-surgical interval nearly twice as long as GBM8, GBM4 tumors developed (Fig 2 and S2 Fig). GBM4 tumors tended to grow more anteriorly towards the olfactory bulbs, which led to weight loss in animals before tumor size was similar to GBM8. In a single animal implanted with GBM4, ctDNA was adequately detected and a similar fragment length and periodicity as seen with GBM8 was identified (Fig 2C).
To evaluate the role of both GBM and the blood-brain barrier in determining fragment length and periodicity, human hepatocellular carcinoma cells (Hep G2) were implanted subcutaneously in the flank of three nude rats. A palpable tumor (approximately 10 mm at maximal diameter), confirmed with histology, formed in a single animal (Fig 2D). The fragment length of ctDNA was consistent with that described in GBM4 and GBM8 (Fig 2E). There was evidence for a similar periodicity, but the relatively low amount of detected ctDNA may have contributed to a noisier distribution of fragment size. Regardless, the replication of results in a xenograft model of hepatocellular carcinoma suggested that the periodicity and reduced fragment length of ctDNA may be general properties of ctDNAs in cancers beyond GBM.
We next considered the effects of the xenograft model on ctDNA fragment length and periodicity and sought to determine if evidence of the observed differences in fragment length were present in other types of solid tumors such as melanoma. In contrast to the xenograft models, the cell-free DNA from tumor patients represented an indistinguishable mix of both ctDNA and cell-free DNA derived from normal healthy cells. By densitometry (TapeStation 2200), the cell-free DNA from melanoma patients had globally shorter fragment lengths compared to healthy controls (Fig 3A).
Potential differences in fragment length size between tumor patients and healthy controls were explored by sequencing the cell-free DNA from a melanoma patient with an elevated concentration of cell-free DNA (36.4 ng/mL plasma; Fig 3A, black arrow) to obtain a large sample for comparison to sequencing results from a pooled sample of control cell-free DNA. The most common fragment length in the melanoma patient was shorter than the most common fragment length in the control sample (145 bp vs. 165 bp, respectively; Fig 3B). There was also evidence for more pronounced fragment length periodicity in the cell-free DNA from the melanoma patient (Fig 3B). In the melanoma patient cell-free DNA, the BRAF V600E allele frequency was increased at shorter fragment lengths compared to the WT allele (Fig 3C). Of note, the broad distribution of the WT allele (Fig 3C, blue line) included a substantial proportion of overlapping fragment sizes with the mutant allele since the BRAF V600E mutation is heterozygous and tumor cells also introduced shorter fragment lengths into the circulation with the WT allele. Subsequently, fragment length for the melanoma patient and the healthy volunteer were binned (e.g. 50 = 50–59 bp; 60 = 60–69 bp, etc.) and the frequency of mutant allele and WT allele was determined, respectively. For a given fragment length, the proportion of the V600E BRAF allele to the WT allele was highest in the 110–140 bp fragment length, which was in contrast to the WT allele in the pooled healthy control sample that occurred at the highest frequency between 160–180 bp (Fig 3D). Importantly, there were limited observations of ctDNA fragments <100 bp in the melanoma patient (Fig 3D, black line), which were likely present but not well recovered by current approaches to library preparation [22]. These collective findings indicated an overall shortening of ctDNA fragment size relative to cell-free DNA that was not an effect of the xenograft model, but rather inherent to ctDNAs across different tumor types.
We then sought to characterize tumor-related differences in cell-free DNA and ctDNA associated with human lung cancer. A comparison of cell-free DNA from 15 lung cancer patients and 9 healthy controls found a statistically significant difference in plasma concentration of cell-free DNA (31.0±23.3 vs. 11.3±4.6 ng mL/plasma, p = 0.006; respectively); however, the range of concentration in tumor patients was broad and overlapped with the concentrations present in healthy controls (S3 Fig). Libraries made with duplex truncated molecular barcoded adapters, which added ~99 bp to each strand of cell-free DNA, were created to enable loading of a consistent concentration (2 ng/μL) of each sample and clear identification of upper/lower markers for direct comparison of densitometry data (TapeStation 2200; S4 Fig). Consistent with our observation in melanoma patients, peak fragment length by densitometry was significantly shorter compared to healthy controls (277.0±4.7 vs. 283.7±4.1 bp, p = 0.002; respectively; S3B Fig) indicating a global shift towards smaller fragments in lung cancer patients. However, there was also considerable overlap of peak fragment length between tumor patients and controls (S3B Fig). There was not an association between peak fragment length by densitometry and plasma concentration of cell-free DNA in the lung cancer patients (Pearson’s r = –0.20, p = 0.47; Fig 4C) or the healthy controls (Pearson’s r = 0.19, p = 0.63; S3C Fig).
For a subset of samples (tumor, N = 7; control N = 5), cell-free DNA was converted to Illumina sequencing libraries and enriched for cancer-relevant genes using a 16-gene capture panel. Utilizing fragment lengths from the complete 16 gene capture panel, we observed that the cell-free DNA from tumor patients ranged from a shorter fragment length with (Fig 4A) and without (Fig 4B and S5A–S5D Fig) a strong periodicity to indistinguishable (Fig 4C and S5E–S5G Fig) in fragment length distribution from healthy controls. Cell-free DNA from the tumor patients did not exhibit a fragment length larger than the controls (S5 Fig), which was consistent with densitometry (S3B Fig).
Cell-free DNA from two lung cancer patients (LC5 and LC10) contained the classic EGFR L858R mutation [23]. Fragments containing the mutant allele, which originate from the tumor rather than from breakdown of normal cells, were shorter than those bearing the WT allele in healthy controls (Fig 4D). This difference was especially pronounced in one sample (LC5; Fig 4E) with a relatively high mutant allele frequency (74.6%; likely due to EGFR amplification, S6 Fig).
Cell-free DNA from six of the lung cancer patients contained the EGFR T790M mutation. In 5 out of 6 patients, the mutant allele frequency was relatively low (0.2–6.6%). However, the general trend in these samples was for mutant alleles to occur at shorter fragment lengths (Fig 4F). In one sample (LC9) with a relatively high mutant allele frequency (25.1%; most likely due to an EGFR amplification, S5 Fig) the mutant allele more commonly occurred at shorter fragment lengths compared to the fragment lengths from healthy controls (Fig 4G). The distribution of fragment lengths of the EGFR WT allele between tumor patients and healthy controls largely reflected differences seen in S5 Fig, although noisier due to fewer total reads (S7 Fig). Within each lung cancer patient with the mutant T790M allele, comparison of the distribution of the EGFR WT allele and the mutant T790M allele fragment lengths identified a general trend for the mutant allele to occur more commonly at shorter fragment lengths (Fig 4H). As with the melanoma patient (Fig 4C), fragment length analysis of the WT allele from tumor patients included an indistinguishable mixture of ctDNA and normal cell-free DNA since the mutant T790M allele is heterozygous. As such, the representative WT allele fragment length distribution from tumor patients included WT alleles derived from tumor cells. This observation may explain, at least in part, why the differences in fragment length between the WT allele and the mutant T790M allele presented in Fig 4H were less pronounced than differences observed between the WT allele from healthy controls and the mutant T790M allele shown in Fig 4F.
We next set out to determine whether selection for shorter fragment lengths could be used to enrich for ctDNA fragments against the large background of cell-free DNA derived from normal cells. Cell-free DNA sequencing libraries from four lung cancer patients (LC1, LC3, LC4, and LC10) with EGFR T790M mutations and one healthy control (C5) were selected for serial fraction collection. By sequencing, LC1 and LC3 had EGFR T790M mutant allele frequencies of 1.2% and 6.6%, respectively, and evidence of overall shorter cell-free DNA fragments compared to healthy controls (Fig 4B and S5B Fig, respectively). LC4 and LC10 had EGFR T790M mutant allele frequencies of 2.3% and 1.7%, respectively, and a similar size distribution of cell-free DNA fragments compared to healthy controls (S5G Fig and Fig 4C, respectively). None of the samples had an EGFR amplification present (S6 Fig). For each sample, 1 μg of sequencing library was loaded onto an 8% native polyacrylamide gel and six adjacent gel fragments were collected (S8 Fig). Extracted DNA (5–10 ng) from each gel fragment was then amplified using the full-length adapter primer and the EGFR T790M mutant allele frequency was determined via digital droplet PCR (Fig 5).
Compared to the mutant allele frequency in the library, three samples (LC1, LC4, and LC10) demonstrated a 2.5-fold to 9.1-fold increase in the mutant allele frequency in a subset of fractions that contained a shorter distribution of cell-free DNA fragments relative to the peak fragment length in the library (Fig 6 and S9–S11 Figs). The fraction associated with the greatest increase in mutant allele frequency for each tumor patient is identified in Fig 6A–6C. In one sample (LC1), the mutant allele frequency did not increase in any fraction relative to the mutant allele frequency in the library (S12 Fig). However, a decrease in the mutant allele frequency was observed in fractions containing longer fragments, while fractions with shorter fragments contained a relatively consistent mutant allele frequency (Fig 6D and S12 Fig). Enrichment for the mutant allele was greatest in fractions that were centered approximately 20–50 bp shorter than the peak fragment length associated with each corresponding library (Fig 6E). The increase in mutant allele frequency was greatest in LC10 (Fig 5 and S11 Fig) and LC4 (S10 Fig), which were the two tumor patients with a similar fragment size distribution profile as that seen in the healthy controls (Fig 4C and S5G Fig, respectively). This finding suggests that the fractional selection of shorter cell-free DNA fragment lengths may improve mutant allele sensitivity when ctDNA is not the predominant component of cell-free DNA. Also notable is that the percentage of mutant allele detected in a sample low in ctDNA prior to enrichment may not represent the true allele frequency present in the tumor due to dilution by normal cell-free DNA. In LC1 (S12 Fig) and LC3 (S9 Fig), the tumor patients with evidence of overall shorter cell-free DNA fragments compared to healthy controls, the increase in mutant allele frequency in fractions 20–50 bp shorter than the peak fragment length associated with each library was not as substantial; however, selecting these fractions also did not diminish the mutant allele frequency (Fig 6E). In contrast, the selection of fractions longer than the library’s peak fragment length substantially reduced the mutant allele frequency in three of the tumor samples (LC1, LC4, and LC10; Fig 6E and S10–S12 Figs). Similarly, the selection of fractions containing cell-free DNA fragments >50 bp shorter than the library’s peak fragment length reduced the mutant allele frequency in all of the samples (Fig 6E). This latter observation may be a consequence of recovery during library preparation as discussed earlier (Fig 3D) [22]. Regardless, these observations provide compelling evidence that the fragment length of ctDNA is shorter than cell-free DNA from healthy cells and selection of shorter cell-free DNA fragments may improve mutant allele frequency. Of note, the EGFR T790M mutant allele was not present above the noise level associated with digital droplet PCR in the fractions obtained from the control sample (S13 Fig).
Our broad observation that the fragment length of ctDNA differs from cell-free DNA is supported by earlier reports that utilized amplicons of varying length to identify large categorical size differences between ctDNA associated with colorectal cancer and cell-free DNA from healthy controls [16, 24]. In addition, deep sequencing has been previously used to identify ctDNA shortening in hepatocellular carcinomas with specific aneuplodies [17]. However, this latter study also identified fragment lengths larger than healthy controls associated with low ctDNA concentrations in patients with hepatocellular carcinoma which is difficult to reconcile [17]. The collective findings described in our study builds upon these previous works by utilizing massively parallel sequencing to define distinct differences in fragment length between ctDNA and cell-free DNA. Specifically, animal models of GBM and hepatocellular carcinoma found that the most common fragment lengths of ctDNA were 134 and 144 bp, which was in contrast to the most common 167 bp fragment length present in normal cell-free DNA. These findings replicated in human patients with melanoma. Moreover, selection of cell-free DNA fractions containing shorter fragment lengths substantially increased mutant allele frequency in human lung cancer patients, particularly when the distribution of cell-free DNA fragment lengths in tumor patients was similar to the distribution seen in healthy controls. As such, the findings described herein provide strong evidence that a more general process that shortens ctDNA fragment length relative to normal cell-free DNA from healthy cells is present and is independent of copy number alterations.
The overall distribution of fragment lengths identified for ctDNA and cell-free DNA in our study was consistent with cellular apoptosis rather than necrosis [25]. In addition, the observed ~10 bp periodicity has been well-described in association with nuclease-cleaved nucleosome activity [26]. However, the etiology of the shorter fragment length associated with ctDNA remains unclear. Lo et al. previously reported similar findings from maternal serum with regards to fragment length differences between fetal cell-free DNA and maternal cell-free DNA [27]. Differences in cell-free DNA fragment lengths between donor-derived and host cell-free DNA in organ transplant patients has also been observed [28]. The extent of cell-free DNA shortening across disparate tissue contexts, in health and disease, suggests that tissue-specific processes may contribute to certain cell-free DNA fragment length sub-populations. One plausible hypothesis is that tissue-specific differences in nucleosome wrapping [29] result in fragment lengths that differ between hematopoietic cells (which contribute the majority of the plasma cell-free DNA) and other tissues of origin. Understanding the specific mechanism behind this phenomenon may prove valuable in oncology. Regardless of etiology, enriching for a specific subset of cell-free DNA fragment lengths may improve detection of ctDNA associated with non-metastatic solid tumors. More sensitive detection of mutations present in ctDNA may lead to non-invasive diagnosis of malignancy, improved detection of tumor recurrence, and better monitoring of response to therapy.
A limitation of this study was that very short rat cell-free DNA fragments (<100 bp) were detected in the GBM41 animal (Fig 1F, red line) and very short human ctDNA fragments (<100 bp) were detected in the Control1 animal (S1D Fig, blue line) that were not present in the other animals. In the former, these very small fragments created a unique bimodal distribution of normal rat cell-free DNA. In the latter, these fragments were associated with an increased proportion of human ctDNA compared to other control animals (Fig 1C). As such, it was unclear if low levels (<0.01%) of ctDNA in tumor-bearing animals were a true signal or noise. Earlier use of a xenograft model for detection of ctDNA via PCR found a very high species sensitivity and specificity [24]. Although the very short fragments identified in our study were most likely secondary to contamination or sample handling, future xenograft-based studies utilizing species specific genomes obtained from massively parallel sequencing to separate ctDNA from cell-free DNA would benefit from determination of sensitivity and specificity. A second limitation is the accuracy of densitometry measurements (S14 Fig). Although densitometry tended to preserve relative differences between samples, we found that estimation of true fragment length was often over-estimated. As such, sequencing results may provide a more accurate measure of fragment length assuming sufficient reads of different sized inserts are available to reduce size profile noise.
All human subject research was approved by the University of Utah Institutional Review Board prior to study initiation. Written informed consent was obtained for samples from melanoma patients and healthy controls according to IRB approved studies 10924 and 7740. Informed consent was not obtained for the lung cancer samples as specimens were obtained from residual clinical samples scheduled for disposal and after de-identification according to IRB approved study 7275. Adult male RNU rats were used in this study. All procedures were approved by the University of Washington Internal Animal Care and Use Committee prior to study initiation. For surgery, rats were anesthetized with ketamine and xylazine administered IP. For imaging, rats were anesthetized with isoflurane mixed with oxygen. Rats were euthanized with Beuthanasia-D administered IP.
Established human GBM stem-like cell lines (GBM4, GBM8) [18, 19] were maintained in serum-free Neurobasal medium (ThermoFisher Scientific) with 2 mM glutamine, 5 μg/mL heparin, 100 U/ml penicillin-streptomycin, N2 at 0.5X, B27 supplement minus vitamin A at 0.5X, and bi-weekly pulsing of FGF and EGF (20 ng/mL each). Human hepatocellular carcinoma cells (Hep G2; ATCC) were maintained in Williams’ medium E with glutamine and 10% fetal bovine serum. All cells were maintained in a humidified incubator at 37°C in 5% CO2.
For implantation, single cell suspensions of GBM4 and GBM8 were achieved using heparin-EDTA and trituration followed by spin and wash ×2 with Neurobasal medium, then spin and resuspend in Neurobasal medium with DNase (4k U/ml), trituration, and incubation ×5 minutes at room temperature. Cells were then washed ×2 with Neurobasal medium to remove DNase and resuspended in Neurobasal medium for cell counting. Cells were counted with a hemacytometer after suspension in Trypan blue. For implantation, 1×106 cells were resuspended in 10 μL of Neurobasal medium. Hep G2 cells were harvested by heparin-EDTA, counted using a hemocytometer after suspension in Trypan blue. For implantation, 5×106 cells were resuspended in 100 μL of Williams’ medium E.
Adult male RNU rats (Charles River Laboratories, Wilmington, MA) were used in this study. All procedures were approved by the University of Washington Internal Animal Care and Use Committee prior to study initiation.
Rats were anesthetized with 60 mg/kg ketamine and 5 mg/kg xylazine administered IP. For intracranial inoculation (GBM4 and GBM8), the head was immobilized in a stereotactic head set with ear bars and a teeth bar. The skull was exposed by a 2 cm midline incision, and a burr hole was created on the right side 1 mm anterior and 2 mm lateral to the bregma. A microsyringe (Hamilton, Reno, NV) was used to inject the 10 μL aliquot of 106 cells into the frontal lobe at a depth of 5 mm from the skull surface over a period of 5 minutes. The needle was kept in place 2 minutes after injection to prevent backflow prior to removal. The burr hole was filled with bone wax (Ethicon, Somerville, NJ). The skin was closed with surgical staples that were removed prior to MR imaging. For flank injections (Hep G2), a 22-gauge needle attached to a TB syringe was used to inject the cells subcutaneously into the right flank.
After the final imaging time point, the rats were anesthetized with Beuthanasia-D (2 mL/kg). A midline abdominal incision followed by thoracotomy was made to access the left ventricle of the heart. A 22-gauge needle attached to a syringe containing heparin was used to remove as much blood as possible (6–10 mL). Subsequently, 4% paraformaldehyde (PFA) was injected into the left ventricle (total volume 150 mL) as the right atrium was opened. Brains were subsequently removed intact, held in 4% PFA×24 hours under gentle agitation, and then maintained in PBS.
After fixation, brains were sectioned to correspond with the anatomic coronal plane. Brains were subsequently embedded in paraffin and sections (5 μm thick) were stained with hematoxylin-eosin. Stained slides were scanned using an Olympus VS110 virtual microscopy system (Olympus, Center Valley, PA) for display on NDP.view (v2.3.1).
Rats were imaged on a 3.0 T Philips Achieva whole-body MRI scanner (Philips Medical Systems, Best, Netherlands) using a dual coil approach. A quadrature transmit/receive head coil (Philips Medical Systems) was utilized for RF transmission, and an in-house-built combined solenoid-surface coil [30] dedicated to high spatial resolution whole-brain rat imaging was used for RF reception. After induction in an anesthesia chamber with 5% isoflurane mixed with oxygen, the rats were positioned within the dual coils and maintained on 2% isoflurane mixed with oxygen via nose cone inhalation. Total scan time for all images was < 1 hour.
Images necessary for construction of bound-pool fraction maps in the rat brain were acquired as previously described [20]. Briefly, Z-spectra data points were acquired for each rat using a 3D MT-prepared spoiled gradient echo (GRE) sequence with TR/TE = 42/4.6 ms, excitation flip angle α = 10°, NEX = 1, and three offset frequencies (Δ = 4, 8, and 96 kHz) of the off-resonance saturation pulse (effective flip angle = 950°). Complementary R1 maps necessary for parameter fitting were obtained using the variable flip angle method [21] with a 3D spoiled GRE sequence (TR/TE = 20/2.3 ms, α = 4 (NEX = 3), 10 (NEX = 1), 20 (NEX = 2), and 30° (NEX = 3)). All Z-spectral and VFA images were acquired with a FOV = 29×29×19.8 mm3, matrix = 97×97×66, acquisition resolution = 0.3×0.3×0.3 mm3 (zero-interpolated to 0.15×0.15×0.15 mm3), and full-Fourier acquisition. Whole-brain 3D B0 and B1 maps were acquired to correct for field heterogeneities. For B0 mapping, the GRE-based dual-TE phase-difference method [31] was used with TR/TE1/TE2 = 20/4.7/5.7 ms and α = 10°. B1 maps were obtained using the actual flip angle imaging method [32] and the following sequence parameters: TR1/TR2/TE = 25/125/6.6 ms and α = 60°. 3D B0 and B1 maps were acquired with a FOV = 29×29×19.8, matrix = 64×64×33, acquisition resolution = 0.45×0.45×0.6 mm3 (zero-interpolated to 0.15×0.15×0.15 mm3), and NEX = 1. All images were acquired in the coronal plane.
For administration of gadolinium (gadopentetate dimeglumine, Bayer HealthCare; 0.5 M/L) a 22 Gauge angiocatheter (Becton-Dickinson, Sandy, Utah) was inserted into the rat tail vein. The catheter was attached to a small bore bifurcated extension (Smiths Medical, Dublin, OH) containing a dilution of gadolinium in one arm and a normal saline flush in the other. The catheter setup was maintained with a saline lock until immediately prior to imaging. At time of injection, 0.2 mmol/kg (0.167 M/L) of gadolinium was manually injected at 50 μL/s followed by a 250 μL flush of normal saline at 50 μL/s. Complementary pre-contrast R1 maps necessary for parameter fitting and post-contrast R1 maps obtained 5 minutes after contrast injection were acquired using the variable flip angle method [21] with a 3D SPGR sequence (TR/TE = 4.6/20 ms, α = 4 (NEX = 3), 10 (NEX = 1), 20 (NEX = 2), and 30° (NEX = 3)) and FOV = 24×24×8.25 mm3, matrix = 64×64×5.5 for an acquisition resolution of 0.38×0.38×1.5 mm3 (zero-interpolated to 0.19×0.19×0.75 mm3). Pre-contrast B1 maps using the actual flip angle imaging method [32] were acquired with the following parameters: TR1/TR2/TE = 25/125/6.7 ms and α = 60°. 3D B1 maps were acquired with a NEX = 1 and an identical resolution as the variable flip angle data points. All images were acquired in the axial plane.
Fast bound-pool fraction parametric maps (f maps) were constructed consistent with a previously described methodology [20] for single parameter determination of f. Briefly, R1 maps were used to define R1F and reconstructed from VFA data using a linear fit to the signal intensities (S) transformed into the coordinates [S(α) / sin α, S(α) / tan α][21] after voxel-based B1 corrections were applied to α. In the MT data, the Δ = 96 kHz Z-spectra images were used to normalize the Δ = 4 and 8 kHz data points and voxel-based B0 and B1 corrections were applied to Δ and α, respectively, during voxel-based fitting for f. The parameters k, T2FR1F, and T2B were constrained to 29 x f/(1-f) s-1, 0.030, 10.7 μs, respectively, as previously determined [20]. R1B, the longitudinal relaxation of the bound-pool, was set to a fixed value of 1 s-1 by convention [33–35].
Pre- and post-contrast R1 maps were similarly constructed from the respective VFA data that was acquired in the axial plane. Corresponding pre-contrast B1 maps were similarly applied for correction of α during fitting of both pre- and post-contrast R1 maps. Image processing dedicated to whole-brain voxel-based determination of f maps and R1 maps was performed using in-house written Matlab (The Mathworks, Natick, MA) and C/C++ language software.
Whole blood acquired from each animal was centrifuged at 1,600 g ×10 minutes at 4°C. The plasma layer was removed and centrifuged at 16,000 g ×10 minutes at 4°C. The buffy coat was then collected and stored at -80°C. After centrifugation, plasma was removed excepting a residual amount near the bottom that may have been in contact with any debris and stored at -80°C. Both plasma samples and the buffy coat were stored at -80°C <1 hour from time of collection from the animal.
DNA was isolated from buffy coat cell pellets using the Qiagen DNeasy Blood and Tissue kit. Shotgun sequencing libraries were constructed with 50 nanograms of gDNA from each animal using the Nextera DNA library prep kit (Illumina). Following the manufacturer’s direction, sample index sequences were added during the PCR step to allow libraries to be pooled for multiplexed sequencing on a single lane.
Cell-free DNA was extracted from rat plasma using the QIAamp Circulating Nucleic Acid kit. DNA yield was measured with a Qubit dsDNA HS assay (Invitrogen) and 1–10 ng of cell-free DNA was used as input for library construction with the Thruplex-LC kit (Rubicon Genomics). For samples with low input concentration (<100 pg/ul), cell-free DNA was first concentrated across Zymo Clean-Concentrate-5 column (Zymo Research).
During library construction, enrichment PCR was performed using a BioRad MiniOpticon real-time thermocycler, with SYBR Green I dye (Invitrogen) added to each reaction at a final concentration of 0.25X. Reactions were individually removed upon entering log-phase amplification as indicated by SYBR signal (7–17 cycles).
Libraries were normalized to 2 nM each and pooled for paired-end 101-bp sequencing across four lanes on an Illumina Hiseq 2000 instrument. A 9-bp index read was also collected and used to demultiplex reads according to input sample, requiring fewer than 2 mismatches to the known indices.
For each buffy coat and cell-free DNA library, adapter sequences were trimmed and paired end reads were mapped to human and rat reference assemblies (hg19 and rn5, respectively) using bwa [36]. For each read pair, the species origin (rat or human) was then determined using the mapping status against both references. Only reads that could be unambiguously mapped to one or the other species were included: reads with low mapping quality score (<30) in both species’ references were discarded, as were reads of comparable mapping quality to both references (absolute difference in map quality scores <20). Tumor DNA abundance in each cell-free DNA and buffy coat fraction was then computed as (# human read pairs) / (# human read pairs + #rat read pairs). Fragment length were then takes as the absolute distances between the outermost bases of each pair of forward and reverse ends.
As a quality control check, an aliquot of each xenografted cell line at the time of implantation was genotyped across a panel of 96 human polymorphisms using a custom BeadArray assay performed by the Northwest Genomics Center. Cell lines with identical genotype calls in ≥ 95 of 96 markers were considered to be identical in origin, whereas all other pairs of cell lines shared genotypes at many fewer markers (34–45; S15 Fig).
All procedures were approved by the University of Utah Internal Review Board prior to study initiation. Blood samples were collected in Streck BCT tubes, stored at 4°C, and processed within 24 hours of collection. Plasma was separated by centrifugation for 10 minutes at 1900g and aspiration to a new tube. Plasma was further centrifuged for 16,000g x 10 minutes to remove any cellular debris, and resulting supernatant was stored at –20°C until cell-free DNA isolation. Custom kits that combined Qiagen lysis and binding buffer with Zymo silica-based columns were assembled to reduce expense during isolation of cell-free DNA. Cell-free DNA was prepared from 8 mL of plasma by adding 800 μL of Proteinase K (20 mg/mL) and 6.4 mL Buffer ACL (Qiagen) followed by incubation at 60°C x 30 minutes. Next, 14.4 mL of buffer ACB (Qiagen) was added to the lysate and incubated on ice for 5 minutes. DNA was isolated from the lysate with Zymo DNA Clean and Concentrator 100 kit according to the manufacturer’s instructions and eluted in 150 μL. A final purification step was performed using two volumes of Ampure XP magnetic beads followed by elution in 25–30 μL 10mM Tris (pH 8.0).
For continuous variables, the means and standard deviations (SDs) were calculated for each group. The student’s independent t-test assuming equal or unequal variance based on Levene’s test was used to compare mean values between tumor patients and healthy controls. Pearson’s r was used to identify correlations between continuous variables. Statistical analyses were performed with SPSS for Windows (Version 12.0, SPSS, Chicago, IL). Statistical significance was defined as P < 0.05.
|
10.1371/journal.ppat.1000202 | Rac1 Is Required for Pathogenicity and Chm1-Dependent Conidiogenesis in Rice Fungal Pathogen Magnaporthe grisea | Rac1 is a small GTPase involved in actin cytoskeleton organization and polarized cell growth in many organisms. In this study, we investigate the biological function of MgRac1, a Rac1 homolog in Magnaporthe grisea. The Mgrac1 deletion mutants are defective in conidial production. Among the few conidia generated, they are malformed and defective in appressorial formation and consequently lose pathogenicity. Genetic complementation with native MgRac1 fully recovers all these defective phenotypes. Consistently, expression of a dominant negative allele of MgRac1 exhibits the same defect as the deletion mutants, while expression of a constitutively active allele of MgRac1 can induce abnormally large conidia with defects in infection-related growth. Furthermore, we show the interactions between MgRac1 and its effectors, including the PAK kinase Chm1 and NADPH oxidases (Nox1 and Nox2), by the yeast two-hybrid assay. While the Nox proteins are important for pathogenicity, the MgRac1-Chm1 interaction is responsible for conidiogenesis. A constitutively active chm1 mutant, in which the Rac1-binding PBD domain is removed, fully restores conidiation of the Mgrac1 deletion mutants, but these conidia do not develop appressoria normally and are not pathogenic to rice plants. Our data suggest that the MgRac1-Chm1 pathway is responsible for conidiogenesis, but additional pathways, including the Nox pathway, are necessary for appressorial formation and pathogenicity.
| The fungus Magnaporthe grisea (M. grisea) is an important pathogen in plants and has a great impact on agriculture. Its infection of rice causes one of the most destructive diseases, the rice blast disease, around the world. M. grisea starts infection by producing conidia, which generate infectious structures and determine disease epidemics. However, the mechanism of conidial production is not well-understood. In this study, we have employed genetic and molecular techniques to silence the function of certain genes in M. grisea and found that the Rac1 gene is required for conidial production. Importantly, we have identified the mechanism for the Rac1 requirement in conidial production, which involves the interaction between Rac1 and its downstream effector Chm1. Furthermore, our study shows that the Rac1/Chm1-mediated conidiation is necessary but not sufficient for the pathogenicity of M. grisea in plants. Additional Rac1 effectors such as the Nox gene products are necessary for M. grisea to cause disease symptoms in rice and barley. Our study provides new insights into the mechanism of conidiation and pathogenicity of M. grisea during its infection in plants.
| Magnaporthe grisea (M. grisea) is a good model organism to study plant pathogenic filamentous fungi [1],[2]. In addition, it is closely related to other prominent non-pathogenic model fungi, such as Neurospora crassa and Aspergillus nidulans [3]. The fungus infects many cereal crops such as rice, barley, and wheat, and causes rice blast, which is one of the most severe rice fungal diseases throughout the world [4],[5]. Under field condition, the infection starts with conidia landing on and attaching to a suitable surface of plant tissues with the help of the mucilage in spore tips [6]. Subsequently, the conidia germinate, form appressoria and invade the plant tissues. This is followed by invasive growth of the fungus [7],[8]. After successful colonization, many conidia are produced on the blast lesions and disseminated to new plant tissues and initiate a new infection cycle within 5–7 d. The severity of the rice blast disease epidemics is proportional to the quantity of spores produced in the lesion [9]. Therefore, many disease control strategies try to target conidiation, especially for the chemical control of the fungus [10]. However, the genetic basis and molecular mechanisms of conidiation are not well understood. Previous studies have identified several loci controlling conidiation [11]. Disruption of con5 and con6 abolishes conidial production. A series of other loci (con1, con2, con4, and con7), acting downstream of con5 and con6, affect the development of conidia and sporulation. However, other than Con7p being shown as a transcriptional factor required for the transcription of several genes important for infection-related morphogenesis of the fungus [12], the other loci have yet to be characterized at the molecular level. Mgb1, a G-protein β-subunit, is involved in cAMP signaling that regulates conidiation, surface recognition, and appressorial formation. mgb1 null mutation reduces conidiation, but does not abolish it [13]. In this regard, several other genes, e.g., chm1, show similar functional phenotype to mgb1 [14]. Therefore, the mechanism governing conidiation needs further characterization.
Rac1, a member of the Rho-family GTPases, exists in many eukaryotes [15], regulates actin cytoskeleton organization and cellular morphogenesis in higher eukaryotes [16]. In mammalian cells, the formation of actin-rich cell extensions termed lamellipodia is regulated by Rac [17]. In plants such as Arabidopsis, RAC/ROP GTPases regulate diverse processes ranging from cytoskeletal organization to hormone and stress responses [18]. Moreover, rice Rac homolog, OsRac1, plays a role in disease resistance by activating reactive oxygen intermediate (ROI) production and cell death [19].
Unlike the other Rho GTPases (CDC42, Rho), Rac orthologs are not found in yeast such as Saccharomyces cerevisiae and Schizosaccharomyces pombe. It is of great interest to study the function of Rac homologs in the development of filamentous fungi. In Penicillium marneffei, CflB, a Rac1 homolog, is involved in cellular polarization during its asexual development and hyphal growth but not involved in its yeast growth state at 37°C [20]. The cflB deletion mutants show cell division (septation) and growth defects in both vegetative hyphal and conidiophore cell types. In the human pathogen Candida albicans, Rac1 is not necessary for viability or serum-induced hyphal growth, but it is essential for filamentous growth when cells are embedded in a matrix [21]. In Cryptococcus neoformans, however, a Rac homolog controls haploid filamentation and high-temperature growth downstream of Ras1 [22]. In the pathogenic fungi of plants such as Colletotrichum trifolii, Rac1 functions downstream of Ras and can restore the hyphal morphology of dominant Ras mutants by regulating MAPK activation and intracellular reactive oxygen species (ROS) generation [23]. In another phytopathogenic fungus Ustilago maydis, Rac1 is required for pathogenicity as well as proper cellular morphology and hyphal growth [24]. Recently, Rolke and Tudzynski [25] reported that Rac1 interacts with Cla4, and regulates the polarity, development and pathogenicity in Claviceps purpurea. Thus, Rac GTPases play an important role in fungal development.
In the current study, we investigate the function of MgRac1, a Rac1 homolog in M. grisea, and show that MgRac1, is essential for conidiogenesis, and contributes to the formation of appressorium and pathogenicity of M. grisea through activating its downstream effectors: the PAK kinase Chm1 and NADPH oxidases.
The M. grisea genome encodes a Rac homolog in the locus MGG_02731.5 [2]. It contains five GTP/GDP binding or hydrolysis motifs (G1 through G5) characteristic of Rho-family small GTPases. The conserved G4 motif has a TKLD sequence characteristic of Rac, and is distinct from that found in Rho (T/NKXD) and Cdc42 (TQXD) [16]. We hereafter named it as MgRac1 (Magnaporthe grisea Rac1). The multiple alignment analysis showed that MgRac1 is highly homologous to Rac1 homologs from other filamentous fungi, including the plant pathogens Colletotrichum trifolii (CtRac1, AAP89013, 94% identity), Fusarium graminearum (FgRac1, EAA72031, 93% identity), and Stagonospora nodorum (SnRacA, SNOG_00327.1, 88% identity).
To study the function of MgRac1 in the fungus, we first generated Mgrac1 deletion mutants by replacement of the MgRac1 ORF with a selective marker [the bacterial phosphotransferase (hph) gene], through transformation of protoplasts of the wild-type M. grisea strain 70-15 with the deletion construct pKRA1 (Figure 1A). Deletion transformants were screened by growing on selection media supplemented with hygromycin and by PCR verification of genomic DNA of the transformants. The putative deletion mutants were further confirmed by Southern blotting (Figure 1B) and RT-PCR (Figure 1C). Two deletion mutants ΔMgrac1-19, ΔMgrac1-21, and one ectopic transformant (Ect), which had the marker inserted into regions other than the MgRac1 gene, were selected for further analysis in this study. Furthermore, we constructed a complementation strain Mgrac1-Com by reintroducing the genome DNA sequence including a 1.2-kb promoter region and the ORF of MgRac1.
Conidiation of the wild-type strain (70-15), Mgrac1 deletion mutants (ΔMgrac1-19 and ΔMgrac1-21) and MgRac1 complement strain (Mgrac1-Com) on 10-day-old oatmeal agar cultures were determined. The most striking finding was that conidiation was dramatically reduced by 3 orders of magnitude in Mgrac1 deletion mutants (Table 1). In contrast, the wild-type strain 70-15 and the complement strain were normal in sporulation under the same conditions (Table 1). Of the few conidia that formed in ΔMgrac1-19 and ΔMgrac1-21, most exhibited abnormal, elongated morphology (Figure 2A), which was also observed in a T-DNA insertion line by Jeon [26]. The constriction at the base of the malformed conidia was incompletely formed, and consequently the conidia could not detach normally from the conidiophore as in wild type (Figure 2A). As a result, a basal appendage (BA, Figure 2A) remained attached, similar to that observed in the chm1 deletion mutant [14]. The data indicate that MgRac1 is essential for the conidiogenesis of M. grisea.
We next examined the MgRac1 gene expression profiles at different growth stages by quantitative real-time PCR. The results showed much higher expression level of MgRac1 in conidium than in mycelium, germ tube and appressorium (Table 2), consistent with its important role in conidiation and conidial morphology. Interestingly, the Mgrac1 deletion mutants could still form conidiophores (Figure 2A), even though conidial production was severely reduced.
Although the few conidia from the Mgrac1 deletion mutants had abnormal morphology, over 90% of them germinated after 24 h of incubation at room temperature (data not shown). However, appressorial formation from these mutant conidia was completely blocked on the hydrophobic side of GelBond membranes by 24 h (Figure 2B). In contrast, over 95% of germ tubes formed appressoria in the wild-type strain 70-15 and MgRac1 complement strain Mgrac1-Com under the same conditions (Figure 2B). Even after prolonged incubation (over 72 h), no appressorium was observed in the Mgrac1 deletion mutants.
Frequent branching and curly tips were observed at the terminal mycelia of the Mgrac1 deletion mutant (ΔMgrac1-19). However, Calcofluor staining of cell walls of mycelia showed that the septa were normal except for shorter intervals (Figure 2C). Like 70-15, the ΔMgrac1-19 mutant had one nucleus in each hyphal compartment, suggesting that nuclear division and cytokinesis were normal in the Mgrac1 mutant (Figure 2D). These data indicate that MgRac1 is dispensable for septal formation in the fungus M. grisea. Furthermore, we compared radial hyphal growth of the wild-type strain (70-15), Mgrac1 deletion mutants (ΔMgrac1-19) and MgRac1 complement strain (Mgrac1-Com) on CM agar media. The Mgrac1 deletion mutants produced typical grayish M. grisea mycelia. But the colonies of the Mgrac1 mutants were coralline-like and slightly smaller, due to slower growth rate (Table 1).
Because the Mgrac1 deletion mutants hardly produced any conidia, and were defective in appressorial formation, we used mycelia plugs of the deletion mutants to inoculate wounded rice leaves (Figure 3A), wounded barley leaves (Figure 3B), and rice roots (Figure 3C). No disease symptoms developed either on wounded leaves and rice roots. In contrast, the wild-type strain (70-15), and MgRac1 complement strain (Mgrac1-Com) caused typical rice blast lesions in the same tissues at 4–5 days post-inoculation (dpi) (Figure 3). The data indicate that Mgrac1 deletion mutants are nonpathogenic, and that MgRac1 GTPase is essential for the pathogenicity of M. grisea.
To further investigate the function of MgRac1 GTPase, we constructed both a dominant negative form of MgRac1 by substituting aspartic acid at position 128 with alanine (D128A, DN), and a constitutively active form of MgRac1 by substituting glycine at position 17 with valine (G17V, CA). After transforming the protoplasts of wild-type strain 70-15 with MgRac1-DN and MgRac1-CA, respectively, positive transformants were identified by Southern blot analysis and further characterized as described above. Real-time PCR analysis indicated that there was a 8-fold and 20-fold increase of Rac1 expression in vegetative hyphae of MgRac1-DN and MgRac1-CA mutants compared with the wild-type strain 70-15, respectively (Table 3), suggesting that the transformants expressed the expected dominant alleles of MgRac1.
Like the Mgrac1 deletion mutants, the MgRac1-DN mutant produced malformed conidia (Figure 4A), failed to develop appressoria after germination (Figure 4B) and failed to penetrate the onion epidermis (Figure 4C), and consequently lost pathogenicity on rice either by spraying (Figure 4D) or inoculating wounded leaves (Figure 4E). MgRac1-CA produced only half amount of conidia (Table 1) and they exhibited small but significant (p<0.01) increase in size (Figure 4A) in comparison to the conidia of wild-type strain 70-15 based on the one way ANOVA analysis. The length and width of MgRac1-CA conidia were 22.87±0.11 µm and 10.11±0.15 µm, while those of 70-15 were 21.25±0.07 µm and 9.13±0.03 µm, respectively, in which the mean values and standard deviations were calculated on measurements of 50 conidia per replicate for 3 replicates in 5 independent experiments by using program SPSS V13.0. However, there was no change in the length and width ratio. The conidia from MgRac1-CA were able to adhere to the surface and germinate, but failed to form appressoria on hydrophobic sides of Gelbond membrane (Figure 4B), and only a few appressoria developed on onion epidermis after 48 hours (Figure 4C). Under the same conditions, the conidia of the wild-type strain 70-15 developed normal and well-melanized appressoria (Figure 4B), which penetrated onion epidermis successfully and developed infectious hyphae (Figure 4C). The MgRac1-CA strain failed to cause disease on rice seedlings (Figure 4D), and wounded rice leaves (Figure 4E), probably due to the defect in appressorial development and infectious growth. Although there were some small brown lesions when sprayed with conidial suspensions, these lesions did not produce any conidia even after prolonged incubation in high moisture after detachment for two days. In contrast, the wild-type strain efficiently generated susceptible lesions that all produced conidia after incubation (Figure 4D and 4E). The data indicate that although MgRac1-CA shows opposite effect on conidiogenesis in comparison to MgRac1-DN, their conidia are nonfunctional and defective in appressorial formation and pathogenicity. To confirm that the phenotypes of DN and CA mutants shown in Figure 4 are indeed due to their constitutively active and dominant negative mutations, as opposed to the elevation in Rac1 protein levels, we constructed over-expression (OE) mutant of MgRac1 and compared their phenotypes. Real-time PCR analysis indicated that there was a 23.88±3.01 fold increase of Rac1 expression in vegetative hyphae of the MgRac1-OE mutant, which also affected expression level of Cdc42, Chm1, Nox1 and Nox2 compared with that of the wild-type strain (Table 3). However, the over-expression of MgRac1 had no obvious effect on conidiogenesis (data not shown) and pathogenicity (Figure 4E) of M. grisea, which indicated that the phenotypes of MgRac1-DN and MgRac1-CA mutants are due to their dominant mutations, rather than the elevation in Rac1 expression levels.
Next we examined the effects of MgRac1-CA and MgRac1-DN on actin organization in condia, since Rac1 was shown to play an important role in actin organization in other organisms [27],[28]. In this case, we employed a heterologous tropomyosin-GFP (TpmA-GFP) fusion protein that was previously shown to bind and label actin cables in the filamentous fungus Aspergillus nidulans [29]. This TpmA-GFP cassette was transferred to M. grisea at the background of the wild-type strain Guy11, which had two copies of TpmA-GFP (provided by Dr. Talbot), and the protoplasts were then transformed with MgRac1-CA and MgRac1-DN, respectively. Conidia were collected and examined by Zeiss LSM 510 confocal microscopy at 1 h and 24 h post-incubation. Strong GFP fluorescence was detected in the cytoplasm. At 1 h after the germination began, the TpmA-GFP-labeled actin structures were mostly distributed in the cytoplasm with some discernable actin filaments in the wild-type strain (WT) (Figure 5). The actin filaments were sometimes found attached to bright TpmA-GFP-labeled spots (Figure 5), which resembled the actin bodies in quiescent yeast cells returning to growth [30]. In the MgRac1-CA mutant, however, the labeled actin structures accumulated at the polarization sites and showed bipolar distribution in each of the three cells in the conidium, with actin filaments more evident than in WT (Figure 5). In the MgRac1-DN mutant, some actin structures also accumulated at both ends of the conidium but most TpmA-GFP-labeled actin filaments appeared abnormally straight and striated in the middle of the cytoplasm (Figure 5), which could contribute to its elongated morphology. After 24 h incubation, most of the TpmA-GFP-labeled actin structures in WT moved from the conidium to the appressorium, but they remained in the conidia of the MgRac1 mutants (Figure 5). The data suggest that in the MgRac1-DN and MgRac1-CA mutants, actin is not properly organized and cannot be mobilized for the formation of appressorium and pathogenicity.
To understand the mechanism of MgRac1-mediated conidiogenesis and pathogenicity in M. grisea, we further investigated functional relationship of MgRac1 with Chm1, which is a Cla4 homolog of the baker yeast Saccharomyces cerevisiae. Cla4 is a p21-activated kinase (PAK), which contains a p21-Rho-binding domain (PBD) and a kinase domain. PAK is known to directly transmit signal from Rac/Cdc42 GTPase by acting as a Rac/Cdc42 effector in yeast [31]. The PBD domain is also known as the CRIB domain (Cdc42/Rac-interactive-binding domain) and responsible for interaction with the active form of Rac/Cdc42 [32]. In chm1 deletion mutants of M. grisea, colony growth rate and conidiation are dramatically reduced and of the few conidia produced, most exhibited abnormal morphology and function [14], similar to the phenotype of our Mgrac1 deletion mutants. Moreover, the hyper-branching phenotype in the growing hyphae of the chm1 deletion mutants is the same as that of the Mgrac1 deletion mutants. Thus we examined the relationship between MgRac1 and Chm1. Real-time PCR analysis indicated that there was a 7-fold increase of Chm1 expression in the MgRac1-CA mutant and a decrease in the MgRac1-DN mutant (Table 3). When MgRac1 was deleted, Chm1 transcript was almost undetectable relative to the wild-type 70-15 transcript (Table 3).
We further investigated whether Chm1 can act as a MgRac1 effector to control conidiogenesis and pathogenicity. If Chm1 is MgRac1 effector, it is expected to physically interact with activated GTP-bound MgRac1 and genetically act downstream of MgRac1. We used the yeast two-hybrid assay to test whether the constitutively active and the dominant negative forms of MgRac1 can interact with either full-length Chm1 or the Chm1ΔPBD mutant in which the PBD domain is removed. The results showed that Chm1 was able to interact with the constitutively active, but not the dominant negative form of MgRac1 (Figure 6A and 6B), indicating that Chm1 is an effector of MgRac1. The results also showed that the PBD domain of Chm1 was responsible for this interaction, since deletion of the PBD domain abolished the Chm1-MgRac1 interaction (Figure 6A and 6B).
We next tested whether Chm1 genetically and functionally acts downstream of MgRac1 in conidiogenesis. As a homolog of PAK kinase, the PBD domain of Chm1 is expected to act as an auto-inhibitory domain to suppress the kinase activity [32]. Upon binding to activated Rac1, the PBD domain is released leading to Chm1 activation (Figure 6C). Thus removal of the PBD domain should make the Chm1 PAK kinase constitutively active. To confirm this, a CHM1ΔPBD construct was made under the control of its native promoter and used for transformation of the Mgrac1 deletion mutant and the wild-type strain Guy11 to generate the double mutants PCA19 and PCG33, respectively. Northern blot analysis confirmed the expression of CHM1ΔPBD transcript in the double mutants, which was smaller than the transcript of wild-type CHM1 (data not shown).
We determined the PAK kinase activity in these mutants. Total protein of vegetative hyphae was subjected to in vitro PAK kinase assay using HTScan PAK1 kinase assay kit. As shown in Figure 6D, PAK kinase activity in both PCA19 and PCG33 mutants was increased by more than two-fold over endogenous PAK activity, indicating that the expressed CHM1ΔPBD was active. In a series of control experiments, we found that the ΔMgrac1-19 and MgRac1-DN mutants significantly reduced the PAK activity relative to the WT strains. In contrast, the constitutively active MgRac1-CA mutant greatly increased the PAK kinase activity (Figure 6D). These data demonstrate that MgRac1-DN and MgRac1-CA are effective dominant negative and positive mutants, respectively.
We then focused on the double mutants to investigate the genetic relationship of MgRac1 and Chm1. Indeed, the double mutant PCA19 recovered in conidiation, produced normal conidia both in morphology (Figure 7A) and in quantity like the wild-type strain (Table 1). In addition, the PCG33 mutant showed no obvious defect in morphology and pathogenicity (Table 1). The data indicate that the constitutively active CHM1ΔPBD can fully rescue the conidiogenesis defect in the Mgrac1 deletion mutant, and that MgRac1 genetically acts upstream of Chm1 to activate the conidiogenesis pathway. However, despite normal production and morphology, the conidia of PCA19 were not functional in terms of further appressorial development and pathogenicity (Figure 7). Although the constitutively active CHM1ΔPBD mutant rescued the condiation defect of the Mgrac1 deletion mutant, the constitutively active MgRac1-CA mutant did not rescue the defect of the chm1 deletion mutant (RCC3 and RCC6 in Table 1). The data further support the assumption that Chm1 is a downstream effector of MgRac1 to control conidiogenesis, but additional effectors of MgRac1 are required for pathogenicity of the fungus M. grisea.
M. grisea genome contains two superoxide-generating NADPH oxidase genes, Nox1 and Nox2. The Nox proteins were described as Rac1 effectors in other organisms [33] and it was shown genetically that each was independently required for the pathogenicity of M. grisea [34]. Thus we further investigated if MgRac1 physically interacts with Nox1 and Nox2 and if the interactions play a role in the conidiogenesis and pathogenicity of M. grisea. We first conducted real-time PCR analysis to examine the relationship between MgRac1 and Nox gene expression. There was a 5-fold increase in the levels of Nox1 and Nox2 transcripts in the MgRac1-CA mutant over the wild-type strain 70-15 (Table 3). In contrast, there was a 6-fold decrease in the levels of Nox1 and Nox2 transcripts in the ΔMgrac1-19 and MgRac1-DN mutants (Table 3). This correlation in gene expression between MgRac1 and Nox is similar to that between MgRac1 and Chm1 and suggests that the NADPH oxidases are also potential MgRac1 effectors in M. grisea.
We then tested whether Nox1 and Nox2 can physically interact with MgRac1 and genetically act downstream of MgRac1 as effectors to control conidiogenesis and pathogenicity. We used the yeast two-hybrid assay to determine if the constitutively active and dominant negative forms of MgRac1 interact with Nox1 and Nox2. The results showed that both Nox1 and Nox2 were able to interact with the constitutively active, but not the dominant negative form of MgRac1 (Figure 8A), indicating that Nox1 and Nox2 are indeed MgRac1 effectors.
To determine the effects of deletion and dominant mutations of MgRac1 on ROS production during mycelial and conidial differentiation, we determined NBT content in vegetative hyphae and conidia of the ΔMgrac1-19, MgRac1-CA and MgRac1-DN mutants, and compared with the wild-type strain 70-15. In support of the contention that the Nox proteins are MgRac1 effectors, there was a strong increase in superoxide production in the hyphal tips of the MgRac1-CA mutant, while there was a significant decrease in the ΔMgrac1-19 and MgRac1-DN mutants, as quantified by a reduction in the mean pixel intensity due to the accumulation of localized formazan precipitates [34] (Figure 8B and 8C). These results are consistent with the real time PCR data in which the Nox genes are up-regulated in the MgRac1-CA mutant but down-regulated in the ΔMgrac1-19 and MgRac1-DN mutants (Table 3). Superoxide production in the MgRac1 complement strain Mgrac1-Com was similar to that of 70-15 in both hyphae and conidia (Figure 8B, 8C, and 8D), indicating full recovery of superoxide production. Interestingly, all mutants including MgRac1-CA generated significantly less superoxide than 70-15 in conidia (Figure 8B and 8D), even though MgRac1-CA produced more superoxide in hyphae (Figure 8B and 8C). At present, it is unclear why Nox activity undergoes such dramatic changes in hyphae and conidia of the MgRac1-CA mutant, but the fact that the conidia derived from the MgRac1-CA mutant are nonpathogenic is consistent with a previous report on Nox deletion mutants, which also produce nonpathogenic conidia [34].
Further epistasis analysis was conducted by over-expression of Nox1 or Nox2 in the ΔMgrac1-19 mutant. NBT staining showed increased superoxide production in both conidia and mycelia of the over-expression mutants (Figure 9A, 9C, and 9D). However, over-expression of Nox1 or Nox2 in the ΔMgrac1-19 mutant did not rescue the defect of conidiation (data not shown) and pathogenicity (Figure 9B), even though there was partial recovery in conidial morphology (Figure 9D).
The rice blast fungus M. grisea is an important pathogen, causing rice blast disease in a staple food for half of the world's population [10]. In this study, we show that the Rac1 GTPase plays a critical role in the formation of conidia and appressoria for infection of rice. M. grisea contains one copy of the Rac1 gene (termed MgRac1), which is highly homologous to its mammalian counterpart [2]. We generated Mgrac1 deletion mutants of M. grisea and found that they have severe defect in conidial production. Of the few conidia formed, most are malformed, elongated, and fail to form appressoria. Consequently the Mgrac1 deletion mutants cannot effectively infect rice leaves and roots, leading to loss of pathogenicity. Furthermore, we generated M. grisea transformants that express dominant negative and constitutively active MgRac1 mutants (MgRac1-DN and MgRac1-CA). In support of the data on Mgrac1 deletion mutants, the dominant negative transformant is also defective in the formation of conidia and appressoria and is nonpathogenic. The constitutively active transformant, on the other hand, produces more conidia, with some enlarged than DN mutants. Although these conidia can germinate normally, they are also defective in further development into appressorium for infection of rice leaves and onion epidermis.
Rac1 is a member of the Rho GTPase family and generally functions in actin cytoskeleton organization and polarized cell growth [16], which plays an important role in many developmental pathways of diverse organisms. Indeed in the filamentous fungus P. marneffei, the Rac homolog CflB is required for cell polarization during asexual development, conidiation and hyphal growth [20]. In the phytopathogenic fungus U. maydis, Rac1 is essential for pathogenicity [24]. These observations are consistent with our findings that MgRac1 is essential in M. grisea development and pathogenicity. In addition to M. grisea, other plant-infecting ascomycetes such as C. trifolii, F. graminearum, and S. nodorum all contain Rac homologs. Our data indicate that MgRac1 plays a critical role in the life cycle of M. grisea, specifically in the development of normal infectious structures that allow successful penetration and initiation of plant infection and disease epidemics.
We further identified a Rac1 signaling pathway required for MgRac1-mediated conidiation during the development of M. grisea. In this pathway, active, GTP-bound MgRac1 interacts with Chm1 via its PBD domain, leading to the activation of Chm1 kinase activity that could subsequently regulate actin organization and polarized cell growth during the conidiogenesis process. We provide several lines of evidence to support that Chm1 is a major effector of MgRac1 for conidiogenesis in M. grisea. First, constitutively active Chm1 corrects the defect of Mgrac1 deletion mutants in conidiogenesis in terms of morphology and quantity of conidia. However, it cannot correct the defect in appressorial formation and pathogenicity, suggesting that these processes require additional MgRac1 effectors. Second, constitutively active MgRac1 cannot rescue the defect of chm1 deletion mutants, indicating that Chm1 functions downstream of MgRac1 in the regulation of conidiogenesis. Chm1 is a homolog of mammalian p21-activated kinase (PAK), which is known to interact with and phosphorylate downstream proteins involved in actin cytoskeleton organization and polarized cell growth in mammalian cells [31]. In the dimorphic human pathogenic fungus P. marneffei, PAK is required for conidial germination [35]. In the ergot fungus Claviceps purpurea, Rac1 and its downstream effector Cla4 function in fungal ROS homoeostasis which could contribute to their drastic impact on differentiation [25]. Cla4 also works as Rac1 downstream effector essential for Rac1-induced filament formation in U. maydis [24]. The importance of the MgRac1-Chm1 signaling pathway in the conidiogenesis of M. grisea reflects an evolutionarily conserved Rac1 pathway that controls various developmental processes across species via regulation of actin organization and polarized cell growth.
Chm1 is also an effector for Cdc42 in M. grisea as shown in the yeast two-hybrid assay (data not shown). Our real-time PCR analysis reveals a potential antagonistic interaction between Rac1 and Cdc42 in M. grisea. There is an increase in Cdc42 expression in ΔMgrac1-19 and MgRac1-DN mutants, while there is a small decrease in Cdc42 expression in the MgRac1-CA mutant (Table 3). However, the conidiogenesis defect in ΔMgrac1-19 and MgRac1-DN mutants is unlikely due to hyperactive Cdc42, because over-expression of Cdc42 has no effect on conidiogenesis (data not shown).
The MgRac1-Chm1 pathway, however, is not sufficient for pathogenesis. Although constitutively active CHM1ΔPBD mutant can rescue the conidiation defect of the Mgrac1 deletion mutant, the resulting conidia remain nonpathogenic, suggesting the involvement of additional effectors, such as the Nox proteins that are NADPH oxidases responsible for ROS production. The nox1 and nox2 deletion mutants of M. grisea are known to be defective in pathogenesis [34]. In the current study, we show that MgRac1-CA but not MgRac1-DN interacts with Nox1 and Nox2 and promotes superoxide production in M. grisea, thus confirming that they are MgRac1 effectors. Consistently, we find that Nox activity is up-regulated in the hyphal tips of the MgRac1-CA mutant and down-regulated in the MgRac1-DN mutant.
The data from real time PCR, yeast two-hybrid assay and epistasis analysis indicate that Nox1 and Nox2 act as downstream effectors of MgRac1. Although the Nox proteins are required for pathogenesis [34], our data indicate that MgRac1-Nox interaction is not required in conidiation. Unlike Chm1, over-expression of Nox1 or Nox2 cannot rescue the conidiation defect of the Mgrac1 deletion mutants. Thus, the two MgRac1 signaling pathways play distinct roles in M. grisea differentiation, with MgRac1-Chm1 interaction specifically controlling conidiogenesis.
Magnaporthe grisea (Herbert) Barr parent strains (70-15 and Guy11) and other derivative strains described in this paper were maintained and cultured on the complete medium plates (CM: 0.6% yeast extract, 0.6% casein hydrolysate, 1% sucrose, 1.5% agar) at 25°C. Cultures for genomic DNA isolation, RNA isolation and protoplast preparation were grown in the liquid starch yeast medium (SYM: 0.2% yeast extract, 1% starch, 0.3% sucrose) in a 150-rpm shaker at 25°C for 3–4 d. Conidia were prepared from 10-day-old cultures grown on the oatmeal agar medium (5% oatmeal, 2% sucrose, 1.5% agar) and rice-polish agar medium (2% rice-polish, 1.5% agar, pH 6.0). The selective top agar medium was supplemented with either 400 µg/ml of hygromycin B (Roche Applied Science) or 300 µg/ml of glufosinate ammonium (Sigma-Aldrich Co.), depending on the selection marker in the plasmid vector. Mono-conidial isolation and measurement of conidiation and growth rate were performed as previously described [36].
Two PCR primers 1F and 1R (Table 4) were designed based on Magnaporthe grisea genome database (www.broad.mit.edu/annotation/genome/magnaporthe.grisea). The MgRac1 gene was amplified from the 70-15 genomic DNA by a 30-cycle PCR reaction (94°C, 1 min; 54°C, 1 min; 72°C, 1 min), followed by 7 min extension at 72°C. PCR products were cloned into the pGEM-T easy vector (Promega Corp.) and confirmed by direct DNA sequencing. The cDNA of MgRac1 was isolated by RT-PCR of total RNA of M. grisea with primers 1F and 1R, followed by cloning into the pGEM-T easy vector and direct DNA sequencing (EF060241).
To replace the gene, a 0.9-kb fragment upstream of the MgRac1 ORF in the M. grisea genome was amplified with primers 2F and 2R (Table 4) and cloned into the XhoI sites on pCSN43, and the resulting construct is named pRAC11. Then a 1.0-kb fragment downstream of MgRac1 ORF was amplified with primers 3F and 3R (Table 4) and cloned between the HindIII and SacI sites in pRAC11, and the resulting construct was the MgRac1 gene replacement vector, pKRA1, which had the selective marker hph gene flanked by the MgRac1 ORF flanking sequences. pKRA1 was then transformed into protoplasts of the wild-type strain 70-15 as described previously [37]. Hygromycin-resistant transformants were screened by PCR with primers 4F and 4R (Figure 1A, Table 4) to confirm that the MgRac1 gene was deleted. These transformants were Mgrac1 deletion mutants.
The complementation vector pCRA1 was constructed by cloning a 2.37-kb fragment containing the native promoter and ORF of MgRac1, amplified by PCR with primers 5F and 5R (Table 4), into the basta-resistance vector pBARKS1. The complementary strain Mgrac1-Com was generated by reintroduction of pCRA1 into the Mgrac1 deletion mutants, followed by screening for basta-resistant transformants and PCR confirmation.
The constitutively active and dominant negative MgRac1 mutants (MgRac1-CA and MgRac1-DN) were generated by site-directed mutagenesis of wild type MgRac1 via a PCR-based approach. Two primers including the forward primer 6F and reverse primer 6R (Table 4) were used to generate MgRac1-CA with 6F containing the substitution of the glycine (G17) of MgRac1 with valine. The dominant negative MgRac1 mutant (MgRac1-DN) was generated by substitution of the aspartic acid (D123) with alanine by recombinant PCR with two pairs of primers 1F/7R and 7F/1R, with 7F and 7R containing the mutation (Table 4). Wild type MgRac1 cDNA was amplified with primers 1F and 1R (Table 4) to construct over-expression MgRac1 mutant. All the mutated and wild-type DNA fragments were amplified with pfu polymerase (Stratagene), confirmed by DNA sequencing, and cloned into the vector pTE11. The expression of MgRac1-CA, MgRac1-DN and MgRac1-OE was driven by the constitutive RP27 promoter built within pTE11, upon transformation of protoplasts of the wild-type strain 70-15, the chm1 deletion mutant and the Guy11 strain expressing the heterologous Aspergillus nidulans tropomyosin-GFP [29].
To generate the CHM1ΔPBD (deletion of the PBD domain185–243 in the Chm1 ORF) construct, the genomic DNA of wild-type strain 70-15 was amplified by recombinant PCR with four primers 8F/9R and 9F/8R (Table 4). The resulting PCR product contained the CHM1ΔPBD sequence driven by the native Chm1 promoter. It was then digested with SacI and cloned into pBARKS1, resulting in the CHM1ΔPBD expression vector pBCP17. After transforming the wild-type strain Guy11 and Mgrac1 deletion mutant with pBCP17, basta-resistant transformants were isolated and screened by PCR with primers 8F and 8R to confirm the CHM1ΔPBD sequence. The expression of CHM1ΔPBD in these transformants was confirmed by Northern blot analysis (see below).
M. grisea Nox1 and Nox2 cDNAs were amplified by RT-PCR with primers 26F/26R and 27F/27R (Table 4) and cloned into the XhoI/BamHI sites of pKNTP vector, which contained the constitutive RP27 promoter and the neomycin gene as a selection marker. The pKNTP vector was derived from pKNTG via insertion of the RP27 promoter, which was amplified from pTE11 by PCR with the primers 25F and 25R (Table 4). The resulting Nox1 and Nox2 expressing constructs were termed pOENO1 and pOENO2, respectively. Upon transformation of Mgrac1 deletion mutants with pOENO1 or pOENO2, 300 µg/ml of neomycin sulfate (Amresco Inc.) was supplemented for selection. Neomycin-resistant transformants were screened and Nox expression was confirmed by NBT staining.
For Southern blot analysis, genomic DNA was isolated from M. grisea wild-type strain 70-15, putative Mgrac1 deletion mutants and ectopic transformants, following the miniprep procedure [37]. DNA aliquots of 5 µg were digested with PstI, separated by electrophoresis on 1% agarose gels and transferred onto a Hybond N+ membrane (Amersham Pharmacia Biotech). Interior probe was amplified with primers 10F and 10R (Figure 1A, Table 4), while exterior probe was amplified with primers 11F and 11R (Figure 1A, Table 4).
For Northern blot analysis, total RNA samples (10 µg per sample), which were isolated from growing hyphae of M. grisea using the RNAiso Reagent (Takara Bio Inc.), were separated by electrophoresis on 1% formaldehyde denaturing gel and transferred onto a Hybond N+ membrane (Amersham Pharmacia Biotech). The probe for Northern hybridization was the 0.5-kb Chm1 exon region amplified by primers 15F and 15R (Table 4). For internal control, a 0.73-kb PCR fragment for 18s rRNA (AB026819) was amplified from M. grisea genomic DNA using primers 16F and 16R (Table 4).
For both Southern and Northern blot analysis, probe labeling, hybridization and detection were performed with DIG High Prime DNA Labeling and Detection Starter Kit I (Roche Applied Science), following the manufacturer's instructions.
First strand cDNA was synthesized with the ImProm-II Reverse Transcription System (Promega Corp.) following the manufacturer's instructions. For RT-PCR, a 2 µl aliquot of first-strand cDNA was subjected to 30 cycles of PCR amplification with MgRac1 ORF primers 1F and 1R. The amount of template cDNA was normalized by PCR with a pair of β-tubulin (XP_368640) primers 12F and 12R (Table 4). Twelve microliters of PCR products were analyzed by 1.5% agarose gel electrophoresis.
In quantitative real-time PCR, MgRac1, MgCdc42 (AF250928), Chm1 (AY057371), Nox1 (EF667340) and Nox2 (EF667341) were amplified by the following pairs of primers: 17F/17R, 18F/18R, 19F/19R, 20F/20R, and 21F/21R, respectively (Table 4). As an endogenous control, an 86-bp amplicon of β-tubulin gene was amplified with primers 22F and 22R (Table 4). Quantitative real-time PCR was performed with the MJ Research OPTICON Real-Time Detection System using TaKaRa SYBR Premix Ex Taq (Perfect Real Time) (Takara, Japan). The relative quantification of the transcripts was calculated by the 2−ΔΔCt method [38].
Conidia were prepared from 10-day-old oatmeal agar cultures. For the measurement of the length and width of conidia, five independent experiments were performed with 3 replicates each time, and 50 conidia were observed in each replicate. Mean and standard deviation were calculated using SPSS V13.0, and one way ANOVA was performed on the data for significant differences between genotypes. Aliquots (50 µl) of conidial suspensions (5×104 conidia/ml) were applied on the hydrophobic side of Gelbond film (Cambrex BioScience). The conidial droplets were incubated in a moist chamber at 25°C. Conidial germination and appressorial formation were examined at 0.5, 1, 2, 4, 8 and 24 h post-incubation. Appressorial penetration on onion epidermal strips was assayed as described previously [39]. Photographs were taken with an Olympus BX51 universal research microscope.
Rice (Oryza sativa L.) and barley (Hordeum vulgare cv. Jinchang 1316) seedlings (15 and 8-day-old respectively) were grown under the conditions described previously [36]. The rice cultivar used for infection assays was CO39 [40]. Conidial suspensions (1×105 conidia/ml in 0.02% Tween solution) were prepared from oatmeal agar cultures for spray or wounded infection assays. Plant incubation and inoculation were performed as described [5]. Root infection assays were carried out as described [41]. Lesion formation was examined at 7 days after inoculation on rice and 5 days after inoculation on barley. The mean of lesion numbers formed on 5-cm leaf tips was determined as described previously [42],[43]. Cell walls and septa of vegetative hyphae were visualized by Calcofluor White (10 µg/ml, Sigma), and nuclei of vegetative hyphae were visualized by DAPI (50 mg/ml, Sigma) as described [44].
The MATCHMAKER GAL4 Two-Hybrid System 3 (Clontech) was used to determine protein–protein interactions. The MgRac1 cDNA was amplified with primers 13F and 13R (Table 4) and inserted into the EcoRI and BamHI sites of the yeast vector pGBKT7 (Clontech). MgRac1 contains the C-terminal CAAL motif that is subject to prenylation at the cysteine residue. This modification makes these Rho-family GTPases membrane associated and difficult to enter the nucleus for protein interactions in the two-hybrid assay. Thus, we constructed MgRac1:C196S mutants that cannot be prenylated and is thus soluble. Constitutively active and dominant negative mutations were generated at the MgRac1:C196S background and the resulting double mutants were used as the baits in the two-hybrid assay. Chm1 ORF was amplified with primers 14F and 14R (Table 4) and cloned between the EcoRI and SacI sites on the yeast vector pGADT7 (Clontech) as the prey in the two-hybrid assay. The CHM1ΔPBD cDNA was amplified by recombinant PCR with two pairs of primers (14F/9R and 9F/14R) from the first-strand cDNA of wild-type 70-15, followed by cloning into the EcoRI and SacI sites of pGADT7 as a prey in the two-hybrid assay. Nox1 and Nox2 ORFs were amplified with primers 23F/23R and 24F/24R, respectively (Table 4), and cloned into the yeast vector pGADT7 (Clontech) as the preys in the two-hybrid assay. The resulting bait and prey vectors confirmed by sequencing were co-transformed in pairs into the yeast strain AH109 (Clontech). The Leu+ and Trp+ transformants were isolated and assayed by X-gal staining. Positive clones were further confirmed by plating onto SD-Leu-Trp-His media for the HIS3 reporter gene expression.
In all assays, the interaction of pGBKT7-53 and pGADT7-T was used as the positive control, and the interaction of pGBKT7-Lam and pGADT7-T as the negative control.
Vegetative hyphae were harvested from 3-day-old CM liquid cultures for protein isolation. About 200 mg of mycelia were resuspended in 2 ml of extraction buffer (50 mM Tris-HCl [pH 7.5], 100 mM NaCl, 50 mM NaF, 2 mM phenylmethylsulfonyl fluoride, 5 mM EDTA, 1 mM EGTA, 1% Triton X-100, 10% glycerol) and centrifuged. Protein concentration was measured by GeneQuant pro spectrophotometer (Amersham Biosciences), and 10 µg of total protein was applied for kinase activity detection. PAK Kinase assay was performed by using the HTScan PAK1 kinase assay kit, according to the manufacturer's instructions (Cell Signaling Technology).
For superoxide detection, hyphae of wild-type strain 70-15 and MgRac1 mutants were collected from 3-day CM agar plates and stained with 0.6 mM NBT (nitroblue tetrazolium) aqueous solution for 2 h. Superoxide production in the hyphal tips was viewed by bright-field microscopy. Conidia were collected from 10-day oatmeal agar plates and stained with 0.3 mM NBT aqueous solution for 1 h. After incubation in NBT, the reaction was stopped by the addition of ethanol, and the pattern of formazan staining was observed by using Zeiss Axiovert 200 M microscope equipped with a Zeiss LSM 510 META system. The intensity of formazan precipitation in conidia and hyphal tips was quantified by using Meta Imaging Series 6.1 software (Universal Imaging Corporation) to calculate mean pixel intensity within regions of interest fitted to the outline structure. Measurements were made on the most intensely stained conidia and hyphae of each strain. Pixel intensity was reduced in areas of formazan precipitation.
GenBank accession numbers for genes or proteins used in this article are EF060241 (MgRac1), AF250928 (MgCdc42), AY057371 (Chm1), EF667340 (Nox1), EF667341 (Nox2), XP_368640 (β-tubulin) and AB026819 (18s rRNA).
|
10.1371/journal.ppat.1004838 | Group Selection and Contribution of Minority Variants during Virus Adaptation Determines Virus Fitness and Phenotype | Understanding how a pathogen colonizes and adapts to a new host environment is a primary aim in studying emerging infectious diseases. Adaptive mutations arise among the thousands of variants generated during RNA virus infection, and identifying these variants will shed light onto how changes in tropism and species jumps can occur. Here, we adapted Coxsackie virus B3 to a highly permissive and less permissive environment. Using deep sequencing and bioinformatics, we identified a multi-step adaptive process to adaptation involving residues in the receptor footprints that correlated with receptor availability and with increase in virus fitness in an environment-specific manner. We show that adaptation occurs by selection of a dominant mutation followed by group selection of minority variants that together, confer the fitness increase observed in the population, rather than selection of a single dominant genotype.
| When RNA viruses replicate, they do so with a high rate of error; hence, their populations are not composed of a single genotype, but of a swarm of different, yet related, genomes. This mutant spectrum has been described as the viral quasispecies, and its composition has important consequences for evolution, adaptation and emergence. In this study, we analysed adaptation in fine detail thanks to the use of the deep sequencing, and we determined the adaptative pathway of a model RNA virus, Coxsackievirus B3, to a new environment, A549 cells. Our results demonstrate that adaptation occurred in response to a differential expression of the virus receptors in the new cellular environment, compared to the former. Our experiments and mathematical analyses established that the corresponding increase in fitness resulted from the selection and contribution of a group of genotypes, including low frequency variants, and not to the effect of a single, dominant genome. Our work underscores the importance of considering group effects when studying RNA virus biology and evolution.
| The extreme mutation rates of RNA viruses and the highly diverse populations they generate in few replication cycles are considered the basis for their rapid adaptation to new environments [1,2]. Such adaptive steps result in the emergence of new variants capable of escaping immune responses, resisting antiviral approaches, altering tissue tropism or crossing species barriers. In the past, experimental evolution of viruses in different host environments has proven to be a useful tool in quantifying fitness increases and the dynamics of adaptation. By classic sequencing techniques, some of the key genetic determinants responsible have been identified [3,4], but until the advent of deep sequencing, analysis of the mutational composition of RNA virus populations was hampered by lack of depth of sequence coverage. The potential to describe the whole virus mutant spectrum and detect variants that otherwise would be overlooked by conventional sequencing is fundamental to studying virus evolution and understanding emergence [5]. Recent work shows that deep sequencing can identify the emergence of escape mutants in experimental and clinical samples [6,7], and can be used to characterize the entire mutant spectrum of a virus population [8].
One of the goals in the field of emerging infectious diseases is to determine whether adaptation to novel hosts (species, tissues or cell types) can be identified for a recently introduced pathogen that is confronted with a less than optimal host environment [9–11]. Viruses are well-suited for studying adaptation and evolution for several reasons: i) high mutation rates ii) short generation time and iii) large population sizes. We used Coxsackie virus B3 (CVB3) as a model, since the genetics of this virus and the interactions between the cell receptors and viral capsid proteins (VP1, VP2 and VP3) are well characterized. CVB3 enters the cell through a primary receptor, the Coxsackie and Adenovirus Receptor (CAR) [12], while certain strains may use as co-receptor the Decay Accelerating Factor (DAF) [13,14], also known as CD55. To study expansion of host tropism, we passaged virus in two cellular environments, a highly permissive one and a less permissive one. By deep sequencing longitudinal samples of experimentally evolved populations, we identify the emergence of host environment-specific mutations undergoing positive selection. We show that Coxsackie virus adapts differently to two cell types according to receptor and co-receptor availability in a multi-step adaptation sequence that involves group selection of minority variants. Importantly, we reveal the significant contribution of several minority variants to the overall fitness of the entire population. Our results underscore the importance of characterizing RNA virus quasispecies during adaptation and virus evolution.
To monitor the evolution of CVB3 towards novel and less permissive host environments, we selected human lung A549 cells, which gave similar final virus yields as the highly permissive HeLa cells, but after two days rather than one day of infection. CVB3 was thus serially passaged 40 times in six biological replicate series in both cell types. Virus yields were constant throughout the passage series suggesting that no significant genetic drift or accumulation of detrimental mutations through population bottlenecking had occurred (Fig 1A and 1B). The time required to reach peak titers was reduced in A549 cells over the first ten passages from 48 hours to 24 hours, suggesting that fitness increases and adaptation occurred in this novel environment. We measured the relative fitness of the passaged viruses by competing the passage 1, 20 and 40 populations from each replicate with a genetically marked neutral CVB3 virus in a quantitative fitness assay [15]. Increases in fitness were observed in both cell types, with the most significant increase found by passage 40 in A549 cells (Fig 1C and 1D), suggesting that adaptation had occurred in this less permissive cell type.
Whole-genome deep sequencing of these populations revealed significant increases in overall genetic variation, throughout the genome, between the 1st, 20th and 40th passages in both HeLa (Fig 1E) and A549 (Fig 1F) cells (p<0.0001). The total number of minority variants within the P1 structural coding region also significantly increased over the passage series (Fig 1G), yet there were no significant differences between the numbers observed in Hela and A549 cell passage. The vast majority (>98%) of these variants were low frequency (<1.0% of the total population), suggesting that moderate genetic drift, rather than positive selection, was responsible for most of this variance. The increase in diversity in both cell types is likely the result of general population expansion in sequence space, since all passage series were started from homogenous in vitro transcribed RNA derived from an infectious clone. The data also confirmed that severe bottlenecking did not occur during the passage series where expansion of diversity should have been stalled or lost.
Although the HeLa and A549 cell passage series presented similar mean genetic diversity at passage 40 (Fig 1E–1G, P = 0.314), we mined the deep sequence data for all minority variants above 1% frequency that might explain adaptation in each condition (Table 1). In both cases, several mutations recurring in multiple replicates mapped to residues known to be part of the CAR receptor and DAF co-receptor footprints [16]; however, the distribution of these mutations was different for the HeLa- and A549-adapted viruses. In HeLa cells, the most abundant variants involved mutations strictly in the CAR footprint (VP1-259, 2–7% frequency), in the CAR/DAF shared footprint (VP2-138, 20–80%), and in the DAF footprint (VP3-234, 2–73% frequency; VP3-63, 3–14% frequency). In contrast, viruses passaged in A549 cells presented no variants strictly involved in the CAR footprint. Instead, in addition to CAR/DAF residue VP2-138 (2–81% total), a larger cluster of exclusively DAF-footprint variants were identified (VP3-63, 3–35% frequency; VP3-234, 3–82% frequency and VP1-271, 2–13% frequency). Strikingly, the most abundant mutation in every replicate passaged in A549 cells, and not observed in HeLa cell passage, was VP3-76 (61–95% total). This residue is not known to participate in either footprint.
Considering these CAR/DAF-specific differences, we hypothesized that expression levels of these molecules must vary between these cells. There were no significant differences in CAR expression in either cell by both flow cytometry (Fig 2A) or Western blot (Fig 2C). On the other hand, while DAF was very highly expressed in HeLa cells (Fig 2B and 2D), expression was one order of magnitude lower in A549 cells by flow cytometry (Fig 2B) and barely detectable by Western blot (Fig 2D). Normalized quantification of Western blot signals revealed that DAF expression was 4-fold higher in HeLa than in A549 cells (Fig 2E). To characterize the localization of CAR and DAF in these cell types, confocal microscopy was performed. In HeLa cells, CAR (Fig 2F) and DAF (S1 Fig) were expressed throughout the surface of the cell. Interestingly, in A549 cells CAR predominantly localized at cell-to-cell contacts (Fig 2G), while DAF was diffused throughout the surface of the cell (S1B Fig), albeit at considerably lower levels than in HeLa cells (see also, S1 and S2 Movies).
It has been shown that adaptation is often a multi-step process, especially in asexual populations like viruses. Usually, until the mutation with the largest beneficial effect becomes fixed, secondary beneficial mutations with less effect are unable to compete, thereby rendering the adaptation process sequential [17–19]. Given the different frequencies of minority variants observed at passage 40, we examined the patterns of emergence and sequential adaptation to novel environments by deep sequencing every second passage in the A549 cell series (Fig 3 and S1 Dataset). A number of step-wise trends were observed across the replicates. In all six replicates, the VP3-E76G mutation was the first mutation to emerge (already above 0.1% at passage 1) and peak by passage 11 to become the most abundant mutation in the population (between 65 and 95%). Furthermore, the increase in frequency of E76G over the first 11 passages (Fig 3) correlated with the increases in fitness over the same period (Fig 4A), underscoring a considerable contribution of this single mutation to population fitness. Indeed, the E76G variant was generated by reverse genetics and found to confer significantly enhanced fitness in A549 cells (Fig 4B). Although the VP3-76 residue was not previously identified as a contact between virus and receptor, this residue maps to the icosahedral three fold region of the capsid (Fig 4C). In the three dimensional virus-receptor structure, there is an interaction between the C-terminal 6-His tags of three symmetry related DAF molecules [20]. This interference by the His-tag may have restricted the normal interactions of DAF SCR3/4 with the virus surface, masking a potential role for residue 76. In a biolayer interferometry assay using BLItz technology [21] (Fig 4D), E76G bound DAF with the same affinity as wildtype virus, but no increased binding could be detected.
Despite the clear fitness advantage conferred by the E76G mutation alone (approximately 10-fold increase with respect to wildtype), it could not account for the more significant fitness increases observed in the passaged populations, particularly at late passage stages (nearly 100-fold increases). Following the fixation of this first mutation, the frequency of VP3-E76G plateaued until a second increase after passage 27, in each of the six replicates, coinciding with the emergence of different combinations of mutations, all mapping to the DAF footprint (Fig 3). Closer examination of the patterns of emerging mutations revealed potential synergistic and antagonistic epistasis among variants. Despite the advantage afforded by extreme depth in coverage, the Illumina technology used here set a limit of 69 nucleotides for read length. In principle, this limitation forfeits the possibility of linking more distant mutations and identifying haplotypes. However, the longitudinal deep sequence data revealed that several mutations increase with parallel kinetics and frequency, suggesting that: i) each mutation appeared in individual genomes and the variants were selected as a group and/or ii) the mutations are accumulating in the same genome, resulting in the selection of single haplotypes.
To distinguish between the above cases, phylogenetic trees describing the mix of haplotypes at each passage were inferred from the longitudinal data using maximum likelihood estimation in a Bayesian model. The best-fit, predicted haplotypes were then generated by reverse genetics and their relative fitness values were measured. As expected, this analysis illustrated the quick rise of the E76G genotype that continues to coexist with, although generally dominating over, the original WT genotype (Fig 5A–5F), with a significant increase in relative fitness (Fig 5G). For residue VP3-234, mutations were predicted to arise on either the WT background (Fig 5C and 5F) or the E76G genotype (Fig 5A and 5E). Mutations in residue VP2-138 were also predicted to arise on either WT (Fig 5B) or E76G (Fig 5C, 5D and 5F) genotypes at approximately the same time in the passage series as residue 234. However, their frequencies on the WT background tended to remain lower than on the E76G background where they seemed to be better tolerated. To confirm these modelled predictions, we generated each mutation on both backgrounds and measured the relative fitness. Indeed, the D138G mutation alone on the WT background conferred nearly a ten-fold drop in fitness and the Q234K mutation conferred up to 100-fold drop in fitness (Fig 5G), while the fitness costs of these mutations on the E76G background resulted in neutral fitness relative to WT. The data suggest that epistasis between these mutations and the E76G mutation rescues the fitness of these double variants and permits their positive selection in the viral population. Interestingly, the D138G and Q234K variants seemed to entirely exclude one another from the population (Fig 5A, 5B, 5D and 5E), and the MLE analysis suggested that if both are present in the population, they must occur on separate genomes. We thus generated the E76G-D138G-Q234K haplotype and confirmed that the triple mutant bears a significant fitness cost, rendering this haplotype less fit than the original WT genotype (Fig 5G).
Finally, shortly after the appearance of residue 138 and 234 variants, a third mutation appeared during the passage series at residue 63 (Fig 5A, 5B and 5E). Once again, computational modelling predicted that it too exists on the E76G background, and on separate haplotypes than the position 138 and 234 mutations. Furthermore, data from two replicates (Fig 5A and 5E) suggested that the E76G-N63Y double variant bears higher fitness than E76G-Q234K, as its frequency increased towards the end of the passage series as the other's decreased. Indeed, the residue 63 variants presented higher relative fitness than both the 138 and 234 variants on the E76G background (Fig 5G).
Although we could not determine whether residue 76 plays a role in receptor interaction directly, and the role of accompanying mutations are inferred from previously published studies, to rule out that these mutations impact virus fitness on activities downstream of receptor binding and entry, we transfected cells with in vitro transcribed RNAs corresponding to some of these variants and assayed virus production at 8 hours (before a new round of infection can occur). The data revealed that the fitness advantages (E76G, E76G-N63Y) and disadvantages (WT-Q234K, E76G-Q234K) observed during infection of cells (Fig 5G) do not appear to exist when the binding and entry step is bypassed (Fig 5H).
Although we initially expected selection of fitness-increasing adaptive mutations to occur by step-wise accumulation of mutations on a single genotype, our computational analysis followed by fitness measurements of double and triple mutation-bearing genotypes suggested otherwise. Importantly, none of these single, double and triple mutations conferred the same fitness increases as those observed in the passage 40 virus population used to identify mutant composition (Fig 5G). Since previous work suggested that overall population fitness results from cooperative interactions among key variants in the mutant swarm, we examined whether a reconstituted composition of the most predominant variants within the passage 40 population could manifest comparable fitness. We thus generated artificial quasispecies presenting mixtures of position 63, 138 and 234 mutations on the E76G background and tested their relative fitness. Interestingly, artificial quasispecies presenting 63:138, 138:234 or 63:234 combinations at 1:1 ratios all presented fitness increases as high as, or higher than, the individual values for each variant (Fig 5G). On the other hand, none of these combinations resulted in the highest fitness values that were observed for the passage 40 populations of each replicate used to identify these individual mutations (Fig 5G, p40 a-f).
To address whether the minority variant composition may dictate observed population fitness, we reconstituted an artificial quasispecies based on the average frequency of each variant in passage 40 populations. In contrast to the individual variants or the 50:50 combinations described above, a mixture of 50:30:10:10 of the four most predominant genotypes E76G, E76G-N63Y, E76G-D138G, and E76G-Q234K, reproducibly reached the same fitness values as passage 40 samples themselves (Fig 5G), demonstrating that the global fitness of a virus population is determined by cooperative contribution of minority variants, rather than the dominant genotype alone.
Taken together, the sequence data and in vitro characterization of CAR/DAF expression provide a model for the adaptive dynamics of CVB3 to the two host environments presented here (Fig 6). In HeLa cells, where both CAR and DAF are highly and ubiquitously expressed on the surface, adaptive mutations mapped to both footprints. It is not clear whether the CAR-specific mutations (e.g. VP1-K259M) observed in HeLa cells increased interactions with CAR, or conversely, decreased interactions to facilitate the appearance of other mutations related to the DAF footprint. Because the particular strain of CVB3 Nancy used to initiate the passage series already contains some DAF-specific binding residues not found on other Nancy clones, we cannot speculate with confidence which way evolution would go in HeLa cells. In retrospect, inclusion of a CAR-exclusive binding strain of Coxsackie virus in the passage series would have helped discern the direction of evolution in HeLa cells. In A549 cells, on the other hand, the principal receptor CAR is mainly located in tight cell-cell contacts and likely inaccessible to the virus during initial stages of infection; while DAF is present throughout the surface, but at relatively low levels compared to HeLa cells. Thus, the focus of selection is entirely on residues involved in DAF binding (VP3-63, VP3-234) or shared in the CAR-DAF footprints (VP2-138). The CAR-specific mutations observed at residue VP1-259 in HeLa cells, were not observed in A549 cells. The dominant mutation, VP3-E76G, occurred in all six A549-passaged replicates and in none of the HeLa passages—a residue that was not identified as a determinant by conventional structural studies of receptor usage [20]. In the original structural studies, the interaction of residue VP3-76 could not be determined due to steric hindrance of the 6 His-tag between the DAF short consensus repeat domains, SCR3/4, and the virus surface. When CAR is not accessible, DAF is thought to facilitate translocation of the virus to tight junctions containing CAR [22,23]. The interaction with CAR mediates the transition to the A-particle, a required entry intermediate of expanded structure relative to native virus [24]. Nevertheless, it is possible that E76G improves the fitness of the virus in aspects not related to receptor binding, even if we could not identify these mechanisms when comparing replication cycles.
Our work reveals the power of deep sequencing to monitor the population dynamics of virus adaptation in new environments. By Sanger sequencing, only the E76G mutation would have been identified in each replicate. The remainder of mutants found to be positively selected during adaptation to host environment were all minority variants that could have otherwise been missed in multiple replicates. An issue inherent to deep sequencing technology is the estimated error of the chemistry (0.1% for Illumina sequencing), which has been regarded as a caveat to properly describe RNA virus quasispecies or mutant spectra. This problem was recently resolved by elegant molecular biology techniques to remove background error [8,25]. By applying more stringent bioinformatic treatment in our study, between 1700 and 2500 individual, statistically significant point mutations were identified in the structural protein region of the six replicates of A549-adapted populations; yet only mutations in positions related to the DAF footprint underwent selection and amplification over time (S1 Dataset). We thus show that robust bioinformatic treatment coupled with longitudinal data can circumvent error-related issues by identifying changes in variant frequency that would indicate positive or negative selection. This is particularly relevant to in vivo or clinical samples where the quantity of RNA genomes may be too low for more direct sequencing approaches and would require PCR amplification. Similar, relatively simple experimental studies could be designed to understand the population dynamics and evolution of pathogens in new environments during host adaptation and host switching, by identifying the selection over time of one or more mutations as single or multiple genotypes. However, one must keep in mind that although in vitro studies, such as ours, using immortalized cell lines may facilitate studying the dynamics of adaptation; but the specific mutations that are identified may not be indicative of the panel of mutations that would arise in vivo.
In this work, adaptation to a novel host environment was a multistep process involving the emergence of an initial mutation (E76G), followed by selection of a number of minority variants on the E76G background. Initially, we expected mutations to accumulate in a step-wise manner in a single genotype; however, fitness assays revealed that genotypes harboring double and triple combinations of these mutations presented fitness decreases relative to the wildtype and/or E76G backgrounds. Instead, computational inference of haplotypes suggested that such variants existed as a heterogeneous population of distinct genotypes. An intriguing observation was that no single variant (nor combination of two variants) conferred the high fitness values observed in the passage 40 populations. Strikingly, only the combination of the four most predominant genotypes, as an artificial quasispecies with the same frequencies observed in the passage 40 populations, was able to confer the same fitness values as the p40 replicate samples. This phenomenon is reminiscent of previous work that suggested that minority variants within an RNA virus quasispecies may contribute significantly to phenotype [26]. In that study, a high-fidelity poliovirus with restricted quasispeces composition was unable to infect the central nervous system (CNS); while the same virus stock that was chemically mutagenized to present wildtype numbers of minority variants restored the ability to disseminate to the CNS. Unfortunately, deep sequencing technology was not yet available and the authors could not uncover the identity of these presumed minority variants by Sanger technology to confirm their hypothesis. Here, we succeed in identifying the variants involved. Importantly, we provide evidence for group selection within the virus population and show that only the group contribution of these positively selected minority variants confers the fitness phenotype observed in the original samples. Our results thus illustrate the significant role of minority genomes in the fitness and phenotype of a virus population, providing further evidence for the quasispecies behavior of RNA viruses under certain conditions. It is important to note, however, that the group selection observed in this study resulted from relatively large population size in passages that reached high MOI by the time cell monolayers were lysed for subsequent passage. It is possible that in conditions where MOI is low, or when population bottlenecks occur (particularly in vivo), that the emergence of such minority variants would be delayed or impeded.
HeLa and A549 cells (American Type Culture Collection) were maintained in DMEM medium with 10% new-born calf serum. Coxsackie virus B3 (Nancy strain) was recovered from a pCB3-Nancy infectious cDNA plasmid [27] that was linearized with Sal I and in vitro transcribed using T7 RNA polymerase. It should be noted that unlike other Nancy strains, this infectious clone already has VP2-138D and VP3-234Q residues, known to facilitate binding to DAF. 4 μg of transcript were electroporated into 4 x 106 Vero cells that were washed twice in PBS and resuspended in PBS at 107 cells/ml. For A549 cells, 15 μg of transcript were electroporated into 10 x 106 cells. Electroporation conditions were as follows: 0.4mm cuvette, 950 μF, 250V, maximum resistance, exponential decay in a Biorad GenePulser XCell electroporator. Cells were recovered in DMEM-10% NCS. For each passage (40 passages total), virus was titrated by TCID50 and 400 μl of medium containing 2 x 105 TCID50 was used to infect 2 x 106 Hela or A549 cells in 6 well plates using a multiplicity of infection (MOI) of 0.1. Cells were incubated with virus for 45 minutes with frequent rocking, the supernatant was removed and monolayers were washed twice with 2ml PBS, then replenished with 2ml of complete medium. For each passage, virus was harvested at total cytopathic effect (CPE) by one freeze-thaw cycle, representing 2–3 viral generations. Six biologically independent stocks and passage series were generated.
Ten-fold serial dilutions of virus were prepared in 96-well flat-bottom plates in DMEM. Dilutions were performed in octuplate and 100 μl of dilution were transferred to 104 Vero cells plated in 100 μl of DMEM-10% NCS. After 5 days living cell monolayers were colored by crystal violet. TCID50 values were determined by the Reed and Muensch method. No significant differences were observed between TCID50 values when using Vero, HeLa or A549 cells as the cellular substrate.
5x108 virion from passaged samples were RNA extracted and RT-PCR amplified by RT (Superscript III) and PCR (Phusion) using primers sets that covered the whole genome, in 3–4 kb fragments. For consensus sequencing, the resulting PCR products were purified, sequenced and analyzed using Lasergene software (DNAStar Inc). For deep sequencing, PCR fragments were purified via the Nucleospin Gel and PCR Clean-up kit (Macherey-Nagel) and total DNA was quantified by Nano-drop. PCR products were then fragmented (Fragmentase), linked to Illumina multiplex adapters, clusterized and sequenced with Illumina cBot and GAIIX technology. Sequences were demultiplexed by CASAVA with no mismatches permitted. Clipping was performed using the fastq-mcf tool, removing common adapter contaminants and trimming low quality bases (Phred<30). Clipped reads were aligned to the Coxsackie virus B3 Nancy sequence as reference with a maximum 2 mismatches per read, and no gaps, using BWA v0.5.9. Alignments were processed using SAMTools to obtain a pileup of the called bases at each position. An in-house pipeline, termed ViVAN (Viral Variant ANalysis) [28] was used to identify statistically significant variants above the background noise due to sequencing error, in every sufficiently covered site (>100x). Briefly, for each position throughout the viral genome, base identity and their quality scores were gathered. Each variant was determined to be true using a generalized likelihood-ratio test (used to determine the total number of minority variants) and its allele rate was modified according to its covering read qualities based on a maximum likelihood estimation. Additionally, a confidence interval was calculated for each allele rate. In order to correct for multiple testing, Benjamini-Hochberg false-discovery rate of 5% was set. The total allele rates passing these criteria, across the whole genome, were used to calculate the mean variation rates (diversity) at different passages. The variation rate at position i is defined as the proportion (F) of significant non-reference alleles (k) and is denoted Vi:
Vi=∑j=1kFij
The region-wide variation rate is the averaged variation rate across all covered positions in the genome (denoted n):
V=∑i=1nVin
Hela and A549 cells were plated onto coverslips, fixed with 2% paraformaldehyde for 20 minutes at room temperature, and then washed with PBS. Before staining, non-specific staining was blocked by incubating the cells with 5% FCS and 0.05% saponin in PBS during 10 minutes. Staining antibodies were diluted also in this buffer, and it was used for washes between antibodies. Cells were incubated with either CAR (Santacruz) or DAF (Abcam) primary antibody, washed, stained with a secondary antibody coupled to the appropriate fluorophore and washed. Cells were analyzed using a Zeiss LSM-700 confocal microscope. 3D reconstruction of the images was performed using the Imaris software.
One million cells of HeLa and A549 cells were lysed using a buffer containing 1% Triton-X and 1% sodium deoxycholate with protease inhibitors (Sigma). A fraction of the lysate was run for 1 hour on a 4–15% gradient gel (Biorad) on denaturalizing conditions. After the run, we performed the transfer to a nitrocellulose membrane. We washed the membrane with PBS-T (PBS 1X and 0.1% Tween-20) and blocked for 1 hour in PBS-T plus 5% milk. After the blocking we washed again with PBS-T and left overnight with each one of the antibodies, anti-CAR and anti-DAF (Santacruz Biotechnology). We washed the membrane and we added the secondary fluorescent antibodies (DyLight 680 and DyLight800 conjugated, Thermo Scientific) for 1 hour. We washed the membrane one last time and measure the fluorescence in the Odyssey system (Li-Cor).
In order to prepare cells for cell cytometry analysis, cells were washed twice with PBS 1x and trysinized, washed again twice in PBS. Cells were stained for 30 minutes with either CAR-PE or DAF-FITC antibodies (Millipore and Abcam) on ice. Unbound antibody was discarded and cells were washed again with PBS 1x. Cells were resuspended in 1% Parafolmaldehyde (PFA, Electron Microscopy Sciences) and kept in the dark for 15 minutes. After the incubation PFA was discarded and cells were resuspended in 200 μl of PBS 1x. Cells were kept at 4C until analysed. For each cell type used (HeLa and A549) specific instrument settings were set according to the size and complexity of the cell type, as well as antibodies fluorescence. Samples were analysed using the MACSquant flow cytometer (Miltenyi Biotec) using 96 well plates and obtaining 10,000 events per sample. Mock samples were also used in each plate to setup the baseline. Results were analysed using Flowjo software v10.
Relative fitness values were obtained by competing each virus population with a marked reference virus that contains four adjacent silent mutations in the polymerase region introduced by direct mutagenesis. Co-infections were performed in triplicate at an MOI of 0.01 using a 1:1 mixture of each variant with the reference virus for 24 hours. The proportion of each virus was determined by real time RT-PCR on extracted RNA from the infection supernanant, using a mixture of Taqman probes labelled with two different fluorescent reporter dyes. MGB_CVB3_WT detects WT virus (including the fidelity variants) with the sequence CGCATCGTACCCATGG and labelled at the 5’ end with a 6FAM dye (6-carboxyfluorescein) and MGB_CVB3_Ref containing the four silent mutations: CGCTAGCTACCCATGG was labelled with a 5’ VIC dye. Each 25 μl-reaction contained 5ul of RNA, 900nM of each primer (forward primer, 5’-GATCGCATATGGTGATGATGTGA-3’; reverse primer, 5’-AGCTTCAGCGAGTAAAGATGCA-3’) and 150 nM of each probe. The relative fitness was determined by the method described by Carrasco et al. [15]. Briefly, the formula W = [R(t)/(R (0))] ^(1/t), represents the fitness, W, of each mutant genotype relative to the common competitor reference sequence, where R(0) and R(t) represent the ratio of mutant to reference virus densities in the inoculation mixture and t days post-inoculation, respectively. The fitness of the normal wildtype to reference virus was 1.019, indicating no significant differences in fitness due to the silent mutations engineered in the reference virus.
CVB3 parental strain, and CVB3-E76G were propagated in HeLa cells and purified as described previously [14]. The membranes of infected cells were broken by three freeze-thaw cycles. Virus was concentrated by pelleting through sucrose and purified by tartrate step gradient ultracentrifugation. The virus bands were collected and exchanged into PBS and virus concentration and quality was estimated by measuring the absorbance at 260, 280, and 310 nm. Binding assays were performed by biolayer interferometry (BLI) measured by the BLItz from Fortebio [21]. Briefly, BLI measures binding by sending white light down a glass fiber-based biosensor, which is reflected back up to the instrument from two interfaces: 1) the interface between the glass fiber and the biosensor, and 2) the interface between the surface chemistry and solution. Since the two reflections come from the same white light source in the instrument, they both contain the same wavelengths. When molecules bind to the surface of the biosensor, the path length of the reflection (the one reflecting from the interface between surface and solution) increases while that of the other reflection remains the same.0.4% BSA and 0.08% Tween was added to virus and DAF to prevent non-specific binding to the sensor during the assay. Purified, His-tagged DAF was diluted to 0.1 mg/ml and attached to a Ni-NTA sensor. The DAF loaded sensor was then dipped into virus diluted to 0.5 mg/ml.
Let X be a phylogenetic tree, taken to include relative population sizes of the branches, and let Xt = (stpt)T be the state of the tree at time t = 1, 2,…,T. Here st denotes the structure of the tree, i.e. the set of branches existing at time t and pt is a vector with the relative population sizes of all existing branches. Furthermore, let Y be the set of measurements, where Yt ∊ ℤ+64xN is the number of reads supporting each of the 64 codons at each of the N different positions in the genome. By Bayes' theorem,
P(X|Y)∝P(Y|X)P(X)
The output probability P(Y | X) follows a multinomial distribution for each position. The dynamics of the phylogenetic tree are naturally assumed to be Markovian, i.e.
P(X)=P(X1)∏t=2TP(Xt|Xt−1),
where the transition probability can be expressed using the structural and population dynamics parts separately,
P(Xt|Xt−1)=P(pt|st,st−1,pt−1)P(st|st−1,pt−1)=P(pt|st,pt−1)P(st|st−1,pt−1).
A simple random walk model is used for the population dynamics P(pt | st, pt-1), the population change from t−1 to t is taken to be lognormally distributed with standard deviation σh where h is the length of the time step. The structural part, P(st | st−1, pt−1), depends on the mutation frequency of the virus and the population size of the haplotype in which the mutation occurs. The posterior probability is thus, assuming that s1 and p1 are known,
P(X|Y)∝P(Y|X)∏t=2TP(pt|st,pt−1)P(st|st−1,pt−1)
To make a maximum likelihood estimation (MLE) of X from the posterior tractable, some approximations are made. The attention is limited to a small set of variants, greatly reducing the number of possible tree structures. The most prevalent variants in the data should be included as they dominate the dynamic behaviour, but minority variants of particular interest can also be added.
Every possible tree structure matching the selected set of variants is generated. The time of appearance for each variant is set to where it is first seen in the measurements, i.e. the first time the frequency is above a small threshold. An MLE of P(pt | st, Yt) is then computed for each tree and time point. The rationale is that there is very little freedom for the population sizes to deviate from a point which can explain the output data. By construction of the tree structure, the number of non-reference haplotypes equals the number of variants at each time point in every tree. Hence, there is at most one convex combination of the haplotypes that match the event probabilities of the multinomial distribution for the output data. Due to the high number of reads in the deep sequencing data, moving away from the optimal point will cause a quick drop in the posterior probability. Dependencies between variants that are close enough to be covered by a single read are included in the model (amino acid residues 63 and 76).
The MLE of X can be found by evaluating the posterior for each generated tree and picking the most likely. All generated trees contain the same number of mutations. Since the value of the posterior probably only needs to be known up to a multiplicative constant, the effect of the overall mutation frequency of the virus on P(st | st−1, pt−1) cancels out. Hence, P(st | st−1, pt−1) is simply proportional to the population size of the haplotypes in which the mutations occur.
|
10.1371/journal.ppat.1000650 | The E3 Ubiquitin Ligase Triad3A Negatively Regulates the RIG-I/MAVS Signaling Pathway by Targeting TRAF3 for Degradation | The primary role of the innate immune response is to limit the spread of infectious pathogens, with activation of Toll-like receptor (TLR) and RIG-like receptor (RLR) pathways resulting in a pro-inflammatory response required to combat infection. Limiting the activation of these signaling pathways is likewise essential to prevent tissue injury in the host. Triad3A is an E3 ubiquitin ligase that interacts with several components of TLR signaling and modulates TLR activity. In the present study, we demonstrate that Triad3A negatively regulates the RIG-I RNA sensing pathway through Lys48-linked, ubiquitin-mediated degradation of the tumor necrosis factor receptor-associated factor 3 (TRAF3) adapter. Triad3A was induced following dsRNA exposure or virus infection and decreased TRAF3 levels in a dose-dependent manner; moreover, Triad3A expression blocked IRF-3 activation by Ser-396 phosphorylation and inhibited the expression of type 1 interferon and antiviral genes. Lys48-linked ubiquitination of TRAF3 by Triad3A increased TRAF3 turnover, whereas reduction of Triad3A expression by stable shRNA expression correlated with an increase in TRAF3 protein expression and enhancement of the antiviral response following VSV or Sendai virus infection. Triad3A and TRAF3 physically interacted together, and TRAF3 residues Y440 and Q442—previously shown to be important for association with the MAVS adapter—were also critical for Triad3A. Point mutation of the TRAF-Interacting-Motif (TIM) of Triad3A abrogated its ability to interact with TRAF3 and modulate RIG-I signaling. TRAF3 appears to undergo sequential ubiquitin “immuno-editing” following virus infection that is crucial for regulation of RIG-I-dependent signaling to the antiviral response. Thus, Triad3A represents a versatile E3 ubiquitin ligase that negatively regulates RIG-like receptor signaling by targeting TRAF3 for degradation following RNA virus infection.
| RNA virus infection is detected through TLR-dependent and TLR-independent mechanisms. Early viral replicative intermediates are detected by two recently characterized cystolic viral RNA receptors, RIG-I and MDA-5, leading to the production of pro-inflammatory cytokines and type I interferons (IFNs). Dysfunctional responses, either failure to respond or hyper-responsiveness, may lead to both acute and chronic immunodeficiency and inflammatory diseases. Thus, the intensity and duration of RLR signaling must be tightly controlled. One general mechanism by which innate immune receptors and their downstream adapters are regulated involves protein degradation mediated by the ubiquitination pathway. Our study demonstrates that the E3 ubiquitin ligase Triad3A negatively regulates the RIG-I-like receptor pathway by targeting the adapter molecule TRAF3 for proteasomal degradation through Lys48-linked ubiquitin-mediated degradation. Thus, Triad3A represents a key molecule involved in the negative regulation of the host antiviral response triggered by RNA virus infection.
| Upon recognition of specific molecular components of viruses, the host cell activates multiple signaling cascades that stimulate an innate antiviral response, resulting in the disruption of viral replication, and the mobilization of the adaptive arm of the immune system. Central to the host antiviral response is the production of type 1 interferons (IFNs), a large family of multifunctional immunoregulatory proteins. Multiple Toll like receptor (TLR)-dependent (TLR-3, -4, -7 and 9) and RIG-I-like receptor (RLR) pathways are involved in the cell specific regulation of Type I IFNs, with accumulating evidence that cooperation between different pathways is required to ensure a robust and controlled activation of antiviral response [1],[2],[3]. RIG-I-like receptors (RLRs) - the retinoic acid-inducible gene-I (RIG-I) and melanoma differentiation-associated gene-5 (MDA-5) - are novel cytoplasmic RNA helicases that recognize viral RNA present within the cytoplasm. Although both TLR7 and TLR9 are critical for recognition of viral nucleic acids in the endosomes of plasmacytoid dendritic cells (pDCs), most other cell types recognize viral RNA intermediates through the RLR arm of the innate immune response [4],[5],[6]. Structurally, RIG-I contains two caspase activation and recruitment domains (CARD) at its N-terminus and RNA helicase activity in the C-terminal portion of the molecule [4]. The C-terminal regulatory domain (CTD) (aa 792–925) of RIG-I binds viral RNA in a 5′-triphosphate-dependent manner and activates RIG-I ATPase inducing RNA-dependent dimerization and structural alterations that enable the CARD domain to interact with other downstream adapter protein(s) leading to the transcription of antiviral genes [7],[8],[9].
RIG-I-dependent signaling to the IKKα/β complex and to TBK1/IKKε is transmitted via a CARD domain containing adapter molecule – alternatively named mitochondrial antiviral signaling (MAVS), interferon-β stimulator 1 (IPS-1), virus induced signaling adapter (VISA), CARD adapter inducing IFN-β (CARDIF) [10],[11],[12],[13]. MAVS localizes to the outer mitochondria membrane via a C-terminal mitochondrial transmembrane targeting domain (TM), and its mitochondrial localization acts as a pivotal point for triggering the antiviral cascade via activation of NF-κB and IRF-3 [3],[14],[15],[16].
Activation of TLRs and RLRs results in the dissemination of an antiviral and antimicrobial cascade necessary to combat invading pathogens [17],[18],[19]. Limiting the intensity and duration of TLR and RLR signaling is likewise essential to prevent this protective response from causing inflammatory or autoimmune injury to the host. Ubiquitination is a post-translational modification by which signaling is suppressed in many regulatory pathways [20]. Lys48-linked ubiquitination is one of the most common pathways to target proteins for 26S proteasomal degradation [21], whereas Lys63-linked ubiquitination is involved in protein-protein interactions, recruitment, and assembly of signaling complexes [22],[23]. It has become clear that ubiquitination of signaling adapters is an integral part of NF-κB and IFN signaling in response to virus pathogen associated molecular patterns (PAMPs). Deubiquitinating enzymes that remove Lys63-linked ubiquitin are also emerging as key negative regulators of the IFN and NF-κB pathways [16],[24],[25],[26],[27]. For example, the deubiquitinating enzyme A (DUBA), a novel OTU-domain DUB negatively regulates IFN signaling following RIG-I, MDA5 or TLR3 stimulation [28]. DUBA specifically removes Lys63-linked ubiquitin chains from TRAF3, resulting in the disruption of interaction between TRAF3 and the downstream kinases IKKε and TBK1 and subsequent blockade of IRF-3 and IRF-7 phosphorylation [28].
The activation of RIG-I/MDA-5 ultimately leads to the TM-dependent dimerization of the MAVS N-terminal CARD domains, thereby providing an interface for direct binding to and activation of the tumor necrosis factor (TNF) receptor-associated factor (TRAF) family members that are involved in both the IFN and NF-κB arms of the innate immune response [29],[30]. TRAF3 is an adapter molecule that is required for the induction of type I IFN and anti-inflammatory cytokine interleukin-10 (IL-10), but is dispensable for expression of pro-inflammatory cytokines in response to viral infection and TLR ligation in bone marrow-derived macrophages (BMMs), plasmacytoid dendritic cells (pDCs), and murine embryonic fibroblasts (MEFs) [31],[32].
TRAF3 was the first TRAF demonstrated to directly associate with CD40. Subsequently, it was shown that TRAF3 negatively regulates CD40 signaling by competing with TRAF2 for CD40 binding, thus impeding CD40-TRAF2 mediated JNK and NF-κB activation [33]. Crystal structure of the binding crevice of TRAF3 bound in complex with a 24-residue fragment of the cytoplasmic portion of BAFF receptor (BAFF-R), revealed two amino acids in TRAF3 -Y440A and Q442- that are involved in BAFF-R interaction [34]. Interestingly, other TNFRs such as CD40 contain similar TRAF-interacting motifs (TIMs), defined by the consensus sequence PxQx(T/S), that interact with the same binding crevice on TRAF3 [35],[36]. In addition, the TRAF family member–associated NF-κB activator (TANK) adapter and the viral oncogene LMP1 of the Epstein Barr Virus also bind to the same structural crevice of TRAF3 [37],[38]. MAVS regulation of type I IFN induction is achieved by direct and specific interaction with the TIM of TRAF3; interestingly point-mutation of the TIM domain completely abrogates TRAF3-mediated IFN-α production in response to Sendai virus infection [39].
Triad3A is a RING finger type E3 ubiquitin-protein ligase that promotes Lys48-linked ubiquitination and proteolytic degradation of TLR4 and TLR9 and negatively regulates their activation by lipopolysaccharide and CpG-DNA, respectively [40]. Triad3A is the most abundant alternatively spliced form of the Triad family. In addition, Triad3A interacts and promotes down-regulation of two TIR domain containing adapter molecules, TIR-domain-containing adapter-inducing IFN-β (TRIF) and TRIF-related adapter molecule (TIRAP). Moreover, Triad3A acts as a negative regulator of TNF-α signaling by interacting with the TIR homologous (TIRH) domain containing protein receptor-interacting protein 1 (RIP1) [41]. This interaction effectively disrupts RIP1 binding to the TNF-R1 complex and impedes RIP-1-mediated NF-κB activation [41].
The identification of a TIM sequence in the N-terminus of Triad3A -using a program written in python language (http://www.biopython.org)- as well as the previously characterized function of Triad3A in TLR signaling, prompted us to investigate the role of Triad3A in the regulation of the RIG-I/MAVS signaling via TRAF3. In the present study, we demonstrate that Triad3A negatively regulates the RIG-I signaling pathway through Lys48-linked ubiquitin-mediated degradation of TRAF3, resulting in the inhibition of the type I IFN response.
The identification of a TIM domain in Triad3A prompted us to examine the ability of Triad3A to inhibit RIG-I mediated activation of IFNB gene transcription; a constitutively active form of RIG-I (aa 1-229, ΔRIG-I), the MAVS adapter or IKKε, were co-expressed together with Triad3A in 293T cells, together with an IFNB promoter luciferase reporter. A low basal activity of the IFNB promoter was not affected by Triad3A expression (Figure 1A), while co-expression of ΔRIG-I, MAVS, or IKKε resulted in 196, 132, 61-fold stimulation of the IFNB promoter, respectively (Figure 1A). Co-expression of Triad3A with ΔRIG-I or MAVS resulted in a complete inhibition of IFNB promoter activity, whereas IKKε mediated activation of the IFNB promoter remained unchanged (Figure 1A). Similar results were also obtained with the NF-κB response (Figure 1B); expression of ΔRIG-I, MAVS or IKKε, (co-expressed together with IRF-7) activated IFNA4 promoter activity 34, 18, 49-fold, respectively, while co-expression of Triad3A blocked IFNA4 activation (Figure 1C). Furthermore, Triad3A blocked interferon stimulated response element (ISRE) activation following Sendai virus infection (Figure 1D). A dose-response curve was performed using the ISRE promoter with increasing amounts of Triad3A and ΔRIG-I, MAVS, TRIF, or TBK1 expression plasmids; ΔRIG-I resulted in 893-fold induction of the ISRE promoter, and Triad3A co-expression diminished activation in a dose dependent manner (Figure S1A). Similarly, MAVS or TRIF adapters activated the ISRE by 785- and 863-fold, respectively; Triad3A again dramatically reduced ISRE activation (Figure S1B, S1C). In contrast, Triad3A did not significantly decrease TBK1-mediated ISRE activation (Figure S1D). Triad3A co-expression with MDA5 or an active form of TLR3 fused to CD4 (CD4-TLR3) resulted in a complete inhibition of IFNB promoter activity (Figure S2A). Triad3A inhibited MDA5-induced NF-κB promoter activity; however Triad3A inhibition of CD4-TLR3 mediated NF-κB promoter activity was less pronounced (Figure S2B). These experiments suggested that Triad3A was a strong inhibitor of RIG-I signaling to IRF-3, IRF-7 and NF-κB and suggested that Triad3A may target an adapter molecule common to both the TLR and RLR signaling pathways.
As a measure of activation of the IFN signaling pathway, the phosphorylation state of IRF-3 was evaluated by immunoblot in the presence of Triad3A using the phosphospecific Ser-396 IRF-3 antibody [42]. ΔRIG-I co-expression induced Ser-396 IRF-3 phosphorylation (Figure 2, lane 3), while co-expression of Triad3A completely blocked IRF-3 phosphorylation (Figure 2, lane 4). MAVS expression likewise induced Ser-396 IRF-3 phosphorylation (Figure 2, lanes 3–5); that was abrogated by Triad3A (Figure 2, lanes 4–6). In contrast, TBK1 co-expression in the presence or absence of Triad3A did not alter the IRF-3 phosphorylation state (Figure 2, lanes 7–8). Complementing the phosphorylation status, Triad3A also inhibited ΔRIG-I and MAVS-induced dimerization of endogenous IRF-3 (Figure 2, lanes 4–6), but did not affect TBK1-induced IRF-3 dimer formation (Figure 2, lanes 7–8), indicating that Triad3A targets RLR signaling upstream of TBK1.
Previous studies demonstrated that the E3 ligase RNF125−a negative regulator of RIG-I− was induced following IFN-α and poly(I∶C) treatment [43]. Endogenous Triad3A protein was induced in human bronchial epithelial A549 cells following dsRNA treatment for 6h, vesicular stomatitis virus (VSV), or Sendai virus (SeV) infection for 16h; correlating with the degradation of TRAF3 protein (Figure 3A). Moreover, Triad3A protein expression is induced following IFN-α/β treatment (data not shown). In addition, it was determined by time-course analysis that 6h dsRNA treatment and 16h virus infection resulted in maximal TRAF3 degradation (Figure S3). Expression of increasing amounts of Triad3A decreased TRAF3 levels in a dose-dependent manner (Figure 3B). Additionally, SeV-mediated degradation of TRAF3 in A549 cells was blocked by the proteasome inhibitors lactacystin and Mg132, but not by the lysosomal protease inhibitor E64 (Figure 3C).
To further confirm the involvement of Triad3A in regulating TRAF3 turnover, two shRNA expression vectors - shRNA1 and shRNA2 that target Triad3A nucleotide sequences 1,532–1,551 and 1,195–1,214, respectively – were used to stably knock-down Triad3A in A549 cells. Knock-down of Triad3A resulted in a 5-fold increase in TRAF3 protein levels (Figure 4A). Interference with endogenous Triad3A also modulated the ISRE promoter; ISRE activity was 3-fold higher in Triad3A knock-down cells infected with SeV, compared to cells expressing scrambled shRNA (Figure 4B). [43]. To investigate the physiological effects of Triad3A inhibition on downstream IFN-stimulated target genes, expression of multiple ISGs were examined by quantitative PCR in A549-Triad3A knock-down cells. SeV infection (40 hemagglutination units/ml (HAU)) in Triad3A knockdown cells were led to a 3–4 fold increase in IFN-β and IFN-α2 mRNA expression 12h post-infection (p.i.) compared to control cells (Figure 4C). Similarly, IP-10 ISG56, IS15 transcripts were increased 3–4 fold at 12h p.i. (Figure 4C), while STAT1 levels remained relatively constant (Figure 4C). In addition, levels of IFN-α and IFN-β released in the supernatant monitored by ELISA increased 2-fold following SeV infection (Figure 4D). Finally, in VSV infected A549 cells, VSV proteins (nucleocapsid (N), surface glycoprotein (G), and matrix (M)) were detected at 8h p.i., whereas in Triad3A knock-down cells, VSV protein expression was delayed, with viral proteins detected only at 16h post-infection (Figure 4E). Notably, in A549 control cells TRAF3 protein levels decreased over time following virus infection, whereas in Triad3A knock-down cells TRAF3 protein levels remained constant (Figure 4E). These results indicate the involvement of Triad3A in regulating IFN and NF-κB dependent gene expression following RNA virus infection.
The functional specificity of TRAFs is dictated by their ability to recognize and bind distinct structural motifs, termed the TRAF-interacting motif (TIM), with the consensus sequence PxQx(T/S). This motif contacts TRAF proteins within a structurally conserved binding crevice within the C-terminal TRAF domain (Figure 5A). Using multiple sequence alignment, we identified an N-terminal motif in Triad3A - amino acid residues 316 -PMQES- 320 - with substantial homology to the consensus TIM that is also found on the adapter molecule MAVS – amino acid residues 143-PVQDT-147 (Figure 5A). Previously, it has been reported that the TIM domain of MAVS interacts with amino acid residues Y440 and Q442 within the TRAF domain of TRAF3. As a result, co-immunoprecipitation experiments were performed to detect an association of Triad3A and TRAF3; following immunoprecipitation of Flag-tagged TRAF3, immunoblot analysis revealed that TRAF3 and Triad3A co-precipitate together (Figure 5B, lane 4). Co-immunoprecipitation of TRAF3 (Y440A/Q442A) and Triad3A revealed that this interaction was impaired, demonstrating that the hydrophobic residues in the TRAF3 binding crevice are important for binding to Triad3A (Figure 5B, lane 5). In the reciprocal experiment, Triad3A S320D was unable to bind TRAF3 in co-immunoprecipitation experiments (Figure 5C, lane 6) and increasing amounts of Triad3A S320D failed to promote TRAF3 degradation (Figure S4). Furthermore, Triad3A S320D no longer inhibited ΔRIG-I-mediated activation of the NF-κB and IFNβ gene transcription but readily inhibited TRIF-mediated activation (Figure 5D,E), thus indicating the specificity of the TIM domain of Triad3A for TRAF3.
To test whether Triad3A-mediated degradation of TRAF3 was promoted by Lys48-linked ubiquitination, an in vivo ubiquitination assay was performed with Flag-tagged TRAF3, HA-tagged wild type or (Lys48 and Lys63) Ub products (Figure 6A), and sub-optimal levels of myc-tagged Triad3A and Triad3A S320D to limit TRAF3 degradation. Following immunoprecipitation of Flag-tagged TRAF3, immunoblot analysis revealed that Triad3A mediated TRAF3 polyubiquitination (Figure 6B, lane 8), with polyubiquitination increasing in the presence of Triad3A and Mg132 (Figure 6B, lane 10), compared to TRAF3 and ubiquitin alone (Figure 6B, lane 7). In contrast, Triad3A S320D did not polyubiquitinate TRAF3 (Figure 6B, lane 9); furthermore, Triad3A promoted Lys48-linked polyubiquitination of TRAF3 (Figure 6B, lane 13) but not Lys63-linked polyubiquitination (Figure 6B, lane 14). Cells expressing optimal levels of Triad3A readily degraded TRAF3 (Figure 6C, lane 2), whereas Triad3A was unable to degrade TRAF3 in the presence of K48R and KO Ub mutants (Figure 6C, lane 3,5).
As both MAVS and Triad3A contain well-characterized TIM domains, the interaction between endogenous TRAF3 and Triad3A was next examined in SeV-infected A549 cells. Following co-immunoprecipitation with anti-TRAF3 antibody, a MAVS-TRAF3 complex was detected at 8h p.i., whereas at 16h, Triad3A disrupted this interaction by associating directly with TRAF3, suggesting that both Triad3A and MAVS compete for the same binding residues on TRAF3 (Figure 7A). Importantly, a kinetic analysis of in vivo TRAF3 ubiquitination demonstrated that endogenous TRAF3 was subject to differential biphasic polyubiquitination; using Lys48 and Lys63 specific Ub antibodies [44], early Lys63-linked polyubiquitination was detected at 4h and 8h p.i. (Figure 7B), whereas a late phase Lys48-linked polyubiquitination of TRAF3 was detected at 12h and 16h p.i. (Figure 7B). Thus, TRAF3- mediated antiviral signaling appears to be regulated by recruitment of TRAF3 to the MAVS TIM, followed by Triad3A competition for the same binding crevice of TRAF3 (Figure 8).
The present study demonstrates that the E3 ubiquitin ligase Triad3A blocks RIG-I-mediated signaling to NF-κB and IRF pathways by targeting the TRAF3 adapter for degradation via Lys48-linked ubiquitinination. Several observations support this conclusion: 1) co-expression of Triad3A blocked ΔRIG-I dependent IRF-3 phosphorylation and dimerization; 2) Triad3A expression decreased TRAF3 protein levels in a dose-dependent manner; 3) knock-down of Triad3A by shRNA increased endogenous TRAF3 protein levels, increased ISG mRNA levels following virus infection, and inhibited VSV replication; 4) Lys48-linked ubiquitination of TRAF3 by Triad3A increased TRAF3 turnover; and 5) Triad3A and TRAF3 physically interacted together, an interaction that was impaired by mutation of TRAF3 (Y440A/Q442A), or reciprocally by point mutation of the TIM domain in Triad3A (S320D). TRAF3 appears to undergo a biphasic ubiquitination following virus infection that is crucial for regulation of RIG-I dependent signaling to the antiviral response. Early Lys63-linked polyubiquitination of TRAF3 leads to the recruitment of TBK1/IKKε and subsequent activation of the antiviral response [28], while late phase Lys48-linked polyubiquitination by Triad3A ultimately degrades TRAF3 and leads to shut-down of the antiviral response (Figure 8).
Recent studies have highlighted the importance of ubiquitination in modulating the innate immune response to invading pathogens via both the TLR and RLR pathways. For example, the RIG-I cytoplasmic RNA sensor undergoes both Lys48-linked and Lys63-linked ubiquitination [43],[45]: the second CARD domain undergoes TRIM25α-mediated, Lys63-linked ubiquitination at Lys-172, resulting in RIG-I/MAVS association and triggering of the antiviral response [45]; RIG-I also undergoes Lys48-linked ubiquitination, leading to RIG-I proteasomal degradation by RNF125 [43]. Additionally, RNF125 conjugates ubiquitin to MDA5 and MAVS, thus inhibiting the assembly of the downstream antiviral signaling complex [43]. Overall, multiple steps in the RLR pathway are regulated by ubiquitination to ensure a properly modulated antiviral cascade.
In addition to the newly described role of Triad3A in the regulation of the RIG-I response, previous studies demonstrated that Triad3A negatively regulates both the TLR and TNF-α pathways by promoting Lys48-linked, ubiquitin-mediated degradation of TLR4, TLR9 and TIR domain-containing adapters TRIF and TRAM [40],[41]. Triad3A regulation of the TNF-α pathway is achieved via a proteolysis-independent mechanism that impedes RIP1 binding to the TNF-R1 [40],[41]. Furthermore, Triad3A promotes ubiquitination and proteasomal degradation of RIP1 following disruption of the RIP-1-Hsp90 complex. Both Hsp90 and Triad3A form a complex that co-ordinates the homeostasis of RIP1; treatment of cells with geldanamycin to disrupt the Hsp90 complex leads to proteasomal degradation of RIP1 by Triad3A [40]. The present study further illustrates the versatility of Triad3A as a negative regulator of innate signaling pathways.
Both TLR and RLR pathways converge upon TRAF3 in the activation of the antiviral cascade. TRAF3 was originally described as a cytoplasmic adapter that interacted with CD40 and LMP1 and modulated the adaptive immune response [46],[47]. The generation of TRAF3 −/− bone marrow-derived macrophages established TRAF3 as a key molecule in signaling to the production of type I IFNs that functioned as a bridge between MAVS and the downstream kinases TBK1/IKKε [32],[39]. Triad3A mediated degradation of TRAF3 results not only in the inhibition of RIG-I signaling, but also inhibition of MDA5 and TLR3 signaling (Figure S2A, B).
The TIM sequence of MAVS (aa 143-PVQDT-147) binds to the hydrophobic C-terminal crevice of TRAF3 (TRAF domain) located between amino acids Y440 and Q442 [39]. The TIM motif represents a binding interface that recognizes different TRAFs with varying degrees of specificity. The binding cleft in TRAF3 has structurally adaptive “hot spots” that can recognize motifs that are divergent from the consensus TIM [36]. Interestingly, Triad3A interaction with TRAF3 was impaired by mutation of residues within the binding crevice (Y440A/Q442A) (Figure 6B). Furthermore, Triad3A disrupts the interaction between MAVS and TRAF3 (Figure 7A), thus highlighting the importance of the TIM domain of Triad3A in regulating TRAF3 interactions by competitive binding.
In contrast to its positive role in the production of type I IFN, TRAF3 negatively regulates noncanonical p100/p52 NF-κB activation through degradation of the NF-κB inducing kinase NIK [48],[49]. In the present study, co-expression of Triad3A decreased IFNB, IFNA4, and NF-κB promoter activity by targeting TRAF3 for degradation. Although it was expected that Triad3A driven TRAF3 degradation would enhance NF-κB promoter activity, the observed decrease in NF-κB activity suggests that Triad3A may disrupt other TRAF family members such as TRAF2 and TRAF6, prevent their association with MAVS, and thus disrupt NF-κB activation. However, it has been previously demonstrated that Triad3A does not target TRAF2 or TRAF6 for proteasomal degradation [41]. It is also possible that some components of the p100/p52 pathway may be engaged downstream of RIG-I; this idea is strengthened by the recent report that TNFR1-associated death domain protein (TRADD) is essential for RIG-I/MAVS signaling, forms a complex with TRAF3/TANK/FADD/RIP1, and leads to activation of IRF-3 and NF-κB [50]. Furthermore, the effect of Triad3A on NF-κB activation was shown to be independent of RIP1 proteolytic degradation [41], thus strengthening the possibility that another TRAF family member associates with the TIM domain of Triad3A.
Previous studies demonstrated that TRAF3 signaling was tightly regulated by the de-ubiquitinase A (DUBA) which removed Lys63 linked Ub residues from TRAF3 and disrupted recruitment of TBK1/IKKε and downstream IFN activation [28]. Dual regulation of TRAF3 by DUBA and Triad3A represents a pivotal point in the control of RLR signaling. The present results suggest a biphasic regulation or “immune-editing”, whereby TRAF3 is Lys63 polyubiquitinated early after virus infection to bridge protein-protein interactions between MAVS and TBK1/IKKε. Later, Lys63 polyubiquitin is removed by DUBA to disrupt TRAF3-TBK1/IKKε interactions [28]; TRAF3 then undergoes a late phase Lys48-linked polyubiquitination by Triad3A, leading to proteasomal degradation (Figure 8). Such a multi-level regulation of TRAF3 underscores its key role in modulating positive and negative antiviral signaling. Furthermore, the complementary functions of DUBA and Triad3A with respect to inhibition of TRAF3 activity and turnover may be subject to stimuli- and tissue-specific regulation, a topic that warrants further investigation. In conclusion, Triad3A acts as a multi-targeting E3 ubiquitin ligase that negatively regulates the TLR, TNF-α and RLR pathways; in the RLR pathway, Triad3A targets TRAF3 for Lys48-linked polyubiquitination, leading to proteasome-dependent degradation, as part of the host-specific mechanism that limits the antiviral response.
Plasmids encoding ΔRIG-I, MAVS, IKKε, TBK1, NF-κB/pGL3, IFNB/pGL3, IFNA4/pGL3, ISRE-luc reporter, and pRLTK were described previously [14],[24],[51],[52]. HA-ubiquitin and other HA-Ubiquitin constructs (HA-Ub-K48, HA-Ub-K63, HA-Ub-K48R, HA-Ub-K63R, and HA-Ub-KO) were kind gifts from Dr. Zhijian Chen (Department of Molecular Biology, University of Texas Southwestern Medical Center, Dallas Texas). MDA5 and CD4-TLR3 were kind gifts from Dr. Stephen Goodbourn (Division of Basic Medical Sciences, St George's, University of London, England) and Dr. Luke A. J. O'Neill (School of Biochemistry and Immunology, Trinity College, Dublin, Ireland) respectively. Human Triad3A cDNA was amplified from pKR5 Flag-Triad3A expression plasmid and cloned into Flag and myc pcDNA3.1/Zeo. The Triad3A point mutant S320D was introduced by Quickchange Kit according to the manufacturer's instructions (Stratagene). DNA sequencing was performed to confirm the mutation. Triad3A shRNA1 targeting nucleotide sequence (1,532–1,551) 5′-GAGCAGGAGTTCTATGAGCA-3′, shRNA2 targeting nucleotide sequence (1,195–1,214) 5′-GGACACTATGCAATCACCCG-3′ and shRNA control have been previously described [40]. Human TRAF3 cDNA was amplified from pKR5 Flag-TRAF3 and pKR5 Flag-TRAF3 Y440A/Q442A expression plasmids provided by Dr. Genhong Cheng (UCLA, USA) and were cloned into Flag pcDNA3.1/Zeo. Mg132, lactacystin and E64 were purchased from Calbiochem. dsRNA was purchased from Invivogen. A549 cells were infected with Sendai virus (40 HAU/ml) for 16h and were treated with either Mg132 (10µM), lactacystin (5µM) or E64 (5µM) 6h p.i.
Transfections for Luciferase assay were carried out in 293T cells grown in Dulbecco's modified Eagle's medium (Invitrogen) supplemented with 10% fetal bovine serum and antibiotics. Subconfluent 293T cells were transfected with 100 ng of pRLTK reporter (Renilla luciferase for internal control), 200 ng of pGL-3 reporter (firefly luciferase, experimental reporter), 200 ng of ΔRIG-I, MDA5, CD4-TLR3, MAVS, TRIF, IKKε, or TBK1 expression plasmids, 200 ng of pcDNA3 or Flag Triad3A/Flag Triad3A S320D pcDNA3, and 100ng of IRF-7 plasmid as indicated by calcium phosphate co-precipitation method. The reporter plasmids were: IFNB pGL3, ISRE-luc, NF-κB pGL3, and IFNA4 pGL-3 reporter genes; the transfection procedures were previously described [53]. At 24h after transfection, the reporter gene activities were measured by Dual-Luciferase Reporter Assay, according to manufacturer's instructions (Promega). Where indicated, cells were treated with Sendai virus (40 HAU/ml) for the indicated time or 16h for luciferase assays. Human A549 cells were cultured in F12K medium (Wisent Inc.) supplemented with 10% fetal bovine serum, glutamine and antibiotics. A549 cells were transfected either with dsRNA (20µg/ml) for 6h or infected with VSV-AV1 (multiplicity of infection of 1 (MOI)) for 16h or Sendai virus (40 HAU/ml) for 16h.
shRNA1 Triad3A and shRNA Control were transfected into A549 cells by using the Fugene 6 transfection reagent (Roche Applied Sciences). Cells were selected beginning at 48h post-transfection for 3 weeks in Dulbecco's modified Eagle's medium containing 10% heat-inactivated calf serum, glutamine, antibiotics, and 2 µg/ml G418 (Invitrogen); individual clones were screened for maximal knockdown of Triad3A by immunoblot.
293T cells were transiently transfected with 2.5 µg Flag-TRAF3, 250 ng myc-Triad3A, 250 ng myc-Triad3A S320D and 1 µg HA-Ubiquitin expression plasmids. At 6h post-transfection, cells were treated with 10 µM of Mg132 where indicated. Samples were harvested 24h post-transfection, lysed using a 1% NP-40 lysis buffer (50 mM Tris-HCL ph 7.5, 150 mM NaCl, 5mM EDTA, 50 mM NaF, 1% NP-40, 10% glycerol, 30 mMβ-glycerophosphate, 1mM orthovanadate (Na3VO4), 1 mM phenyl-methyl-sulfonyl fluoride (PMSF)) supplemented with 0.1% protease inhibitor cocktail (Sigma-Aldrich, Oakville, Ont.) and the deubiquitinase inhibitor N-ethylmaleimide (NEM, 10 mM, Sigma-Aldrich, Oakville, Ont). Samples were boiled for 10 minutes in 1% SDS and diluted 10 times in lysis buffer. 250 µg of proteins were then immunoprecipitated overnight at 4°C with constant agitation with 0.5 µg of anti-Flag (M2; Sigma-Aldrich) crosslinked to 30 µl of protein A/G PLUS-Agarose (Santa Cruz Biotechnology). Immunoprecipitated protein was washed 4 times with supplemented lysis buffer, denatured in 2% SDS-loading dye, and loaded onto a 7.5% acrylamide gel for SDS-PAGE analysis followed by transfer to nitrocellulose membrane. Polyubiquitination was detected by immunoblotting with a monoclonal anti-HA antibody (Sigma-Aldrich, Oakville, Canada). A549 cells were infected with Sendai virus (40 HAU/ml) in the presence of 5 µM of lactacystin and samples were collected every 4h p.i. Samples were lysed as previously described and samples were boiled for 10 minutes in 1% SDS and diluted 10 times in lysis buffer. 500 µg of proteins were then immunoprecipitated overnight at 4°C with constant agitation with 0.5 µg of anti-TRAF3 (sc-6933 Santa Cruz, USA) crosslinked to 30 µl of protein A/G PLUS-Agarose (Santa Cruz Biotechnology). Immunoprecipitated protein was washed 4 times with supplemented lysis buffer, denatured in 2% SDS-loading dye, and loaded onto a 7.5% acrylamide gel for SDS-PAGE analysis followed by transfer to nitrocellulose membrane. Polyubiquitination was detected by immunoblotting with polyclonal Lys48 and Lys63 anti-ubiquitin specific antibodies (Millipore, USA).
Cells were lysed in lysis buffer (50 mM Tris-HCl, pH 7.5, 250 mM NaCl, 0.5% NP-40) supplemented with 0.1% protease inhibitor cocktail (Sigma-Aldrich, Oakville, Canada). 250 µg of proteins were then immunoprecipitated overnight at 4°C with constant agitation with either 0.5 µg of anti-myc (9E10; Sigma-Aldrich) or 0.5 µg of anti-Flag (M2; Sigma-Aldrich) or 0.5 ug of anti-TRAF3 crosslinked to 30 µl of protein A/G PLUS-Agarose (Santa Cruz Biotechnology). After extensive washing with lysis buffer, the immunocomplexes were analyzed by immunoblotting as described.
Whole cell extracts (20–40 µg) were separated in 7.5–12% acrylamide gel by SDS-PAGE and were transferred to a nitrocellulose membrane (BioRad, Mississauga, Canada) at 4°C for 1h at 100 V in a buffer containing 30 mM Tris, 200 mM glycine and 20% (vol/vol) methanol. Membranes were blocked for 1h at room temperature in 5% (vol/vol) dried milk in PBS and 0.1% (vol/vol) Tween-20 and then were probed with primary antibodies. Anti-Flag (M2), anti-Hemagglutinin HA (H7), or anti-myc (9E10) each at a concentration of 1 µg/ml were purchased from Sigma-Aldrich (Sigma-Aldrich, Oakville, Canada); anti-MAVS (1∶1000, in-house previously described [14]) were prepared in blocking solution plus 0.02% sodium azide. Anti-IRF-3 (1∶5000, IBL, Japan), anti-β-Actin (1∶5000, MAB1501 Millipore, USA), anti-Triad3A (1∶1000, ProSci Inc. USA), anti-RIG-I (1∶1000, rabbit polyclonal Ab, previously described [14]), anti-VSV (1∶3000, rabbit polyclonal Ab raised against VSV proteins G, N, and M), anti-ISG56 (1∶1000, gift from Dr. Ganes Sen, Cleveland Clinic), anti-TRAF3 (1 µg/ml, sc-6933 Santa Cruz, Cal, USA), anti-IRF-3 Ser 396 (1∶1000, rabbit anti-peptide Ab, previously described [54]), and Lys48 and Lys63 anti-ubiquitin specific antibody (1∶1000, Millipore, USA) were prepared in 3% BSA/PBS/0.03% sodium azide.
Whole cell extracts were prepared in Nonidet P-40 lysis buffer (50 mM Tris, pH 7.4, 150 mM NaCl, 30 mM NaF, 5 mM EDTA, 10% glycerol, 1.0 mM Na3VO4, 40 mM β-glycerophosphate, 0.1 mM phenylmethylsulfonyl fluoride, 5 µg/ml of each leupeptin, pepstatin, and aprotinin, and 1% Nonidet P-40), and then were subjected to electrophoresis on 7.5% native acrylamide gels, which were pre-run for 30 min at 4°C. The electrophoresis buffers were composed of an upper chamber buffer (25 mM Tris, pH 8.4, 192 mM glycine, and 1% sodium deoxycholate) and a lower chamber buffer (25 mM Tris, pH 8.4, 192 mM glycine). Gels were soaked in SDS running buffer (25 mM Tris, pH 8.4, 250 mM glycine, 0.1% SDS) for 30 min at 25°C and were then electrophoretically transferred on Hybond-C nitrocellulose membranes (Amersham Biosciences) in 25 mM Tris, pH 8.4, 192 mM glycine, and 20% methanol for 1 h at 4°C. Membranes were blocked in phosphate-buffered saline containing 5% (vol/vol) nonfat dry milk and 0.05% (vol/vol) Tween 20 for 1 h at 25°C and then were blotted with an antibody against IRF-3 (1 µg/ml) in blocking solution for 1 h at 25°C. After washing the membranes five times in phosphate-buffered saline/0.05% Tween, they were incubated for 1 h with horseradish peroxidase-conjugated goat anti-rabbit IgG (1∶4000) in blocking solution. Immunoreactive bands were visualized by enhanced chemiluminescence (Amersham Biosciences).
Quantitative PCR assays were performed in triplicate using the AB 7500 Real-time PCR System (Applied Biosystems). The primers used were as follows: IFN-ß, 5′-TTGTGCTTCTCCACTACAGC-3′ (forward) and 5′-CTGTAAGTCTGTTAATGAAG-3′ (reverse); IFN-α2, 5′-CCTGATGAAGGAGGACTCCATT-3′ (forward) and 5′-AAAAAGGTGAGCTGGCATACG-3′ (reverse); ISG15, 5′-AGCTCCATGTCGGTGTCAG-3′ (forward) and 5′-GAAGGTCAGCCAGAACAGGT-3′ (reverse); ISG56 5′-CAACCAAGCAAATGTGAGGA-3′ (forward) and 5′-AGGGGAAGCAAAGAAAATGG-3′ (reverse); CXCL10 5′-TTCCTGCAAGCCAATTTTGTC-3′ (forward) and 5′-TCTTCTCACCCTTCTTTTTCATTGT-3′ (reverse); STAT1 5′-CCTGCTGCGGTTCAGTGA-3′ (forward) and 5′-TCCACCCATGTGAATGTGATG-3′ (reverse); ß-Actin, 5′-CCTTCCTGGGCATGGAGTCCT-3′ (forward) and 5′-AATCTCATCTTGTTTTCTGCG-3′ (reverse). All data are presented as a relative quantification with efficiency correction based on the relative expression of target genes versus ß-Actin as reference gene. Standard curves and PCR efficiencies were obtained using serial dilutions of pooled cDNA prepared from stable shRNA1 Triad3A and shRNA control A549 cells infected with Sendai virus (40 HAU/ml) for 12h. Data were then collected using the AB 7500 Real-time PCR System (Applied Biosystems) and analyzed by comparative CT method using the SDS version 1.3.1 Relative Quantification software.
The supernatants from stable shRNA Triad3A and shRNA Control cells infected with Sendai virus (40 HAU/ml) were collected at 12h p.i. The concentrations of IFN-β and IFN-α in the supernatants were measured using ELISA kits (PBL Biomedical Laboratories, Piscataway, NJ).
|
10.1371/journal.pgen.1003108 | The GATA Transcription Factor egl-27 Delays Aging by Promoting Stress Resistance in Caenorhabditis elegans | Stress is a fundamental aspect of aging, as accumulated damage from a lifetime of stress can limit lifespan and protective responses to stress can extend lifespan. In this study, we identify a conserved Caenorhabditis elegans GATA transcription factor, egl-27, that is involved in several stress responses and aging. We found that overexpression of egl-27 extends the lifespan of wild-type animals. Furthermore, egl-27 is required for the pro-longevity effects from impaired insulin/IGF-1 like signaling (IIS), as reduced egl-27 activity fully suppresses the longevity of worms that are mutant for the IIS receptor, daf-2. egl-27 expression is inhibited by daf-2 and activated by pro-longevity factors daf-16/FOXO and elt-3/GATA, suggesting that egl-27 acts at the intersection of IIS and GATA pathways to extend lifespan. Consistent with its role in IIS signaling, we found that egl-27 is involved in stress response pathways. egl-27 expression is induced in the presence of multiple stresses, its targets are significantly enriched for many types of stress genes, and altering levels of egl-27 itself affects survival to heat and oxidative stress. Finally, we found that egl-27 expression increases between young and old animals, suggesting that increased levels of egl-27 in aged animals may act to promote stress resistance. These results identify egl-27 as a novel factor that links stress and aging pathways.
| Stress is a fundamental aspect of aging, but it is unclear whether the molecular mechanisms underlying stress response become altered during normal aging and whether these alterations can affect the aging process. In this study, we found a GATA transcription factor called egl-27, whose targets are significantly enriched for age-dependent genes and stress response genes, and whose expression increases with age. In contrast to previous work describing factors that are causal for aging, we found that egl-27 activity is likely beneficial for survival since egl-27 overexpression extends lifespan. egl-27 promotes longevity by enhancing stress response; specifically, increased levels of egl-27 protect animals against heat stress, while reduced egl-27 activity impairs survival following heat and oxidative stress. These results suggest that aging is not simply a process of constant decline. Some factors, such as egl-27, are more active in old animals, working to restore organismal function and to improve survival.
| Responses to various forms of stress play an important role in aging and longevity. Several types of stress result in damage that can accumulate over time (e.g. oxidative stress results in damaged proteins that often accumulate with age) [1]–[3]. Responses to these stresses have protective effects that can alleviate the effects of damage accumulation. Consistent with this idea, previous studies have found that mutants with extended longevity often exhibit increased stress resistance [4]–[7]. For example, mutations that disrupt the insulin/IGF-1 like signaling (IIS) pathway not only extend longevity, but also increase resistance to many types of stress including heat, oxidative, and pathogenic stress [8]–[11].
As the first genetic pathway in Caenorhabditis elegans that was linked to longevity, the IIS pathway is a conserved endocrine component that controls important aspects of development, metabolism, and stress response [12]. Activation of the IIS receptor (DAF-2) causes phosphorylation of a phosphatidylinositol 3-kinase (AGE-1), which initiates a cascade of signals resulting in phosphorylation and inactivation of a FOXO transcription factor (DAF-16). Reduction of IIS through knockdown of daf-2 or through the presence of certain environmental stresses, results in activation of DAF-16/FOXO, which triggers a transcriptional program that promotes both stress resistance and longevity [10], [12]. Many genes in the DAF-16 transcriptional response are involved in various stress responses, and some of these also change in expression during aging. For example, heat shock proteins are induced by many types of stress including heat and pathogenic infection [13]–[17], and expression levels of certain heat shock genes are increased in C. elegans mutants with reduced IIS and extended longevity. Furthermore, heat shock proteins in C. elegans increase in expression between young and old animals, although expression is reduced in very old populations in which 90% of the population is dead [18], [19].
The increased expression of stress genes during aging is not confined to worms. Studies have shown that genes induced by oxidative stress increase with age in flies [20]; p53-related damage response genes increase with age in mice [21]; also, genes that are involved in immunological complement activation, which are generally induced in response to oxidative and pathogenic stress [22]–[26], increase expression in old age across four human tissues [27]. While these results suggest that stress response pathways become increasingly activated in old organisms, it is unclear whether this activation has a protective function and is beneficial for longevity or whether it represents a misregulation of stress pathways and is a contributor to organismal decline.
In C. elegans, the upstream regions of the genes that constitute the DAF-16 transcriptional program are enriched for both the DAF-16 binding site and a GATA-like transcription factor binding site [28]. One of the GATA factors that may be involved in the DAF-16 mediated IIS transcriptional program is ELT-3, as elt-3 expression is increased in age-1 mutants. Furthermore, elt-3 is required for the longevity phenotype of daf-2 mutants, suggesting that the elt-3/GATA transcription factor functions downstream of the IIS pathway [29].
The GATA family of transcription factors may also play important roles in regulating the molecular changes that accompany normal aging. Transcriptional profiling of young and old animals has revealed that the promoters of age-dependent genes are enriched for GATA motifs. The GATA transcription factor elt-3 is responsible for some of the age-dependent changes in gene expression. Expression of elt-3 declines as worms age, resulting in decreased expression of its downstream targets. Low levels of elt-3 have a deleterious effect on survival and stress response suggesting that this decline in elt-3 levels hastens the aging process [29].
In this work, we identify another GATA transcription factor, egl-27, that functions to promote stress survival and to delay aging. In addition to its homology to GATA factors, egl-27 is also homologous to the MTA1 component of NuRD chromatin remodeling complex [30]–[33]. Previous studies show that egl-27 is expressed in most somatic cells during development and in adult worms [30], [34]. We show that egl-27 expression increases with age and that increased levels of egl-27 through overexpression are sufficient to extend lifespan and to increase survival in response to heat stress. In contrast, reducing egl-27 activity suppresses the longevity and thermotolerance phenotypes of reduced insulin/IGF-1 like signaling. Moreover, egl-27 can respond to the presence of stress as its expression is induced by a variety of different stresses. EGL-27 binds upstream of genes involved in both stress and aging, but interestingly EGL-27 targets are enriched for genes whose expression decreases with age. Finally, egl-27 expression is regulated by the GATA transcription factor elt-3 and the IIS transducing gene daf-16. These results define egl-27 as a novel factor that promotes both longevity and stress response.
Reduction of egl-27 activity by RNAi knockdown was previously shown to partially suppress the longevity phenotype of the long-lived IIS receptor mutant daf-2(e1370) [29]. We extended this result by showing that the cold-sensitive egl-27(we3) allele could fully suppress the longevity phenotype of daf-2(e1370). Specifically, we found that daf-2(e1370); egl-27(we3) double mutants have a median lifespan that is 2.5 fold shorter than daf-2(e1370) single mutants and 35% shorter than wild-type worms (Figure 1A). egl-27(we3) mutants live only 10% shorter than wild-type worms, suggesting that the combination of the egl-27(we3) mutation and the daf-2(e1370) mutation results in a slight synthetic lethality (Figure 1A).
As a control, we compared the extent of daf-2(e1370) suppression by egl-27(we3) to that of daf-16/FOXO, a well-characterized suppressor of daf-2 [35]. We found that daf-2(e1370); daf-16(mu86) double mutants have a median lifespan that is two-fold shorter than daf-2(e1370) single mutants and 19% shorter than wild-type worms (Figure 1A). These data show that egl-27(we3) suppresses daf-2(e1370) longevity to approximately the same extent as daf-16(mu86).
egl-27(we3) is cold-sensitive for lethality [30], and so we tested whether temperature affects the suppression of daf-2 longevity by egl-27(we3). We hatched worms at the developmentally permissive temperature (20°C) and then shifted them at day 2 of adulthood to 15°C, 20°C, or 25°C. We found that egl-27(we3) suppresses daf-2(e1370) longevity at all three temperatures; specifically, daf-2(e1370); egl-27(we3) worms have a median lifespan that is 2.9 fold shorter than daf-2(e1370) mutants at 15°C, 1.9 fold shorter at 20°C, and 1.7 fold shorter at 25°C (Figure S1A). These results show that egl-27(we3) is temperature-sensitive for developmental arrest but not for suppression of longevity by daf-2(e1370).
We next tested whether increased levels of egl-27 are sufficient to increase longevity. To do this, we engineered three different constructs containing egl-27 and generated strains overexpressing each construct. The first construct is from the modENCODE project and contains GFP-tagged egl-27 in a fosmid with 18 kb of sequence upstream and 8 kb of sequence downstream of egl-27. This fosmid also contains the full coding sequence for three other genes: F31E8.6, F31E8.1, and tbc-1. We found that worms expressing egl-27::GFP had a lifespan extension of 13% (Figure 1B). The second construct contains mCherry-tagged egl-27 with full intergenic regions covering 5 kb of sequence upstream and 152 bp of sequence downstream of egl-27. Worms overexpressing egl-27::mCherry had a lifespan extension of 15% (Figure 1B). Finally, we cloned the egl-27 genomic region containing the we3 temperature-sensitive mutation from egl-27(we3) worms. This construct also contains full intergenic regions. We generated three transgenic worm strains containing egl-27(we3) on an extrachromosomal array in order to create strains that conditionally overexpress egl-27 at the permissive temperature. We grew worms at either the non-permissive or permissive temperature starting at day two of adulthood, and then measured their lifespans. Interestingly, we found that overexpression of egl-27(we3) extended lifespan at both the permissive and non-permissive temperatures. Specifically, at 20°C, median lifespan was increased 23–31%(Figure S1B) and at 15°C, the median lifespan was increased 11–21% (Figure S1C). These results suggest that the addition of low levels of egl-27(we3) activity at 15°C is sufficient to extend lifespan or that egl-27(we3) is temperature sensitive for development but not for its life-extending functions.
To determine whether egl-27 is expressed at higher levels in these overexpression lines compared to control worms, we used qRT-PCR to measure levels of egl-27 mRNA expression in the different overexpression strains and in control worms. We found that egl-27 expression is increased 2.4 fold in the egl-27::GFP strain versus control worms (p = 0.0008, Figure S5E). egl-27 expression is increased 4.1 fold in the egl-27::mCherry strain (p = 0.02, Figure S1D). Finally, egl-27 expression is increased 23%, 11%, and 2.4 fold in the three egl-27(we3) overexpression strains although none of these increases are significant, possibly due to high expression variability caused by extrachromosomal array expression (Figure S1D). We did not observe any abnormal developmental phenotypes in any of the egl-27 overexpression lines, suggesting that these levels of increased egl-27 do not adversely affect development. Furthermore, we validated that transgenic egl-27::GFP can function like endogenous egl-27(+), as we showed that egl-27::GFP can rescue the egl-27(we3) lethal mutant phenotype (Figure 1C).
Increased longevity is strongly correlated with increased stress resistance [4], [36]. To determine whether egl-27 can promote stress survival, we assessed the relationship between egl-27 activity and survival to heat and oxidative stress. To assay the phenotype of an egl-27 reduction-of-function mutation, we used the hypomorphic mutation egl-27(we3). To assay the phenotype from overexpression of egl-27, we used the egl-27::GFP strain described above.
We assayed heat-stress survival by subjecting worms to 8 hours at 35°C and then measuring survival (Figure 2A). egl-27 reduction-of-function mutants die more quickly after heat stress than wild-type worms; egl-27(we3) has a median survival time following heat-stress that is 2.9 fold shorter than for wild-type worms (log rank p-value = 7.8×10−8). egl-27 gain-of-function worms survive longer after heat stress than control worms; they have a median survival time following heat stress that is 1.4 fold longer than for control worms (p = 5.8×10−6) and they also have a time to 95% mortality that is 2.6 fold longer than for control worms.
daf-2 insulin-like receptor mutants are resistant to many types of stress [9], [11]. We showed that egl-27(we3) partially suppresses the heat resistance phenotype of daf-2. daf-2(e1370); egl-27(we3) double mutants have a median survival that is 1.7 fold shorter than daf-2(e1370) single mutants, but still 7.5 fold longer than wild-type worms. In contrast, a well-characterized suppressor of daf-2, the FOXO transcription factor daf-16(mu86) fully suppresses the heat stress resistance conferred by daf-2 mutants. daf-2(e1370); daf-16(mu86) double mutants have a median survival that is 12.9 fold shorter than daf-2(e1370) single mutants and that is the same as wild-type worms (Figure 2B). Additionally, egl-27(we3) has a less pronounced effect on the heat stress survival of daf-2(e1370) mutants than it does on wild-type worms (1.7 fold shorter vs. 2.9 fold shorter respectively) suggesting that egl-27 may be required for some but not all of the heat-stress resistance of daf-2(e1370) mutants.
We examined whether egl-27 has a functional role in mediating the response to oxidative stress. To do this, we grew worms on plates supplemented with 10 mM paraquat and then measured their survival time. We observed that altering levels of egl-27 did not have a large effect on oxidative stress survival as neither reduction nor gain of egl-27 activity affects median lifespan compared to control worms (Figure 2C). However, we found that egl-27(we3) partially suppresses the oxidative stress resistant phenotype of daf-2(e1370); specifically, daf-2(e1370); egl-27(we3) double mutants have a median lifespan that is 2.5-fold shorter than for daf-2(e1370) mutants and that is 1.4-fold longer than wild-type worms (Figure 2D). Although the median survival time for daf-2(e1370); egl-27(we3) double mutants is similar to that of daf-2(e1370); daf-16(mu86) double mutants (which have a median time of survival that is 2.3-fold shorter than for daf-2(e1370) mutants and that is 1.5-fold longer than for wild-type worms), the time to 95% mortality is different between the two lines. While daf-2(e1370); daf-16(mu86) double mutants have a time to 95% mortality that is 4.1 fold shorter than for daf-2(e1370) single mutants, daf-2(e1370); egl-27(we3) double mutants have a time to 95% mortality that is 16% shorter than for daf-2(e1370) single mutants (Figure 2D). These results suggest that egl-27 activity is important for some of the oxidative stress resistance conferred by reduced activity of the insulin-like receptor gene daf-2.
The results presented above suggest that egl-27 functions downstream of daf-2 in the IIS pathway. To test whether the IIS pathway can modulate egl-27 expression, we examined whether egl-27 expression is altered in daf-2 mutants. Using fluorescence microscopy, we compared the intestinal expression of an integrated egl-27::mCherry transcriptional reporter in a strain with reduced daf-2 activity to control worms in two day old adult, hermaphrodite worms (Figure 3A). We found that egl-27 expression increased 58% in daf-2(e1370) mutants compared to control worms, providing molecular evidence that egl-27 is regulated by daf-2 in the IIS pathway (Figure 3B).
To further define where egl-27 acts in the IIS pathway, we examined whether egl-27 acts downstream or upstream of the IIS pathway modulator daf-16/FOXO. To do this, we tested whether a mutation in daf-16 suppresses the increased levels of egl-27 found in daf-2 mutants (Figure 3A). We found that levels of egl-27::mCherry in daf-2(e1370); daf-16(mu86) double mutants are 2.1 fold lower than in daf-2 mutants (Figure 3B). This indicates that daf-16 is required for increased egl-27 expression in daf-2 mutants, which suggests that egl-27/GATA is regulated by daf-16/FOXO in the insulin signaling pathway. Further supporting this idea, we found a canonical DAF-16/FOXO binding element (GTAAACA) [28], [37] 614 bps upstream of the egl-27 translational start site, suggesting that egl-27 may be a direct target of DAF-16.
In addition to daf-16/FOXO, the GATA transcription factor elt-3 modulates IIS, as elt-3 expression is increased in age-1 mutants with reduced IIS and elt-3 is partially required for the longevity phenotype of daf-2 mutants [29]. To determine whether egl-27 acts downstream of the GATA transcription factor elt-3, we tested whether egl-27 expression is affected by a null mutation in elt-3 (Figure 3A). We found that levels of egl-27::mCherry are 3.0 fold lower in elt-3(vp1) mutants compared to control worms, suggesting that egl-27 acts downstream of elt-3 (Figure 3B). We next tested whether elt-3 is required for increased levels of egl-27 expression found in daf-2 mutants. We found that levels of egl-27::mCherry in daf-2(e1370); elt-3(vp1) double mutants are 6.7 fold lower than in daf-2 mutants and 4.8 fold lower than control worms (Figure 3B). Furthermore, levels of egl-27::mCherry in daf-2(e1370); elt-3(vp1) double mutants are similar to levels in elt-3(vp1) mutants (Figure 3B), suggesting that elt-3 is necessary for heightened egl-27 expression in the context of reduced IIS.
In the expression experiments above, we measured egl-27::mCherry expression in the anterior intestine. To determine whether egl-27 is regulated in a tissue-specific manner, we also examined egl-27 expression in the head region, which is composed of several different cell types – hypodermal, neuronal, muscle, and pharyngeal cells. We found that egl-27::mCherry levels are 35% higher in daf-2(e1370) mutants and 34% lower in elt-3(vp1) mutants compared to control worms. egl-27::mCherry levels are 37% lower in daf-2(e1370); daf-16(mu86) double mutants and 68% lower in daf-2(e1370); elt-3(vp1) double mutants compared to daf-2(e1370) worms (Figure S2A). These results suggest that IIS and elt-3 regulate egl-27 expression across multiple tissues, although the magnitude of this regulation may vary slightly across tissues.
We also determined whether egl-27 regulation by IIS and elt-3 GATA transcription occurs only during adulthood, or whether the same regulation occurs during development. To do this, we examined whether egl-27 expression is affected by mutations in daf-2 and elt-3 at the L2 larval stage of development. We found that egl-27 levels are 80% higher in daf-2(e1370) mutants and 3.1 fold lower in elt-3(vp1) mutants compared to control worms (Figure S2B). Because egl-27 is regulated by daf-2 and elt-3 to approximately the same degree during development and adulthood, the genetic networks that regulate egl-27 expression are likely developmental programs that persist into adulthood. To confirm our fluorescent microscopy results, we also examined how endogenous egl-27 levels are affected in daf-2 and elt-3 mutants. Using qRT-PCR, we showed that egl-27 levels are 67% higher in daf-2(e1370) mutants (p = 0.009) and 2.0 fold lower in elt-3(vp1) mutants (p = 0.003) compared to control worms during the L2 larval stage of development (Figure S2C, S2D).
To determine whether egl-27 forms a feedback loop with daf-16 and elt-3, we examined whether egl-27 can act upstream of these regulators. To examine whether egl-27 can regulate either DAF-16 or ELT-3 activity, we examined how reduction of egl-27 affects levels of expression of sod-3::GFP, an established transcriptional reporter for DAF-16 [38], [39] and ELT-3 [29], [39] (Figure 3C). We found that levels of a sod-3::GFP transcriptional reporter are not significantly different between egl-27(we3) mutant worms and control worms (Figure 3D). However, DAF-16 activity is low in wild-type worms, so we examined whether reduction of egl-27 affects levels of sod-3::GFP expression in daf-2 mutants where DAF-16 is highly activated. Similar to previous reports [38], [39], we found using fluorescent microscopy that sod-3 expression is 4.3 fold higher in daf-2(e1370) worms compared to control worms, and that increased sod-3 expression is suppressed in daf-2(e1370); daf-16(mu86) double mutants (Figure S2E). We found that sod-3 levels are equally high in daf-2(e1370); egl-27(we3) double mutants (Figure 3D). These results indicate that reduction of egl-27 activity does not affect expression of sod-3::GFP, suggesting that neither DAF-16 nor ELT-3 activity are affected in egl-27(we3) mutants.
Finally, we examined whether egl-27 can regulate its own expression. To do this, we examined how reduction of egl-27 activity affects the expression of an egl-27::mCherry transcriptional reporter (Figure 3E). We found that egl-27::mCherry expression is reduced by 32% in egl-27(we3) mutants compared to control worms (Figure 3F). This suggests that egl-27 activates its own expression in a feed-forward loop. Supporting this, we found several GATA-like binding motifs in the promoter region of egl-27 (TTATC/GATAA 107 bps upstream, TATCA/TGATA 728 bps upstream, and CTTATCA/TGATAAG 800 bps upstream of the translational start site). These results suggest a role for GATA transcription factors such as ELT-3 or EGL-27 itself, in directly regulating egl-27 expression.
In response to many types of stress, DAF-16/FOXO becomes activated [40], [41]. This leads to increased protection from the stress itself, and may also lead to increased levels of egl-27 expression. According to this model, various types of stresses could also lead to changes in expression of egl-27, via activation of DAF-16/FOXO. To test this possibility, we exposed worms carrying an egl-27::mCherry transcriptional reporter to six different stresses (osmotic shock, gamma radiation, starvation, heat stress, oxidative damage, and UV damage). We then compared egl-27::mCherry expression under each stress condition to its expression in controls using fluorescent microscopy (Figure 4A). We found that egl-27::mCherry expression is induced after exposure to starvation, heat stress, oxidative damage, and UV stress (p value<0.001) (Figure 4B). Interestingly, none of the stresses increase egl-27 expression by more than two-fold. This suggests that egl-27 acts differently from canonical stress-induced genes, such as heat shock proteins, which are expressed at low levels under normal conditions but can be induced up to 100-fold following heat stress [42], [43].
As a control, we showed that the increase in egl-27 expression is specific and not caused by a general increase in transcription in response to these stresses. We used fluorescence microscopy to measure expression levels of myo-3::GFP, a muscle-specific myosin gene (Figure S3A), and showed that its expression is not induced in response to starvation, heat stress, oxidative damage, or UV stress (Figure S3B).
To determine how egl-27 expression changes during aging, we used qRT-PCR to measure egl-27 expression in day 4 and day 14 adult worms. We found that egl-27 expression increases two-fold between young and old worms (Figure 4C). This is consistent with previous work showing that stress genes such as heat shock proteins increase in expression as worms age before declining in expression in extremely old worms [18], [44]. These data suggest that egl-27 expression, like the expression of other stress-related genes [18], [20], [21], [27], [45], increases in old worms.
To better understand the mechanism by which egl-27 promotes longevity and stress survival, we identified where EGL-27 binds in the genome. To do this, we prepared lysates of the egl-27::GFP worm strain described above, at the L2 larval stage of development. These lysates were used by the modENCODE consortium to generate binding site data for EGL-27, using a GFP antibody to immunoprecipate the GFP-tagged EGL-27. We chose to perform ChIP-seq using L2 larval stage worms rather than adult worms because the majority of datasets generated by modENCODE were for larval stages. Because we showed that egl-27::GFP can rescue the egl-27(we3) lethal mutant phenotype (Figure 1C), the sites that are bound by EGL-27::GFP are likely to be representative of the sites that are bound by endogenous EGL-27.
By examining ChIP-seq data from the modENCODE project, we identified 4113 DNA regions showing significant binding by EGL-27::GFP [46]. Previous work has shown that some DNA regions are bound by one or a few transcription factors (factor-specific) while other DNA sites are associated with a large number of transcription factors, and that the specific functions of each transcription factor are better defined by its factor-specific targets than by these redundantly bound targets [46]. Because we were interested in the factor-specific functions of EGL-27, we removed 2306 sites located within redundantly bound regions from further analysis. Of the factor-specific binding sites, 481 are located within gene promoters, defined as 1000 bps upstream to 500 bps downstream of a translational start site. 426 binding sites are located within exons and 466 binding sites are located within introns. 78 are located within 1000 bp downstream from the translational stop site. Finally, 516 are located in intergenic regions. Because we were interested in putative targets of EGL-27, we focused on the 481 peaks located within gene promoters.
To identify the consensus sites that may be directly bound by EGL-27, we examined the 481 ChIP-seq peaks that fall within gene promoters for the presence of enriched DNA motifs. We used the Gibbs sampling program BioProspector [47] to perform a de novo motif search on the center 100 bp sequence of EGL-27 ChIP-seq identified binding sites. The top 10 motifs found by the program are variations of two motifs: the GATA motif and a novel RGRMGRWG motif (Table S2). The GATA motif (GAKAAG) is found in 32% of EGL-27 target peaks and the novel RGRMGRWG motif is found in 25% of target peaks. Both are significant when compared to background sets consisting of randomly generated 100 bp sequences centered from 1000 bp upstream to 500 bp downstream of translational start sites (Figure 5). Both motifs are also significant when compared to scrambled sequence derived from EGL-27 peaks that preserve nucleotide frequencies (Figure 5).
The novel RGRMGRWG motif is not a consensus-binding site for any known class of transcription factors. However, this motif was previously identified as enriched in the promoters of differentially-expressed genes in insulin signaling daf-2 mutants as well as sirtuin pathway mutants [48].
Since EGL-27 contains a GATA DNA-binding domain, the enrichment for GATA motifs in the EGL-27 binding sites supports its function as a GATA transcription factor. Previous studies have shown that GATA motifs are important for osmotic and pathogenic stress response [49]–[51] and aging [28], [29], which is consistent with our model that egl-27 regulates stress and aging genes by binding the GATA motif.
The 481 EGL-27 promoter peaks are located in the promoter regions of 501 unique genes (Table S3). We conducted gene ontology (GO) analysis [52], [53] on the set of 501 EGL-27 target genes in order to determine if the EGL-27 target genes are enriched for specific biological pathways. We found that EGL-27 targets are enriched for hypodermal genes that form the cuticle as well as intestinal genes involved in aromatic compound catabolic process (Table S4).
We examined whether target genes bound by EGL-27 significantly overlap genes that change expression with age. We obtained a list of 1132 age-regulated genes as previously defined by DNA microarrays [29]. Specifically, we computed the hypergeometric p-value for the overlap between the set of 501 EGL-27 target genes and the set of 1132 age-regulated genes. We found that 67 EGL-27 targets show age-dependent expression, which is a 2.8 fold enrichment over the number expected by chance (p = 9.5×10−15) (Table 1). Even though the ChIP-seq analysis was performed at the L2 larval stage of development, we were still able to find a strong enrichment for age-regulated genes. Interestingly, 61 of the 67 targets decline in expression with age (Figure S4) suggesting that increased levels of egl-27 during normal aging are insufficient to prevent the age-dependent decline of these genes. Using GO analysis, we found that these 67 age-dependent EGL-27 targets are even further enriched for hypodermal genes that form the cuticle as well as several metabolic categories including aromatic amino acid family metabolic process, oxoacid metabolic process, and organic acid metabolic process (Table S4).
Because we previously showed that egl-27 expression is induced in response to several types of stress, we examined whether egl-27 targets are differentially-expressed in response to different types of stress. To do this, we acquired published transcriptional responses to 16 different stress conditions involving response to six pathogens, six environmental stresses, and four toxins (Table S5). We then compared each set of genes that are differentially-expressed in that particular stress condition to the set of EGL-27 targets to determine whether the two sets significantly overlap by a hypergeometric test. Even though the ChIP-seq experiment for EGL-27 was not performed under stress conditions, we found that EGL-27 target genes are significantly (p<10−5) enriched for differentially-expressed genes from 11 of the 16 different stress conditions (Table 1). This supports the idea that egl-27 is involved in the response to many types of stress.
To better characterize how egl-27 mediates heat-stress survival, we assessed how egl-27 regulates heat-stress target genes. To do this, we examined the expression of four egl-27 target genes (grd-3, T14B1.1, Y37A1B.5, and lpr-3) that were previously shown by DNA microarray experiments to be differentially expressed following heat stress [54]. We first wanted to see that expression of these genes change as expected after heat stress in wild-type worms. To do this, we extracted RNA from wild-type second larval stage worms that were exposed to 90 minutes of heat at 35°C followed by a 2-hour recovery. We then used qRT-PCR to determine how the expression of each gene changes in heat stressed worms compared to unstressed worms. Similar to the DNA microarray data, we found that grd-3, T14B1.1, and Y37A1B.5 expression increases while lpr-3 expression decreases after heat stress (Figure 6A–6D).
We examined how reduced egl-27 activity affects target gene expression before heat stress. To do this, we compared expression levels of EGL-27 target genes in egl-27(we3) mutants to their expression levels in wild-type worms. We found that two genes (grd-3 and T14B1.1) had higher expression while two genes (Y37A1B.5 and lpr-3) had lower expression in egl-27 loss-of-function worms than in wild-type worms (Figure 6A–6D). This shows that egl-27 plays a role in modulating target gene expression under normal conditions when worms are unstressed.
Next, we assessed how reduction of egl-27 activity affects the heat stress response of these four genes, where the heat stress response is defined as the ratio of expression after heat stress compared to before stress. We found that reduction of egl-27 activity resulted in attenuation of the heat-induced changes of all four genes (Figure 6E). Interestingly, the reduced heat stress response is not caused by lower levels of induced expression, but rather by higher levels of basal expression. Expression of all four target genes was significantly altered in egl-27 reduction-of-function mutants prior to heat stress (Figure 6A–6D). For example, in egl-27 reduction-of-function worms, the expression of two targets (grd-3 and T14B1.1) is induced even in pre-stressed animals. Although expression of these genes remains unchanged following stress, both pre-stress and post-stress expression levels in egl-27 reduction-of-function worms are similar to their levels in post-stress wild-type worms. In the case of these two genes, egl-27 may function to alter baseline expression levels in unstressed worms. However, expression levels of two other targets (Y37A1B.5 and lpr-3) are significantly different in egl-27 reduction-of-function worms compared to wild-type worms both pre- and post- heat stress. For these two genes, egl-27 may function to alter baseline expression levels in unstressed worms and is required for additional activation of expression following heat stress. These results suggest that endogenous egl-27 is required to reduce basal stress levels in wild-type worms.
To complement the reduction-of-function results, we examined how increased basal egl-27 activity in the egl-27::GFP overexpression strain affects differential expression of egl-27 in response to heat-stress (Figure S5). First, we used qRT-PCR to measure the combined RNA levels of egl-27::GFP and endogenous egl-27. As expected, the egl-27::GFP overexpression strain had both increased levels of combined egl-27 expression before and after stress when compared to control worms (Figure S5E). Additionally, we observed that the heat stress response of egl-27 is greater in egl-27::GFP worms compared to control worms (2.9 fold versus 2.2 fold respectively)(Figure S5E). This result suggests that worms with egl-27::GFP have increased levels of egl-27 activity, which may make the expression of egl-27 more sensitive to heat stress as part of a positive feedback loop.
However, the egl-27::GFP strain showed variable effects on the expression of EGL-27 target genes both before and after heat stress (Figure S5A–S5D). Expression of grd-3 was unaffected by increased levels of egl-27 before heat stress. Following heat stress, expression of grd-3 was higher in the egl-27::GFP strain than in the control strain, such that induction levels were 1.8 fold in the egl-27::GFP strain compared to 1.2 fold in the control strain. Expression of T14B1.1 was unaffected by increased levels of egl-27 both before and after heat stress. Expression of Y37A1B.5 was increased in egl-27 gain-of-function worms before heat stress but repressed in these worms following heat stress. Finally, expression of lpr-3 was 19-fold lower in egl-27 gain-of-function worms before heat-stress compared to control worms and 3.6-fold lower following heat stress. Whereas heat stress causes this gene to decrease expression slightly in control worms, heat stress causes its expression to increase in the egl-27::GFP strain. While expression of lpr-3 was significantly down-regulated following heat stress in wild-type worms, lpr-3 levels were not significantly changed in transgenic control worms, underscoring the background differences that can exist between wild-type and transgenic animals. These data suggest that increased levels of egl-27 affect the expression of some target genes both prior to and after heat stress.
We have shown that a GATA/MTA1 transcription factor [29]–[33], [55], egl-27, is an important mediator of stress response and longevity in C. elegans. Using binding site data from chromatin immunoprecipitation followed by ultra high throughput sequencing (ChIP-seq), we have shown that EGL-27 binds upstream of genes that are enriched for those that increase with age and that change in response to diverse stresses.
We examined whether EGL-27 has a beneficial or detrimental effect on longevity and stress response. We found that overexpression of egl-27 increases both longevity as well as survival in response to heat stress. In contrast, egl-27 mutants have a shortened lifespan and reduced survival in response to heat stress. Furthermore, reduction of egl-27 activity suppresses the longevity phenotype as well as the heat and oxidative stress resistance phenotypes of daf-2 mutants. These results suggest that egl-27 may promote longevity through promoting stress resistance. This possibility is supported by other studies showing that increased expression of genes that confer resistance to specific stresses also extend lifespan. For example, increased levels of zebrafish lysosyme which confers antimicrobial defense [56], heat shock factor 1 which confers heat resistance [5], [56], and skn-1 which confers resistance to oxidative damage [6], [57], are all capable of extending lifespan when overexpressed in transgenic worms. Here, we identify a GATA transcription factor, egl-27, that plays a role in the response to multiple stresses and has a beneficial effect on longevity when overexpressed.
Altering levels of egl-27 activity affects the expression of heat stress genes both in unstressed as well as heat-stressed worms. From these experiments, we infer that egl-27 normally maintains a program to resolve cellular stress, and that altering levels of egl-27 alters baseline stress levels in worms. Similar to this, a previous study showed that certain stress response genes are expressed at lower levels in stress-resistant daf-2 mutants [58]. This suggests that that changes that alter baseline levels of stress can also alter baseline expression of stress response genes.
egl-27 expression is regulated by a variety of stress pathways. We found that egl-27 expression is induced by multiple stresses: heat stress, oxidative damage, UV irradiation, and starvation. Next, we showed that egl-27 acts downstream of the GATA transcription factor elt-3 and two IIS pathway components (IIS receptor daf-2 and the FOXO transcription factor daf-16). egl-27 expression is induced in long-lived daf-2(e1370) mutants and this induction is suppressed in daf-2(e1370); daf-16(mu86) and daf-2(e1370); elt-3(vp1) double mutants. These data support our finding that egl-27(we3) can fully suppress daf-2(e1370) longevity. Furthermore, previous work has shown that elt-3 is a pro-longevity factor whose expression is confined mainly to hypodermal cells [29], [59], [60]. Our finding that elt-3 can regulate egl-27 expression in several tissue types including the intestine suggests that elt-3 can affect gene expression in a cell non-autonomous manner.
Surprisingly, expression of sod-3, which acts as a readout for DAF-16 activity [38], [39], is unchanged in daf-2(e1370); egl-27(we3) double mutants compared to daf-2(e1370) single mutants, suggesting that activated DAF-16 is not sufficient for extended longevity in the absence of functional egl-27. Finally, we found that egl-27 can regulate its own expression in a feed-forward loop. This evidence for auto-regulation supports the idea that egl-27 may be involved in a complex circuit with feedback mechanism for regulating target gene expression.
Interestingly, egl-27 expression increases with age in wild-type worms. Our finding that increased egl-27 expression extends lifespan and improves stress resistance suggests that the way that egl-27 expression changes during aging is beneficial for the organism. In contrast, previous studies have focused on age-dependent changes in expression that are detrimental for the organism. For example, the GATA transcription factor elt-3 is an important regulator of the transcriptional changes that occur between young and old. elt-3 declines in expression with age and low levels of elt-3 have a deleterious effect on survival and stress response, suggesting that declining levels of elt-3 may act as a driver of aging [29]. Another study found that NF-κB acts as a key regulator of age-dependent gene expression differences in nine types of human and mouse tissues [61]. NF-κB expression increases in old animals, and this increase is detrimental as blocking NK-κB in old skin results in a partial reversal of the aging transcriptome and more youthful skin [61], [62].
In contrast to elt-3 in C. elegans and NF-κB in mice, our studies suggest that changes in egl-27 expression during aging may act to improve stress response and to promote longevity. However, most EGL-27 binding targets from ChIP-seq experiments decline in expression with age, suggesting that this increased expression is insufficient to prevent the age-dependent decline of these genes. Because increased levels of egl-27 extend lifespan, increased expression of egl-27 in old worms appears to delay the aging process instead of causing it. Our work offers novel insight into the interplay between stress and aging, and suggests that aging is not simply a process of moving from an ideal young transcriptome to an inadequate old transcriptome. Rather, age-dependent changes in gene expression are likely comprised of a mix of beneficial, detrimental, and neutral changes.
All C. elegans strains (Table S6) were handled and maintained as described previously [63]. Genotyping primers are described in Table S7.
Lifespan experiments were conducted as previously described [35], [64]. All experiments were done at 20°C unless otherwise noted. Age refers to days following adulthood and p-values were calculated using the log-rank (Mantel-Cox) method [65].
daf-2(e1370) and egl-27(we3) are single base pair change mutants, so we used the tetra-primer ARMS-PCR procedure [66] to design 4 primers for each SNP of interest. daf-16(mu86) and elt-3(vp1) are deletions so we used a combination of 3 primers (two flanking the deleted region and one inside of the deleted region) to probe for the deletion. To generate DNA for genotyping, 1–10 worms were lysed in 1× PCR buffer with 1.5 mM MgCl2 and Protease K. A single PCR reaction was setup using this DNA and all (three or four) primers and then visualized on a 2% agarose gel. For egl-27(we3), all reactions will produce a 396 bp product; mutant allele will produce a 213 bp product while wild-type allele will produce a 240 bp product. For daf-2(e1370), all reactions will produce a 300 bp product; mutant allele will produce 155 bp product while wild-type allele will produce a 203 bp product. For daf-16(mu86), mutant allele will produce a 400 bp product while wild-type allele will produce a 634 bp product. For elt-3(vp1), mutant allele will produce a 145 bp product while wild-type allele will produce a 401 bp product. Primers and temperatures are detailed in Table S7.
Worms carrying the integrated transgene egl-27::GFP (OP177) were crossed to egl-27(we3) (JA1194) to generate egl-27::GFP; egl-27(we3) (SD1751) in the F2 generation, which was identified by PCR genotyping (genotyping primers for egl-27(we3) detailed in Table S7; methods for single nucleotide genotyping above). N2, egl-27(we3), and egl-27::GFP; egl-27(we3) worms were synchronized using hypochlorite and grown to day 1 of adulthood. 10 hermaphrodites from each strain were individually placed onto NGM plates seeded with E. coli and allowed to lay eggs for 1 hour. The adult worms were removed and the number of eggs counted. The numbers of hatched worms were scored 1 day, 2 days (not shown), and 6 days (Figure 1C) after the beginning of egg laying. % survival was computed as the percentage of hatched worms divided by the number of total eggs. Error bars represent the standard error between the 10 replicates.
To quantify mCherry and GFP expression, we imaged at least 15 worms for each condition at 20× using a Zeiss Axioplan microscope. Images were analyzed using ImageJ [67].
All assays were performed using day 1 adult hermaphrodites. In all cases, E. coli refers to the OP50 strain. Control worms were always transferred the same number of times, and imaged at the same time as experimental worms. All strains used for imaging and survival assays are detailed in Table S6.
Two independent cultures of worms expressing egl-27::GFP (OP177) were synchronized using sodium hypochlorite to isolate eggs followed by hatching in S basal overnight. Arrested L1 stage larva were collected and grown on NGM plates seeded with E. coli until mid L2 stage. ChIP-seq was performed by the modENCODE consortium [46], [69], [70]. The program PeakSeq [71] was used to identify EGL-27::GFP binding sites (q value<0.00001). Peaks bound by five or more out of the original 23 transcription factors were removed from further analysis. Significant, factor-specific peaks were then compared to the C. elegans genome to identify putative EGL-27 target genes. A gene is identified as a target if the center of a peak occurs 1.5 kb upstream or 500 bp downstream of its translational start site.
The center 100 bp sequences from the 481 promoter peaks were filtered for low complexity regions using RepeatMasker [72] and then submitted to BioProspector to identify overrepresented cis regulatory sequences [47]. To eliminate motifs or amino acid distributions that are generally enriched in all promoters, a random set containing 481 masked 100 bp promoter sequences was submitted to BioProspector as background input. We report the highest 10 motifs from BioProspector (Table S2), but since only 2 are unique, we ran further analysis on these two.
To determine the fold enrichment and probability of these two motifs occurring in our dataset compared to the probability expected by chance, 1000 datasets containing 481 randomly-generated 100 bp promoter sequences (to match the number of promoter peaks) were created. We scored the number of times these two motifs were found in our promoter sequences and the average number of times they were found in each of the 1000 background datasets. An enrichment factor for each motif was calculated using (# of ChIP-seq peaks containing at least one instance of the motif)/(average # of background peaks containing at least one instance of the motif). A chi-squared p-value for each motif was calculated using the # of ChIP-seq peaks containing at least one instance of the motif as observed and # of background peaks containing at least one instance of the motif as expected (Figure 5).
As a second control, we generated a set containing the 481 original sequences but scrambled the order of the nucleotides in order to create a random sequence while preserving the nucleotide frequency. We generated 1000 iterations of this set of 481 scrambled sequences and for each motif, scored the average number of times they were found in this scrambled sequences. Again, we calculated an enrichment factor and a p-value for each motif in comparison to the scrambled sequences (Figure 5).
EGL-27 target genes and age-regulated EGL-27 target genes were submitted to GOrilla [52], [53] for Gene Ontology (GO) analysis. A background gene list was also submitted. For EGL-27 target genes, the background list consisted of all annotated C. elegans genes. For age-regulated EGL-27 target genes, the background list consisted of all genes represented on the Stanford C. elegans microarray platform. GOrilla outputs a FDR q-value and an enrichment score for each GO category. The FDR q-value is the hypergeometric p-value corrected for multiple hypotheses testing using the Benjamini and Hochberg method [73]. Enrichment score is computed as (# of inputted genes in GO category/# of total inputted genes)/(# of genes in each GO category/# of background genes) [52], [53].
Age-regulated genes were obtained from [29] and probe IDs were remapped to WormBase gene IDs (WS228) using annotations from Stanford Microarray Database. Probes that could not be mapped to gene IDs were removed from further analysis. Differentially-expressed stress response genes for each of 16 stress conditions were obtained from publications detailed in Table S5. When probe names were supplied, probe IDs were remapped to WormBase gene IDs (WS228) using annotations from Affymetrix, Stanford Microarray Database, and Washington University Genome Sequencing Center. Probes that match multiple genes were assigned all associated gene IDs. Probes that could not be mapped to gene IDs were removed from further analysis. If only gene names were supplied, gene IDs were filtered using WormBase gene annotations (WS228). Unmatched probe IDs or gene IDs were removed from further analysis.
To determine whether EGL-27 target genes were significantly enriched for age-regulated or stress-response gene sets, we first determined the number of EGL-27 targets that are significantly differentially-expressed in each age or stress response set. For each comparison, the background gene count is the number of genes included in the platform for that microarray. Using this number, we computed, both a hypergeometric p-value and an enrichment score for each comparison, where enrichment score is defined as (# of EGL-27 targets/# of differentially expressed genes)/(# EGL-27 targets/# of background genes).
N2, egl-27(we3), control, and egl-27::GFP (OP177) worms (full genotypes and strain information in Table S6) were synchronized using hypochlorite to isolate eggs followed by hatching in S basal overnight. Arrested L1 stage larva were collected and grown on NGM plates seeded with E. coli until mid-L2 stage before splitting into an experimental and control set. Experimental worms were then incubated at 35°C for 90 minutes and allowed to recover at 20°C for 2 hours before collecting in Trizol (Invitrogen). Control worms were kept at 20°C and collected in Trizol at the same time. Total RNA was isolated using Trizol reagent, treated with DNaseI to degrade genomic DNA, and purified using RNeasy kit (Qiagen). cDNA was synthesized using oligo dT primers and SuperScript II First Strand Kit (Invitrogen). qPCR reactions were performed using RT2SYBR Green qPCR Mastermix (Qiagen) and the 7900HT Fast Real-Time PCR System (ABI). Melting curve analysis was performed with each primer pair to ensure that quantification is the result of only one product. A serial dilution was performed for each primer pair to generate a standard curve. act-1 was used as an internal control to normalize expression levels as previously described [74], [75]. All primers are detailed in Table S7.
|
10.1371/journal.pgen.1006121 | Network Analysis of Genome-Wide Selective Constraint Reveals a Gene Network Active in Early Fetal Brain Intolerant of Mutation | Using robust, integrated analysis of multiple genomic datasets, we show that genes depleted for non-synonymous de novo mutations form a subnetwork of 72 members under strong selective constraint. We further show this subnetwork is preferentially expressed in the early development of the human hippocampus and is enriched for genes mutated in neurological Mendelian disorders. We thus conclude that carefully orchestrated developmental processes are under strong constraint in early brain development, and perturbations caused by mutation have adverse outcomes subject to strong purifying selection. Our findings demonstrate that selective forces can act on groups of genes involved in the same process, supporting the notion that purifying selection can act coordinately on multiple genes. Our approach provides a statistically robust, interpretable way to identify the tissues and developmental times where groups of disease genes are active.
| Some genes are extremely intolerant of mutations that alter their amino acid sequence. Such mutations are highly likely to drive disease, and previous reports have implicated these genes in multiple diseases. To better understand the function of these constrained genes and their place in cellular organization, we developed a framework to ask if these genes form biochemical networks expressed in specific tissues and developmental timepoints. Using clustering analysis over protein-protein interaction maps, we show that 72/107 such genes form a densely connected network. Using another new method, we found that these 72 genes are coordinately expressed in fetal brain and early blood cell precursors, but not other tissues, in the Roadmap Epigenomic Project, and then show that this gene module is active in very early developmental time points of the hippocampus included in the Brainspan Atlas. We also show that these genes, when mutated, tend to cause genetic diseases. Thus we demonstrate that evolution constrains mutation of key mechanisms that must therefore require careful control in both time and space for development to occur normally.
| Genetic variation is introduced into the human genome by spontaneously arising de novo mutations in the germline. The majority of these mutations have, at most, modest effects on phenotype; they are thus subject to nearly neutral drift and can be transmitted through the population, with some increasing in frequency to become common variants. Conversely, de novo mutations with large effects on phenotype may be subject to many different selective forces, both positive and negative, with the latter resulting in either the variant being completely lost from the population or maintained at very low frequencies [1].
Large-scale DNA sequencing can now be used to comprehensively assess de novo mutations, with many current applications focusing on the protein-coding portion of the genome (the exome). This approach has been used to identify causal genes and variants in rare Mendelian diseases: for example, exome sequencing of ten affected individuals with Kabuki syndrome identified the methyl transferase KMT2D (formerly MLL2) as causal, after substantial post hoc data filtering [2]. In complex traits, this approach has successfully identified pathogenic genes harboring de novo mutations in autism spectrum disorders [3], intellectual disability [4] and two epileptic encephalopathies [5]; notably, all these studies sequenced the exomes of parent-affected offspring trios and quantified the background rate of de novo mutations in each gene using formal analytical approaches. They were thus able to identify genes harboring a statistically significant number of mutations, which are likely to be causal for disease [5,6].
These large-scale exome sequencing studies have demonstrated that the rate of non-synonymous de novo mutations is markedly depleted in some genes, and that these genes are more likely to harbor disease-causing mutations [6]. As synonymous de novo mutations occur at expected frequencies, this depletion is not driven by variation in the local overall mutation rate; instead, these genes appear to be intolerant of changes to amino acid sequence and are thus under selective constraint, with non-synonymous mutations removed by purifying selection. These genes represent a limited number of fundamental biological roles, which suggests that entire processes, rather than single genes, are under selective constraint. This is consistent with the extreme polygenicity of most human traits, where hundreds of genes play a causal role in determining organismal phenotype [7,8]. These genes must participate in the same cellular processes, but uncovering the relevant connections and the cell populations and developmental stages in which they occur remains a challenge. We and others have described statistical frameworks to test connectivity within a nominated set of genes [9–11] by considering how genes interact either in annotated pathways or in networks derived from protein interactions or gene co-expression across tissues, and these approaches have been successfully applied to detecting networks of genes underlying neurodevelopmental disease [12]. These studies have demonstrated that genes underlying complex diseases tend to aggregate in networks; we hypothesize that the same is true of constrained genes. However, unlike disease traits where the relevant organ system is known and hypotheses about pathogenesis can by formulated, the phenotypic targets of selective forces are usually unknown. Thus, systematic genome-wide approaches to assessing connectivity between a set of genes of interest and to identify relevant tissues are required to investigate how selective constraint acts on groups of genes and uncover the relevant physiology.
To address these issues we have developed a robust, unbiased framework and applied it to genome-wide selective constraint data derived from exome sequences of 6,503 individuals [6]. We identified a single, statistically significant subnetwork of 72 interacting genes highly intolerant of non-synonymous variation, with no other interacting groups of genes showing evidence of such coordinate constraint. To establish biological context for this subnetwork, we developed a robust approach to test for preferential expression of the module as a whole, rather than the individual constituent genes. Using gene expression data from the cosmopolitan atlas of tissues in the Roadmap Epigenome Project [13,14], we found that this subnetwork is preferentially expressed in several early-stage tissues, with the strongest enrichment in fetal brain. To more carefully dissect the role of this subnetwork in the central nervous system, we analyzed expression data from BrainSpan [15], an atlas of the developing human brain, and found that the constrained gene subnetwork is preferentially expressed in the early development of the hippocampus. Consistent with this observation, this module is enriched for genes mutated in neurological, but not other, Mendelian disorders. We thus show that selective constraint acts on a set of interacting genes active in early brain development, and that these genes are in fact intolerant of mutation. Our Protein Interaction Network Tissue Search (PINTS) framework is publicly available at https://github.com/cotsapaslab/PINTSv1.
We have previously described a framework to assess selective constraint across coding sequences in the genome [6]. Briefly, we calibrated an expectation for all possible conversions of one base to another by mutation from non-coding sequence. For each transition, we modeled the effect of the surrounding sequence and its conservation across species to correct for context effects. We then counted the number of synonymous and non-synonymous variants in the coding sequence of each gene in the genome and derived a statistic of constraint on each class of variation compared to this global expectation. We found that a number of genes show decreased rates of non-synonymous substitution but expected rates of synonymous substitution, consistent with purifying selection removing the non-synonymous alleles from the population.
If constrained genes lie in biologically meaningful networks, we expect them to (i) interact and (ii) be expressed in the same tissues. We developed a robust, modular workflow (PINTS–Protein Interaction Network Tissue Search; Fig 1) to test both of these hypotheses at a genome-wide level. To detect interactions between constrained genes we used a high-confidence protein-protein interaction network (InWeb [16]), and employed a clustering algorithm previously validated on such networks [17]. We assessed significance empirically by randomly reassigning constraint scores to genes (see Materials and Methods and S1 Text). We then tested any significant subnetworks for preferential expression in the diverse tissue atlas provided by the Roadmap Epigenome Project (REP), which assays gene expression in 27 human primary samples across the developmental spectrum [14]. Our final dataset is comprised of 9729 genes both present in InWeb and detected in at least one REP tissue.
Our workflow is both modular and flexible: clustering algorithms, gene-gene relationships and tissue atlases can be replaced as required, so that analyses can be tailored to suit specific biological problems. A flexible implementation, including all data described here, is freely available as an R package at https://github.com/cotsapaslab/PINTSv1.
We define highly constrained genes as those with evidence of constraint on non-synonymous de novo substitutions (p < 5 x 10−6, Bonferroni correction for the number of genes in our InWeb dataset) but null synonymous constraint scores, indicating intolerance to functionally relevant mutation rather than fluctuations in the local mutation rate [6]. Of these, 107/9729 genes pass this stringent threshold (binomial p < 2.2 x 10−16; S1 Table), and form the core of the analysis presented here. We found that 67/107 form a connected subnetwork (Fig 2A; Table 1). Five additional genes are included as our cluster detection algorithm by design looks for a backbone of null nodes connected to many signal nodes. To assess the significance of this observation, we randomly distribute constraint scores to InWeb nodes 1000 times and find that the constrained subnetwork is larger (number of nodes: p < 0.001) and more densely connected (number of edges: p < 0.001; clustering coefficient: p = 0.008) than expected by chance (Fig 2B). As such, it also explains more total constraint in the genome than expected (sum of constraint scores: p < 0.001). After accounting for the genes forming this subnetwork, we found no evidence that the remaining 35 genes form statistically significant subnetworks by our criteria.
The genes in the constrained subnetwork appear to represent several fundamental cell processes, most notably mitosis and cell proliferation (SMC1A, SMC3, CTNNB1) and transcriptional regulation (CHD3, CHD4, SMARCA4). We performed a formal pathway analysis to further test this and found enrichment of several annotated pathways reflecting these fundamental processes (Table 2). Encouraged that our detected subnetwork represents one or more biological processes under constraint, we sought to add cellular context to our observations. In particular, we wanted to determine if this group of genes is preferentially expressed in particular tissues, indicating a likely site of action. We thus developed an approach to estimate the joint probability of preferential expression of the genes in the subnetwork in each tissue of an atlas of expression data, while accounting for how frequently each gene is detected across the entire atlas. We applied our approach, which uses Markov random fields, to the expression data on 27 primary tissues and cell lines available from the Roadmap Epigenome Project. Using two conservative permutation-based significance tests, we find the constrained subnetwork is preferentially expressed in a number of fetal and immune tissues (Fig 2C and Table 3), including fetal brain (permuted p < 0.001), the immune cell subpopulations marked by CD34 (permuted p < 0.001) and CD8 (permuted p = 0.017) and fetal thymus (permuted p = 0.048). We note that, whilst only a subset of genes are expressed in any one tissue, the combinations of genes expressed in these tissues is highly statistically significant: each gene is only expressed in a small subset of the tissues interrogated, so the cumulative probability of seeing these genes coordinately expressed in any one tissue is small.
As several tissues are enriched for subnetwork expression, we sought to understand whether we were capturing the same signature across multiple tissues reflecting a shared process. We assessed whether the same genes are preferentially expressed in each tissue, and found a distinct signature in the fetal brain and heart samples and the immune cell subpopulations (CD34+, CD8+, CD3+, thymus; pairwise p < 0.05 hypergeometric test; S2 Table). To ensure our tissue expression results are not an artifact of the threshold we set for preferential expression, we repeated the entire analysis with a range of threshold values and found consistent results across tissues; this is most notable in fetal brain (Fig 2D and S3 Table), which remains significant irrespective of threshold used.
Genes under selective constraint are more likely to harbor pathogenic mutations causing Mendelian diseases, consistent with intolerance of functional mutations [6]. Accordingly, we found that our subnetwork of 72 genes is significantly enriched for OMIM annotations (Fisher’s exact p = 0.0013). To further elucidate this observation, we mapped all OMIM entries to Medical Subject Headings (MeSH) disease categories and assessed enrichment per organ system category. We found that our subnetwork is significantly enriched for genes mutated in Mendelian diseases affecting the central nervous system (Fisher’s exact p = 0.0017, S5 Table), validating our observation of enrichment in fetal brain. We note that this enrichment is not in the inflammatory/immune neurological disease sub-category, suggesting no overlap with the discrete immune signature we found. Samocha et al [6] have previously reported that constrained genes are also enriched for de novo mutations associated with autism spectrum disorders, further strengthening our conclusion that this constrained subnetwork represents a brain-related biological process.
To further elucidate the relevance of our constrained module to brain physiology, we interrogated expression data for multiple brain structures across developmental stages from the BrainSpan project [15]. We found a strong signature of preferential expression in very early stages of development, which declines rapidly and is absent by mid-gestation and remains inactive after birth into adulthood (Fig 3A and Table 3). Several transitional structures in the early brain exhibit significant preferential expression levels, including the ganglionic eminences that eventually form the ventral forebrain and the early structures of the hippocampus. The latter structure shows the most consistent signature across developmental time, with the module’s pattern of expression gradually weakening and becoming non-significant by mid gestation (post-conception weeks 16–18; Fig 3B). These results, taken with the likely involvement of constrained genes in fundamental processes of mitosis and transcriptional regulation, suggest this gene module is relevant to developmental patterning at crucial time points in early brain development.
We have shown that selective constraint influences sets of interacting genes involved in core cellular control processes, and that these have elevated expression levels in early stages of central nervous system development. We found the strongest enrichment in the early hippocampal stages at post-conception weeks eight and nine, with additional signals in ventral forebrain structures and the parietal cortical wall. This stage of development involves neuronal proliferation through carefully orchestrated sequences of cell differentiation during developmental patterning across the brain. As the constrained subnetwork we have detected is enriched for genes involved in the control of mitosis and transcription, we speculate that it plays a fundamental role in these processes. Our finding that neurological Mendelian disease genes are over-represented, combined with previous reports of de novo mutations affecting autism spectrum disorders [6,18], intellectual disability [6] and epileptic encephalopathy [5], further support this notion, indicating that most perturbation leads to severe phenotype. This strong limitation in tolerance may also explain our observation of enrichment in immune cell populations, as precise control of developmental decisions is crucial to the correct differentiation of the lymphoid and myeloid lineages throughout life. As the selective constraint scores are by design corrected for both coding sequence length and GC bias [6], constraint is more likely to be due to intolerance of changes to protein function rather than structural characteristics of the encoded proteins.
Network analyses have been used to identify interacting groups of genes conserved across species [19], and to identify groups of co-expressed genes in both healthy individuals [20] and groups of genes whose expression is coordinately altered in neurological disease [21]. In particular, network analyses of expression data across species suggest that co-regulated genes form stable interaction networks that evolve in a coordinate fashion [19]. These diverse analyses all suggest that functionally linked genes form stable networks and are targets of natural selection due to their group contribution to specific biological processes [22]. Our own results support this notion, demonstrating that interacting protein networks are under remarkable constraint within the human species, presumably because they underlie carefully orchestrated biological processes.
More broadly, our results present a glimpse into how natural selection may affect entire groups of genes involved in central homeostatic functions. Most studies of selection aim to identify specific alleles inconsistent with the nearly neutral model of drift, with particular success in studies of recent positive selection [23,24]. We suggest that the majority of these effects represent near-Mendelian effects on relevant phenotypes, which are the actual targets of selective forces: for example, variability in lactase persistence is almost entirely explained by any one of handful of necessary and sufficient alleles [25]. However, the majority of human traits are polygenic, and selection would likely exert far weaker effects on risk alleles, most of which have been revealed by GWAS to only explain a fraction of phenotypic variance. Although such polygenic adaptation [26] has proven difficult to detect thus far, our data provide confirmation that selective forces can act on groups of genes involved in the same process, supporting the notion that purifying selection can act coordinately on multiple genes. We describe how selective constraint acts on groups of genes, suggesting such coordination, though we note that the constraint statistics contain no information about whether multiple genes are targets of the same pressure. We further note that the substantial preferential expression we see does not apply to the entire constrained subnetwork—this may be due either to imprecise specification of the network itself or limitations in detecting preferential expression in a limited tissue atlas. However, our results clearly support a coherent physiological role for this network in early fetal development.
We have presented a robust approach to identifying sets of interacting genes under selective constraint and placing these into biological context, using the wealth of genome-scale data produced by large-scale public projects. Our approach builds on robust statistical frameworks to interrogate single variants or genes and thus provides previously lacking biological context from which further hypotheses can be drawn. The approach is flexible and not restricted to studies of constraint: per-gene measures derived from studies of other forms of natural selection, non-human hominid introgression, common and rare variant disease association can be analyzed in our framework. Further, as PINTS is modular, appropriate tissue atlases can be used to meaningfully interpret results. We believe our work represents a new class of approaches that can leverage multiple genome-scale datasets to gain new insight into biological activities responsible for health and disease.
We have used selective constraint scores as previously described [6]. Briefly, we used a mutation rate table—containing the probability of every trinucleotide XY1Z mutating to every other possible trinucleotide XY2Z—based on intergenic SNPs from the 1000 Genomes project and the sequence of a gene to determine that gene’s probability of mutation. These sequence context-based probabilities of mutation were additionally corrected for regional divergence between humans and macaques as well as the depth of coverage for each base in an exome sequencing study. Given the high correlation (Pearson’s r = 0.94) between the probability of a synonymous mutation in a gene with the number of rare (MAF < 0.01%) synonymous variants in that gene seen in the NHLBI’s Exome Sequencing Project, we used a linear model to predict the number of rare missense variants expected per gene in the same dataset. The difference between observation and expectation was quantified as a signed Z score of the chi-squared deviation. The missense Z score was used as the basis for determining selective constraint. In this study, we took a conservative approach to assessing selective constraint, using the Bonferroni correction for number of InWeb genes to derive a significance threshold of pc < 5 x 10−6.
We used InWeb, a previously described comprehensive map of protein-protein interactions, containing 169,736 high-confidence interactions between 12,687 gene products, compiled from a variety of sources [16]. By mapping ENSEMBL IDs, we were able to identify 9729 genes with constraint scores from Samocha et al [6] also present in the REP expression data (below), to which we restricted our analysis.
To detect clusters of interacting constrained genes, we used a heuristic form of the prize-collecting Steiner tree (PCST) algorithm [27,28], which has been previously applied to protein-protein interaction data[17]. The canonical form of the PCST algorithm takes a connected, undirected graph G(V,E,w,u) with V vertices and E edges, with vertex weights w and edge weights u; it then finds the connected subgraph T(V’,E’) with maximal profit(T), which is some function of (w’-u’). By definition, T is a minimal spanning tree. The algorithm thus identifies the set of nodes with the strongest signal given the cost of their connecting edges. The classical PCST algorithm is, however, NP-hard, which makes it computationally intractable on the scale of InWeb [27]. Several heuristic simplifications have been proposed, including one previously validated as suitable for protein-protein interaction networks which we use here [17]. This approach partitions the set V into null (with weights w < 0) and signal (with weights w > 0) vertices (genes) and equal edge weights e before searching for T. Beisser et al have implemented this approach in the BioNet package for the R statistical language [29]. Here, we define signal genes as those with constraint scores passing the Bonferroni threshold of pc < 5 x 10−6, and calculate the weights as w = -log(pc) + log(5 x 10−6). The PCST algorithm returns a single, maximal T solution; to discover further independent subnetworks, we apply the method iteratively after we assign gene nodes in the previously discovered solution to be null.
The algorithm always returns a solution for T, so we sought to assess the significance of our observations empirically. To understand if the observed solution is unlikely by chance, we permuted the constraint scores of genes 1000 times and for each iteration ran the heuristic PCST to generate 1000 random resampled subnetworks (these are also used in the tissue-specificity analyses described below). We then quantified the following key parameters and assessed how many random subnetworks had values exceeding those of the true discovered subnetwork: size (number of gene nodes); density (number of connections); clustering coefficient and total amount of constraint explained (sum of constraint scores). To address the possible contribution of degree bias to these results, we also performed biased permutations to select signal nodes with the same degree distribution as we had previously done for DAPPLE [9]. We found weak correlation between degree and significance (S1 Fig) and opted for random permutations where the number of combinations of random genes selected as signal nodes is much larger.
We obtained gene expression data for a cosmopolitan set of tissues from the Roadmap Epigenome Project (REP) [14]. The REP data consists of 88 samples across 27 tissue types from diverse human organs, profiled on the Affymetrix HuEx-1_0-st-v2 exon array, which we downloaded on 9/25/2013 from http://www.genboree.org/EdaccData/Current-Release/experiment-sample/Expression_Array/. We processed these data using standard methods available from the BioConductor project [30,31]. Briefly, we removed cross-hybridizing probesets, applied RMA background correction and quantile normalization and then summarized probesets to transcript-level intensities. We then mapped transcripts to genes using the current Gencode annotations for human genes (version 12). Transcripts with no match in Gencode were removed and the remaining transcripts we again quantile normalized. We then assigned transcript expression levels to their matching genes. Where multiple transcripts mapped to the same gene we used the transcript with maximum expression over all cell types.
The Brainspan atlas [15] data are available as processed, gene-level expression levels from http://www.brainspan.org/static/download.html. We mapped these genes to the InWeb gene set using ENSEMBL IDs, and quantile normalized data for the overlapping genes. We then grouped replicate data by developmental stage and brain structure and calculated preferential expression as described above.
We used a previously described approach to detect tissue-specific expression across each tissue atlas [32]. Briefly, we group together replicates from the same cell type and compute pairwise differential expression between all pairwise combinations of tissues, using an empirical Bayes approach to account for variance shrinkage [33]. Thus, for each gene there are 26 linear model coefficients and associated p values for each tissue, quantifying the comparison to all other tissues. For each gene in each tissue, we then capture the overall difference in expression from all other tissues as the sum of these coefficients. To reduce noise, only coefficients with p < 0.0019 (p < 0.05 with Bonferroni correction for 26 tissues) are considered. Rescaling all coefficient sums across all genes values to the range [−1,1] gives us a final preferential expression score. Intuitively, a gene highly expressed in only one tissue would get a high positive enrichment score in that tissue, as it is differentially expressed compared to all other tissues. The score is directional, strong negative values indicate very low expression in one tissue compared to all others. We partition the overall distribution into deciles and define preferential expression in a tissue if a gene has a score > 0.1.
To score the tissue specific expression of a subnetwork, we detect which genes in the subnetwork are preferentially expressed in each tissue of our expression atlas and assess the joint probability of this observation. Rather than ask if some nodes of the subnetwork are preferentially expressed in a given tissue, we developed an approach to account for the connections between genes; we thus assess whether the pattern of preferential expression across the whole subnetwork is unusual for a given tissue, suggesting the subnetwork is operational. Formally, we consider the subnetwork as a Markov random field with a particular configuration of preferentially expressed nodes in each atlas tissue. We compute a score for each configuration using a standard scoring function [34]:
P (x1,…,xn)=1Z∏(i,j)∈EdgesΦ(xi, xj)
The partition function Z is defined as:
Z=∑x1,…,xn∏(i,j)∈EdgesΦ(xi, xj)
where xi(i = 1, …, n)represents a binary tissue specificity of the genes in the subnetwork for a given tissue with values either 1 (expressed) or 0 (not expressed). The Φ(xi,xj) factor lists the co-occurrence of two connected nodes across tissues. This is calculated from the thresholded preferential expression data, and each pair of connected nodes is assigned exactly one configuration in each tissue, so that
Φ(xi=0, xj=0)+Φ(xi=1, xj=0)+ Φ(xi=0, xj=1)+ Φ(xi=1, xj=1)=number of tissues
We assess the significance of these scores using two conservative permutation approaches. First we assess how likely we are to see each observed configuration (i.e. each pattern of detected/not detected nodes) in each tissue of the atlas. We do this by permuting the preferential expression scores across tissues for each gene independently and rescoring the configuration found in each tissue. This alters the co-expression structure across genes and empirically assesses how likely we are to see a particular configuration of a specific subnetwork by chance. Second, we estimate the probability of observing the extent of tissue specificity in each tissue. We construct the null expectation by scoring the resampled subnetworks generated by permutation above in each tissue and compute the empirical significance from this distribution of scores.
To ensure our results are not artifacts of a specific preferential expression threshold, we repeat this analysis across a spectrum of preferential expression thresholds (See S3 Table).
To test if any biological pathways are over represented in a subnetwork, we use the Gene Set Enrichment Analysis (GSEA) approach [35]. We obtained the full list of curated canonical pathways from the GSEA website (http://www.broadinstitute.org/gsea/msigdb/collections.jsp) and mapped the 9729 genes to each pathway using HUGO IDs. We then test for enrichment of subnetwork members over background using the hypergeometric test.
To test if genes in the subnetwork are more likely to harbor pathogenic mutations causing Mendelian diseases than expected by chance, we retrieved OMIM records for all 9729 genes using the biomaRt package in BioConductor [31]. We then tested whether the proportion of 107 subnetwork genes with OMIM entries was higher than the background proportion of the full set of 9729 in our analysis using Fisher’s exact test (S4 Table). We then mapped all OMIM entries to Medical Subject Headings (MeSH) disease categories using the Comparative Toxicogenomics Database (CTD) MEDIC disease vocabulary [36] and assessed enrichment in any disease category, again using Fisher's exact test (S6 Table).
|
10.1371/journal.pntd.0006125 | Tegumentary leishmaniasis and coinfections other than HIV | Tegumentary leishmaniasis (TL) is a disease of skin and/or mucosal tissues caused by Leishmania parasites. TL patients may concurrently carry other pathogens, which may influence the clinical outcome of TL.
This review focuses on the frequency of TL coinfections in human populations, interactions between Leishmania and other pathogens in animal models and human subjects, and implications of TL coinfections for clinical practice. For the purpose of this review, TL is defined as all forms of cutaneous (localised, disseminated, or diffuse) and mucocutaneous leishmaniasis. Human immunodeficiency virus (HIV) coinfection, superinfection with skin bacteria, and skin manifestations of visceral leishmaniasis are not included. We searched MEDLINE and other databases and included 73 records: 21 experimental studies in animals and 52 studies about human subjects (mainly cross-sectional and case studies). Several reports describe the frequency of Trypanosoma cruzi coinfection in TL patients in Argentina (about 41%) and the frequency of helminthiasis in TL patients in Brazil (15% to 88%). Different hypotheses have been explored about mechanisms of interaction between different microorganisms, but no clear answers emerge. Such interactions may involve innate immunity coupled with regulatory networks that affect quality and quantity of acquired immune responses. Diagnostic problems may occur when concurrent infections cause similar lesions (e.g., TL and leprosy), when different pathogens are present in the same lesions (e.g., Leishmania and Sporothrix schenckii), or when similarities between phylogenetically close pathogens affect accuracy of diagnostic tests (e.g., serology for leishmaniasis and Chagas disease). Some coinfections (e.g., helminthiasis) appear to reduce the effectiveness of antileishmanial treatment, and drug combinations may cause cumulative adverse effects.
In patients with TL, coinfection is frequent, it can lead to diagnostic errors and delays, and it can influence the effectiveness and safety of treatment. More research is needed to unravel how coinfections interfere with the pathogenesis of TL.
| Infectious diseases are often studied one by one, but people can have more than one infection at the same time. This is likely to happen when different microorganisms are linked to specific geographical regions or living conditions. In this paper, we summarise the literature about infections occurring together with tegumentary leishmaniasis (TL), a disease of skin and mucosal tissues that is caused by Leishmania parasites. We found that in Latin America, patients with TL are often also infected with helminths or with Trypanosoma cruzi (the parasite that causes Chagas disease). Information from other parts of the world is scarce. Animal studies and observations in humans show that one infection can change the course of another infection, but how this happens is not well understood. When different infections affect the same patient at the same time, the diagnosis can be difficult, especially when different microorganisms are biologically similar, when they cause similar lesions, or when they are present in the same lesions. Treatment can also be difficult because some coinfections reduce the efficacy of the treatment against Leishmania and because some drug combinations can lead to cumulative adverse effects.
| Tegumentary leishmaniasis (TL) is a disease of the skin and mucosal tissues caused by several species of the genus Leishmania (Protozoa, Trypanosomatida, Trypanosomatidae) that are transmitted by the bite of phlebotomine sandflies [1]. Parasites belonging to the subgenus Leishmania are found in the Old and the New World, whereas those of the subgenus Viannia are restricted to the New World [1–3]. Leishmania parasites produce a wide spectrum of clinical manifestations in humans and other mammals, ranging from asymptomatic infection to life-threatening disease [1–3]. Yearly, an estimated 1 million people develop TL, mainly in Bolivia, Brazil, Colombia, Peru, Algeria, Tunisia, Saudi Arabia, Syria, Iran, Afghanistan, and Pakistan [4].
The overlapping geographical distribution of TL with many highly prevalent (e.g., helminthiasis) [5] and some less common (e.g., leprosy) [6] infectious diseases, as well as experimental studies [7], together indicate the importance of understanding how coinfections may alter the outcome of TL and vice versa. Indeed, several infectious diseases linked to poverty, housing conditions, hygiene, or to vectors that thrive in similar circumstances tend to affect the same populations [8–12]. It is therefore likely that in the tropical and temperate regions where TL occurs, many people carry more than one pathogen at once, although the epidemiology of such coinfections is not well known. Furthermore, the clinical outcome of Leishmania infection depends on characteristics of both the Leishmania parasite and the human host immune response [13–16]. Pathogens other than Leishmania may modulate this host immune response and consequently influence the natural history of TL as well as the response to antileishmanial treatment [12,16].
The most frequently studied coinfection is that between Leishmania and human immunodeficiency virus (HIV), in that the natural history of each of the two infections is modified by the presence of the other [17]. HIV increases the risk of severe and disseminated TL, and some HIV-infected patients develop visceral leishmaniasis in the presence of Leishmania species that are usually only dermotropic [17–19]. HIV also increases the risk of TL recurrence and treatment failure [18,19]. On the other hand, leishmaniasis interferes with monocyte and macrophage function in such a way that it facilitates HIV progression [20]. Interactions between TL and infections other than HIV have not been comprehensively reviewed before.
The objectives of the present review are to summarise the evidence about the (i) frequency of TL and coinfections other than HIV in human populations, (ii) interactions between Leishmania and other pathogens in animal models and human subjects, and (iii) implications of TL coinfections for clinical practice.
We searched the medical literature to identify publications about TL and coinfections. For the purpose of this review, we defined TL as all forms of cutaneous (localised, disseminated, or diffuse) and mucocutaneous leishmaniasis. Records about the skin manifestations caused by L. donovani and L. infantum/L. chagasi (such as post–kala-azar dermal leishmaniasis) were not included because the main clinical outcome of these infections is visceral leishmaniasis, which is outside the scope of this review.
Records about HIV/AIDS and TL were not included because this topic has already been extensively reviewed elsewhere [17–19]. Records about the contamination or superinfection of TL lesions with gram-positive or gram-negative bacteria of the skin such as Staphylococcus aureus or Streptococcus pyogenes were also excluded. Review papers were not included. We did not restrict the search by geographical region, study design, language of publication, or publication date.
The databases MEDLINE, Embase, LILACS, Scielo, Cochrane, and African Index Medicus as well as local library databases, searched in August 2017, were the information sources for this review. We used search terms indicating (groups of) infections, pathogens, and diseases caused by these pathogens. The detailed search strategy for MEDLINE is given in S1 File. We also reviewed the reference lists of selected articles.
Two reviewers extracted the data from the included records; any doubts and discordances were resolved through discussion. Specific points of interest while reading and summarising the articles were (i) frequency of coinfection in humans, (ii) mechanisms of interaction and effect of coinfection on TL progression, and (iii) potential implications for clinical management. We described the information the same way the authors of the original publications did, using mainly counts, proportions, and medians.
We used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [21] to prepare this review, but it was not possible to follow all the recommendations because PRISMA mainly focuses on the evaluation of healthcare interventions and our focus was broader than that. The PRISMA checklist is given in S1 List.
The MEDLINE search retrieved 669 records, and searching other databases yielded 348 additional records. After reading titles or abstracts or both, we removed 79 duplicates and discarded 841 records because they were not relevant (Fig 1). The most frequent reason for dropping records was that, while leishmaniasis and another infection were mentioned in the same text, the publication was not about coinfection (e.g., a paper about different infections occurring in the same region but not affecting the same persons). We assessed the remaining 97 full-text records for eligibility and retained 73 for the present review (Fig 1).
The 73 articles included in this review had different study designs (Table 1). There were 21 original research papers about experimental studies of coinfection in animal models and 52 original research papers about coinfection in human patients. The 52 studies about human subjects included 1 clinical trial, 2 cohort studies, 13 cross-sectional or prevalence studies, 7 studies on the development or performance of diagnostic tests, 24 case series or case reports with a clinical focus, and 5 case series or reports with an immunological focus. The coinfecting pathogens for which we found the highest number of records were Trypanosoma cruzi (n = 18), Mycobacterium leprae (n = 14), helminths (n = 12), and M. tuberculosis (n = 9). Two records addressed coinfection of Leishmania with more than one pathogen (Table 1).
The studies providing information about the frequency of coinfection in human populations are summarised below and in Table 1.
This is the first comprehensive review of the literature about TL and coinfections other than HIV. Coinfection adds to the complexity of TL: the outcome of a single Leishmania infection in humans is difficult to predict, and the impact of coinfection on the course of TL is even more puzzling. Nevertheless, coinfection is clinically relevant because it is frequent, it can lead to diagnostic errors and delays, and it can influence the effectiveness of treatment and drug side effects. Therefore, it is crucial to gain a better understanding of the interaction between TL and other infectious diseases.
The frequency of coinfections has been studied mostly in Latin America so far. There is relatively good evidence about T. cruzi infection in Argentina (an estimated 41% of TL patients also carry T. cruzi) [36] and about helminthiasis in Brazil (an estimated 15% to 88% of TL patients also carry helminths) [5,12].
Several hypotheses have been explored about the mechanisms of interaction between the different microorganisms, but no clear answers have emerged so far from a literature that is scattered and still developing. Such interactions may involve one or all components of innate immunity coupled with the complexity of regulatory networks that affect the quality and quantity of the acquired immune responses (e.g., T-cell subset bias or regulatory cytokine production). Given that TL pathology is fundamentally an immunopathology reaction, coinfections could paradoxically lead to exacerbated TL disease by enhancing immune responses against Leishmania parasites in lesions. The impact of Plasmodium coinfection on TL in animal models is clearly detrimental; the impact of all other coinfections in animal models or human studies is less clear or less consistent.
Diagnostic problems occur when concurrent infections cause similar lesions (e.g., TL and leprosy), when different pathogens are present in the same lesions (e.g., Leishmania and S. schenckii), or when cross-reactions induced by phylogenetically close pathogens affect the accuracy of diagnostic tests (e.g., serology for leishmaniasis and Chagas disease). Regarding treatment, some coinfections seem to reduce the efficacy of antileishmanial drugs (i.e., helminthiasis), and there may be cumulative adverse effects caused by drugs or drug combinations (e.g., antimonial treatment in patients with chagasic cardiomyopathy, and combinations of antileishmanial and antimycobacterial drugs).
The strengths of this review are the broad search of the literature and the fact that the reporting follows PRISMA guidelines [21]. On the other hand, because the search strategy had few restrictions, we retrieved information in heterogeneous formats. As a consequence, we could not systematically assess the risk of bias in the individual records and decided to include all the available information. Most animal studies predate the introduction of the Animals in Research: Reporting In Vivo Experiments (ARRIVE) guidelines for reporting animal research [102]; therefore, issues related to experimental design and the avoidance of bias may not have been explicitly recorded in the publications reviewed.
Despite the broad search including several databases other than MEDLINE, the retrieved information was fragmented, and the evidence was insufficient to give firm answers to all the review questions. For example, all the evidence about TL and malaria came from animal studies without validation in humans. By contrast, all the information about tuberculosis came from human case reports with limited information about pathogenesis. In total, only 3 out of the 73 included records were cohort studies or clinical trials specifically designed to investigate the impact of coinfection on the course of TL in humans. Furthermore, there was not enough information available to look into the effect of coinfections on different clinical forms of TL (i.e., localised, diffuse, disseminated, and mucosal) separately. This is an important limitation because the host immune responses underlying these different forms of TL are contrasting and may be differentially modified by coinfections. For example, coinfections that induce a strong proinflammatory response could be beneficial in early cutaneous but detrimental in mucosal leishmaniasis. Finally, there was almost no information about coinfection in human subjects from Africa or Asia.
Several factors may have contributed to the lack of evidence about coinfections. First, coinfections tend to get less attention than single infections. Second, TL and many of the relevant coinfections are neglected diseases that affect poor populations and are typically under-researched and underreported. Finally, the complexity of TL together with other infections may lead to negative results or findings that are difficult to explain, which may reduce the chance of publication.
From a clinical point of view, several questions remain to be resolved. Even if the interactions between pathogens are complex, these clinical questions are fairly straightforward. For each of the coinfecting microorganisms, we need to better document (i) how frequent it is among patients with TL in different settings, (ii) whether TL patients with the coinfection fare better or worse than patients without it, (iii) whether the presence of the coinfection affects the accuracy of diagnostic tests, and (iv) what the best way to treat the coinfected patient is. With advances in the development of vaccines for leishmaniasis, including TL, an understanding of how vaccine responses might be modulated due to coinfection also becomes a question of some significance.
With regard to the interaction between pathogens, additional mechanisms, unexplored in the literature to date in relation to TL, are worthy of consideration. First, metabolic disturbances resulting from coinfection may alter the capacity of the immune system to appropriately respond during TL or vice versa [103,104]. Second, coinfections, in particular with helminths, may lead to a dysbiosis (i.e., alterations in the development or composition of the microbiota) that consequently impacts immune health [97,104,105]. Therefore, the answer to how the clinical outcome differs between single- and coinfected patients may not lie in understanding how two specific sets of immune responses interact but rather in how these responses are linked via complex regulatory circuits established and maintained by our commensal microbiota.
Several elements of the design of future experimental research deserve consideration. First, it is important to clarify what the outcomes of interest are, i.e., the risk of symptomatic disease, the time between infection and lesion appearance, the size of the lesion, time to healing, response to treatment, or risk of metastasis and comorbidities. The impact of coinfections on these different clinical outcomes may vary. Second, the species, the infective doses, and the timing of Leishmania and coinfection may also matter. Finally, animal models differ from each other, and they do not always represent what happens in human coinfection.
In patients with TL, coinfection with other pathogens may be the rule rather than the exception. More research is needed to unravel how other infections interfere with the pathogenesis of TL. It is important that clinicians bear in mind the possibility of coinfection because this can complicate diagnosis and treatment.
|
10.1371/journal.pntd.0006411 | Serosurveillance of Coxiellosis (Q-fever) and Brucellosis in goats in selected provinces of Lao People’s Democratic Republic | Goat raising is a growing industry in Lao People’s Democratic Republic, with minimal disease investigation to date, especially zoonoses. This study determined the proportional seropositivity of two zoonotic diseases: Q fever (causative agent Coxiella burnetii) and Brucellosis (Brucella species) in goats across five provinces (Vientiane Capital, Xayaboury, Xiengkhuang, Savannakhet and Attapeu). A total of 1458 goat serum samples were tested using commercial indirect ELISA for both pathogens, plus Rose Bengal agglutination test for Brucellosis. Overall individual seropositivity of C. burnetii was 4.1% and Brucella spp. was 1.4%. A multiple logistic regression model identified that province (Vientiane Capital, p = 0.05), breed (introduced Boer mixed breed, p = 0.006) and age (goats ≥3 years old, p = 0.014) were significant risk factors for C. burnetii seropositivity. The results of the survey indicated that province (Vientiane Capital, p<0.001), breed (introduced Boer mixed breed, p<0.001), production system (commercial, p<0.001), age (adult, p = 0.004), and farm size (large, 0.001) were all significant risk factors seropositivity for Brucella spp. It was concluded that Lao goats have been exposed to both C. burnetii and Brucella spp. however the risk of clinical disease has not yet been determined and there is an urgent need to determine human health risks and economic losses caused by Q fever and Brucellosis.
| Goat raising is a growing industry in Lao People’s Democratic Republic however there is very little information whether or not goat raising poses a disease threat to farmers and the general population through diseases that may be transmitted between animals and humans (i.e., zoonotic diseases). To determine this, we tested goats for antibodies against two zoonotic diseases: Q fever (causative agent Coxiella burnetii) and Brucellosis (Brucella species) in Lao goats across five provinces (Vientiane Capital, Xayaboury, Xiengkhuang, Savannakhet and Attapeu). The presence of antibodies does not necessarily indicate active disease but that animals have been previously exposed to Q fever and Brucellosis. A total of 1458 goat serum samples were tested and the overall antibody positivity of the goats for C. burnetii was 4.1% and Brucella spp. was 1.4%. The highest risk of having Q fever antibodies was the goats being based in Vientiane Capital, of Boer mixed breed and ≥3 years old. The highest risk of having Brucella spp. antibodies was being based in Vientiane Capital, of Boer mixed breed as well as factors related to production system, age, and farm size. There is an urgent need to determine human health risks and economic losses caused by Q fever and Brucellosis.
| Lao People’s Democratic Republic (Laos) is a landlocked country in the Greater Mekong Sub-region with an economy greatly dependent on agriculture [1]. Livestock have become increasingly important for improving rural livelihoods in Laos, providing a source of high quality protein, manure as fertiliser for plant growth, a means of household wealth storage, and income to buy food, education and healthcare [2].
Goats are becoming increasingly important for smallholder food farming in Laos [3, 4], providing livestock products that are perceived to require lower inputs than cattle and buffaloes. Furthermore, following regional economic growth there has been an increase in regional demand for goat meat in Vietnam and China, leading to rapidly increasing smallholder goat population and appearance of several commercial farms throughout Laos. Anecdotal reports suggest that there is no commercial milk or cheese production. The 2011 census reported that 45,000 farm households raised goats [3], however, it is difficult to report accurately on diversity of goat farming in Laos as it is the smallest livestock sector and is not always included in demographic reports. There is a need to focus on goat and goat farmer health in Laos as this has largely gone without investigation.
Human health is closely linked to livestock health. Healthy livestock can provide food, wealth and financial security, whereas unhealthy or diseased livestock cannot, and may be a reservoir for diseases infectious to humans (i.e., zoonoses). The close working relationship of farmers and their families with their animals allows for zoonotic disease transmission [1] with Coxiella burnetii (causing Q fever in humans) and Brucella melitensis considered potentially important bacterial zoonotic pathogens associated with goats in Laos [5]. Both agents can cause undulant fever and chronic disease in humans [5, 6]. These pathogens have the ability to cause large-scale outbreaks due to their low infectious dose, resistance in the environment and ability to travel via aerosolisation of the pathogens [5–7]. Q fever and Brucellosis are difficult to diagnose and treat in humans due to their non-specific presentation and intracellular nature [5, 7, 8]. Furthermore, C. burnetii and Brucella spp. can economically impact rural livelihoods as they reduce productivity due to reproductive loss in livestock herds [5, 8, 9]. C. burnetii and Brucella spp. are considered biothreats and classified as “Select Agents” in the USA [10, 11].
C. burnetii seroprevalence studies in Laos have revealed the pathogen is not widely distributed in cattle, despite the likelihood of an epidemiological hotspot in the Thailand-bordered province of Xayaboury [12, 13]. Thailand is a significant trading partner with Laos, and C. burnetii antibodies are reportedly present in 4% of Thai goats [14]. Surveys for Brucella spp. antibodies in Laos has revealed only a limited distribution of this pathogen in the Lao cattle herd [12, 13]. However, studies have revealed 1.4% of goats and 12.1% of goat herds in Thailand are seropositive for Brucella spp. [15]. Human Q fever and brucellosis case reporting is increasing in Thailand, with close contact with goat parturition material and consumption of raw goat meat considered as risk factors for contracting both diseases [14, 16]. One Thai study estimated that 12.6% of occupationally-exposed persons are seropositive for C. burnetii antibodies in selected provinces [14]. A report from 2004 outlines two Brucellosis cases, the first report in scientific literature since the 1970s, with one case having a history of drinking raw goat’s milk [17]. Following an outbreak of undulant fever in 2006 in a province northeast of Bangkok, village members were tested for exposure to Brucellosis of which 43.5% had antibodies to B. melitensis, with risk factors including recent contact with goats during their parturition and eating raw goat meat [16]. There are no overall human population seroprevalence or incidence estimates for Q fever or Brucellosis in Thailand available in the literature. This increase in reporting could be caused by increased disease awareness rather than an increase in disease incidence. Even though there are no reports of human Q fever or Brucellosis in Laos to date, this may be due to under-reporting rather than lack of disease manifestation, as there have been no specific investigations into these pathogens reported in the literature to date.
Considering the high risk of zoonoses due to factors including to the close working nature of farmers and livestock in Laos, porous borders, rapidly increasing smallholder goat population and appearance of commercial enterprises, it is important to investigate the prevalence of C. burnetii and Brucella spp. in Lao goats. This study reports on a serological survey of that aimed to determine the seroprevalence of C. burnetii and Brucella spp. antibodies in Lao goat herds in selected provinces and to identify potential risk factors associated with presence of infection. Furthermore, an understanding of infectious diseases in Lao goats enables development of disease prevention and control strategies and public health policies, supporting smallholder livestock farmers to increase their productivity, and therefore income, whilst minimizing human disease.
This study was conducted in compliance with State Acts and National Codes of Practice for Ethical Standards, with animal and human ethics approval obtained from The University of Sydney Ethics Committee (project no. 2015/765 and 2014/783, respectively). Verbal consent was obtained from all of the goat owners prior to the collection of the sample.
This study was conducted in five provinces in Laos: Vientiane Capital, Xayaboury, Xiengkhuang, Savannakhet and Attapeu (Fig 1). Samples were collected between October 2016 and May 2017. Vientiane Capital was selected due to the emergence of commercial goat enterprises and the convenience of proximity to the National Animal Health Laboratory (NAHL). Xayaboury was selected as a follow up to a previous study suggesting there was an epidemiologic “hot spot” for Q fever in cattle located in districts bordering Thailand [12, 13]. Samples were collected from Xiengkhuang, Savannakhet and Attapeu as part of necropsy training workshops run by the same team.
To provide some guidance regarding the number of samples to be collected, sample size was calculated using the formula: n = (Z2 P(1-P))/e2, where Z is the value from a standard normal distribution corresponding to the desired confidence level, P is the expected true proportion and e is the desired precision. A sample size to estimate individual level seroprevalence with a precision of 0±1% with an expected herd prevalence of 3.5% and confidence level of 95% was calculated using “AusVet Epitools” website [18]. As no goat seroprevalence studies for C. burnetii and Brucella spp. had been performed in Laos previously, estimated herd prevalence was extrapolated from caprine C. burnetii seroprevalence studies in neighbouring Thailand [14]. In 2011, the Lao agriculture census estimated that there were 215600 goats in Laos [3]. This resulted in a target sample size of 1291 goats however this was an estimate for the whole of Laos rather than selected provinces/districts. Using this sample size guidance, a total of 69 villages within 15 districts were selected (Fig 1). Districts were chosen with close assistance from staff from the Department of Livestock and Fisheries in Laos (DLF). Districts were visited if they had the following criteria: accessible via road vehicle; close working relationship with DLF staff; and households or farms that raised goats. Upon arrival at each district, the sampling team consulted provincial and district staff as to which villages raised goats. The sampling team visited participating different households and farms within each village following discussions between the sampling team, DLF staff, village chiefs and farmers. The team aimed to sample as many goats as possible within each village and household, this included every goat presented to the team.
Blood (3–5ml) was collected from the jugular vein into a sterile syringe and allowed to clot. The serum was removed and stored at 4°C for transport back to NAHL where it was centrifuged at 5000 rpm for 5 minutes and stored at -80°C until further analysis. Samples collected in 2017 obtained epidemiological data including date of collection, owner name, age via examination of teeth [19] and gender. Additionally, for the provinces of Vientiane Capital, Xayaboury and parts of Attapeu, breed (native Kambing-Katjang or introduced Boer/ Boer mixed), the number of goats per household owned, and production system (commercial enterprise with employed persons to raise goats or small holder, family raised household goats) were recorded. However, no such data was recorded for the 704 samples collected in 2016.
Data was entered into Microsoft Excel and analysed using Stata/SE version 15.0 for Macintosh (StataCorp, College Station, TX). For both pathogens, serological prevalence was calculated as the proportion of animals that had detectable antibodies in the same population, with 95% confidence intervals. For both C. burnetii and Brucella spp., measures of association for categorical data were assessed using either Pearson’s Chi-squared test or Fisher’s exact test as appropriate.
For calculations regarding C. burnetii seropositivity only, univariable logistic regression models were fitted to obtain unadjusted estimates of odds ratios (OR). Furthermore, a multivariable logistical regression model was performed to determine factors independently associated with C. burnetii seropositivity. Only data with full epidemiological information (754 samples) were included in this model, as the remaining samples were collected prior to 2017 without full epidemiological and clinical information. Variables with univariable significance (p≤0.05) were entered into the multivariable model. All tests of significance were performed at 5% level of significance.
The distribution of provinces, districts and villages sampled are outlined (Fig 1, Table 1). The number of serum samples collected per village ranged from 2 to 80, with a median of 19.
A total of 1458 goats were sampled. Overall, 60/1458 (4.1%; 95% CI 3.0, 5.0) of goats sampled were seropositive for C. burnetii antibodies (Table 2). Notably, the OR of individual goat C. burnetii seropositivity within Vientiane Capital was 33.4% (95% CI 4.6, 243.7, p = 0.001) and in Xayaboury was 8.4% (95% CI 1.0, 67.6, p = 0.046) respectively, with both statistically significant compared to the chosen reference province of Savannakhet (Table 2).
Within Vientiane Capital, all districts had some seropositive animals (Table 3) although there was a significant difference in seroprevalence between the districts sampled in Vientiane Capital (p<0.001), with the highest seroprevalence in Parkgnum (25.4%) (Table 3). Vientiane Capital had 75% (8/12) villages with a seropositive result, with significant difference between villages (p<0.001) including Ponsavan village with the highest seropositivity of 29.6% (Table 3). Within Xayaboury, Kentao (7.3%) was the only district to display any seropositivity.
Where full epidemiological information was available (n = 744; 51%), univariable and multivariable logistic regression was performed. The OR for Boer crossbred goat seropositivity for C. burnetii was significantly greater when compared to native Kambing-Katjang goats (OR 9.2; 95% CI 4.3, 19.8, p<0.001) (Table 2). The OR for goats sampled from commercial farms for C. burnetii seropositivity was significantly greater than goats sampled from smallholder farms (OR 5.5; 95% CI 2.8, 10.6, p<0.001) (Table 2). The OR for goats sampled from large farms indicated C. burnetii seropositivity was significantly greater when compared to other farms (OR 3.5; 95% CI 1.4, 9.0, p<0.001) (Table 2). The likelihood of seropositivity increased with age, with antibodies detected in 10.5% of goats ≥3 years of age when compared with only 2.3% of young adults (ages 1–2) and 2.1% of kids (aged <1). The adult goats (aged >3 years) had significantly increased exposure to C. burnetii than kid goats (OR 5.5; 95% CI 2.4, 12.5, p<0.001) (Table 2). However, there was no statistical difference between the young adult goats (aged 1–2) and the kid goats, (OR = 1.1; 95% CI 0.4, 2.9, p = 0.855). Overall, age was a significant variable in the univariable analyses, p = 0.014 (Likelihood ratio test, not in the table).
All variables (province, breed, production system, age category, gender and farm size) were significantly associated with C. burnetii seropositivity on univariable analysis (Table 2). There was a significant difference between genders (p = 0.025) with 4.7% of female goats being seropositive and only 1.5% of male goats demonstrating seropositivity (Table 2). Subsequently, all variables were included in the multivariable analysis. The following variables had multivariable significance: introduced Boer breed goats (OR = 6.9; 95% CI 1.7, 27.3, p<0.001); goats 3 years or older (OR 4.1; 95% CI 1.3, 12.5, p = 0.004); and goats located in Vientiane Capital had marginal significance (OR 5.4; 95% CI 1.0, 29.3, p = 0.05) (Table 2).
Overall, 20/1458 (1.4%; 95% CI 0.8, 2.2) goats tested demonstrated Brucella spp. seropositivity having serial positivity to both ELISA and Rose-Bengal agglutination tests (Table 4), despite 3.0% of goat samples returned seropositive ELISA results alone. Significant differences were noted between provinces (Pearson’s chi2 p<0.001), with the highest seroprevalence noted in Vientiane Capital (4.0%), and in Attapeu (1.6%). Brucella spp. seropositivity was not detected from Xayaboury, Savannakhet or Xiengkhuang (Table 4).
There was a significant difference of seropositivity to Brucella spp. between districts sampled in Vientiane Capital (p = 0.009), with the highest seroprevalence in Naxaythong (9.5%), followed by Xaythany (3.4%) (Table 3). Parkgnum district had no Brucella spp. positive samples. Within Vientiane Capital, only three (25.0%) villages were seropositive for Brucella spp: Ponsavan village (37.0%), Hongngua (22.6%), and Nagnang (3%) (Table 4). This finding was significant compared with other villages within Vientiane Capital (Fisher’s exact p<0.001). Within Attapeu province, there was a significant difference in seropositivity detected between districts (Fisher’s exact p = 0.014), with 2/9 (18.1%) animals seropositive in Phouvong and 1/120 (0.8%) of animals seropositive in Sahmakisai. The other two districts within Attapeu had no Brucella seropositive animals.
Where full epidemiological information was (n = 754) available, there was a significant difference demonstrated between goat breeds (p<0.001), with 5.6% of introduced Boer or Boer cross breeds found positive, and no seropositivity detected in native breeds (Table 4). There was a significant difference demonstrated between production systems (p<0.001), with 5.6% of goats on commercial farms seropositive for Brucella spp. and no seropositivity detected in smallholder systems (Table 3). Animals were more likely to have positive antibodies detected if they were sampled on a large farm with over 40 animals (5.3%), with no seropositivity detected on farms with 40 or fewer goats (p<0.001).
There was a significant difference in age of the goats (p<0.001), where adults ≥3 years recorded the highest proportion of seropositivity (3.3%) followed by goats 1–2 years old (1.1%) (Table 4). Only 0.3% of kids aged <1 year old was found to be seropositive. There was no association detected between genders (p = 0.23) (Table 3).
This study investigated the presence of exposure of Lao goats to zoonotic pathogens, C. burnetii and Brucella spp, and determined their seropositivity. The seroprevalence of both C. burnetii and Brucella spp. was much higher and more widespread in goats compared with previous cattle seroprevalence studies within Laos [12, 13]. Similar results were found in Thai goats for C. burnetii seroprevalence (4%)[14] and Brucella spp. seropositivity (1.4%), although Brucella spp. seropositivity appears more widespread in Thailand (12.1% of herds) [15]. The Coxiella study in Thailand utilised a similarly prepared ELISA from a different company (IDEXX) to this study however samples were also taken using convenience methodology so comparability of seroprevalence results between studies is limited [14]. The Brucella study utilised compliment fixation, Rose-Bengal and ELISA and if any test were positive the animal was considered positive [15] which may have artificially increased seropositivity when compared to the study presented here.
The results presented here clearly demonstrate a spatial difference of C. burnetii seropositivity with Vientiane Capital having significantly higher individual goat seropositivity than other provinces. Similar to previous studies, a hotspot of C. burnetii seroprevalence was demonstrated in cattle in Xayaboury province, located on Laos-Thailand border [12, 13]. Furthermore, goats in Vientiane were much more likely to be exposed to Brucella spp. than goats in other areas. This spatial distribution has not previously been reported for either pathogen in Laos, although is probably associated with Vientiane Capital being a major thoroughfare for international trade and the location of emerging commercial goat enterprises. In other global seroprevalence surveys for both pathogens, areas with high international trade are considered high risk for exposure to C. burnetii and Brucella spp. [22, 23]. It is a known problem that illegal movement of animals occurs throughout South East Asia through “porous borders” as demonstrated through Foot and Mouth disease outbreak studies [24]. It is possible that goats brought into Laos from other counties were already exposed to the pathogens, or that travel and intensification of production has contributed to possible infection. Nevertheless, there appears a significant public health risk to goat farmers and possibly consumers within Vientiane Capital in Laos, as recently identified with Orf virus infection [4].
It is interesting that for both pathogens, introduced Boer mixed bred goats were significantly more likely to be seropositive than native Kambing-Katjang goats, with Brucella spp. seroprevalence only reported in the introduced goats. Reasons for C. burnetii being higher in Boer mixed bred goats may be that these animals tended to be more intensively raised and on commercial farms and were likely to have been or descendants of imported animals. Intriguingly, resistance in native goats to Brucella spp. has been previously suggested with seroprevalence studies in Mexico [25] and Malaysia [22] reporting increased Brucella spp. exposure in imported breed goats compared to native animals. It has been suggested that different breeds of cattle may also be resistant to Brucella spp. infection through genetic innate immunity [26, 27]. Further studies are necessary to determine the possible role of genetics of goat immunity to a variety of pathogens.
Age was independently associated with seropositivity of C. burnetii, and there was a significant difference between age groups as risk for Brucella spp. seropositivity. Adult animals ≥3 years were more likely to be exposed to both pathogens, a finding consistent with literature and likely due to increasing opportunities of pathogen exposure [5, 28]. Similarly, female goats were more at risk of having antibody titer against C. burnetii than male goats, likely representing to the tropism for both pathogens to the placenta and mammary lymph nodes [6, 8].
Results demonstrated here indicating risk factors for infection in Laos included commercialization in comparison with smallholder systems and association of larger farms with higher seroprevalence are in contrast with studies in Thailand and elsewhere [15, 29, 30]. It is thought that smallholder farmers with free ranging goats were at higher risk for Brucella spp. exposure as the mobility of wandering herds favours spread of infectious disease when allowed to mix with naïve herds [15, 29, 30]. Intensification has been reported as a risk factor in seroprevalence studies [25, 31], where close contact within herds may also favour pathogen spread. Although herd size and commercialisation were significant risk factors on univariable analysis for C. burnetii seroprevalence, neither variable was significant on multivariable analysis and hence were potential confounding variables, despite the likelihood that these herds were introduced to infection before or after importation.
To accurately estimate prevalence of a disease in the population, sensitivity and specificity of a test must be known to approximate the occurrence of false results. The ID-Vet Q fever iELISA has been determined at 100% sensitivity and 100% specificity for Coxiellosis in cattle located in France, however no sensitivity and specificity reports have been performed for small ruminants [25, 31] so these estimations may not be accurate for small ruminant studies. Furthermore, as the status of a disease (endemic or not endemic) in a population can alter the sensitivity and specificity of a test in a region, analysis of sensitivity and specificity is required in Laos, and South East Asia as a whole.
Internal company testing of the ID-VET Brucella iELISA found 100% specificity on a herd of 160 goats in France, and 100% sensitivity on 5 goats in Southern Italy [21]. The small sample size brings into question the reliability of the estimations, which additionally might not be applicable to the region of South East Asia as sensitivity and specificity, can differ by regions [32, 33]. The Rose-Bengal test has been assessed at 94% sensitive and 99% specific [34] yet there is no estimation of sensitivity or specificity for the combined tests specific to the region. Latent class analysis can be used to determine the sensitivity and specificity of tests in the absence of a gold standard test for the region. Furthermore, it can be estimated using latent class analysis which test is most suitable, in what order it should be performed, and also if the tests should be utilised in parallel (all positives are included), or serially (only positives on all tests included, as in this study). For example, studies of cattle serology in Zambia indicated a competitive-ELISA paired with Fluorescent Polarisation test results in the highest sensitivity and specificity [35], and for sheep in Europe the blocking-ELISA is most accurate [36], and finally Rose-Bengal plus competitive ELISA test gave the best results for cattle testing in Zimbabwe [37]. However, it is a weakness of this study that this analysis has not been performed to determine the utility of testing in parallel or serially.
Brucella spp. and Q fever serology results must be interpreted in the context that Lao goats are not vaccinated for either disease, nor are there active control programs and as such it is highly unlikely that any of the positive serology results were attributable to local vaccine strains. Nevertheless, it is a limitation of this study that the farmers were not queried about vaccination. A positive Brucella spp. serology result can be caused by a cross-reaction with a range of bacterial species including Yersinia spp. giving rise to results that may not be fully accurate [38] although repeating tests can increase the reliability of results [34]. It is possible that the use of serial testing with Rose-Bengal and ELISA for Brucella antibodies reduced the number of sites with Brucella seropositivity than would have been reported by ELISA alone. Furthermore, following discussions with the ELISA kit manufacturers, a decision was made to classify "suggestive" positive samples as negative thereby decreasing the possibility of false-positive results. While the decision to employ serial ELISA and Rose-Bengal in this study was based on methodologies of previous Brucella spp. studies within Laos [12, 13], there is potential for latent class analysis to be utilised on this data set to determine the best combination of tests for the most appropriate sensitivity and specificity result in Laos.
In a preliminary investigation during this study, vaginal swabs were collected from goats located within Vientiane Capital, with one found to be positive for Brucella spp. DNA with real time PCR and none to have C. burnetii DNA (personal communication, Dr. Reka Kanitpun). This preliminary finding indicated that farmers and their families were at potential risk for contracting Brucellosis from their goats, although further investigations are needed to understand shedding patterns and speciation of Brucella spp., preferably using molecular diagnostic tools. This study did not differentiate the species of Brucella spp. as both the serological tests utilised were genus specific only. Goats are generally associated with B. melitensis, the species that is most pathogenic to humans, yet infection with B. abortus or B. suis is also common [5].
Although this present seroprevalence data demonstrates previous exposure of animals to these pathogens, serology alone does not provide a complete picture of infection status within an animal population. Seropositivity does not indicate disease manifestation, current shedding of pathogens, or consequently current risk of transmission. Studies have suggested that a significant proportion of animals that shed C. burnetii or Brucella spp. are not seropositive; furthermore animals can be seropositive and not be shedding [5, 6]. Furthermore, the proportion of animals shedding C. burnetii is independent of abortion history in a herd, and shedders might represent clinically unapparent infections [9, 39].
Outbreaks of human Q fever and Brucellosis are commonly linked to seasonal parturition in small ruminant production system [5, 9, 40, 41]. In developing nations, many fevers presenting to medical clinics go undiagnosed due to their general “ill-thrift” nature [42, 43]. It is imperative that medical practitioners in Laos are aware of Q fever and Brucellosis as differential diagnoses, especially for at risk populations, including livestock farmers with recent exposure to animal parturitions, pregnant women and people consuming raw goat products.
Control of zoonotic and transboundary disease pathogens proves difficult in developing nations where veterinary support and resource may be limited or unavailable. Vaccination of animals for either C. burnetii or Brucella spp. is not recommended as it must be given according to parturition calendars. Furthermore, vaccination will only reduce but not stop shedding, can cause goat abortions if given at incorrect times and can cause human disease if self-inoculated [5, 40, 42]. Test and culling is recommended for Brucella spp. seropositive farms and may have a role in this study where very few villages were considered likely to be affected. It is acknowledged this is currently difficult politically and financially. Disinfecting farms quarterly may reduce disease spread for both pathogens[22] although may not be applicable to rural village settings with free ranging goats.
Despite these issues, this study has addressed important knowledge gaps on C. burnetii and Brucella spp. seroprevalence in Lao goats whilst raising a number of other questions. Further investigations of the potential risk factors for transmission of the different species and farming practices are necessary to determine why Boer crossbred goats have higher seroprevalence. Further studies investigating shedding of both C. burnetii and Brucella spp. are required for speciation and potential trace back of disease transmission, plus collection of caprine placentas for PCR and pathology. There is urgent need to determine current Q fever and Brucellosis seroprevalence and occurrence of the diseases in humans, especially in at -risk populations including livestock farmers, others exposed to goat effluent, plus people consuming raw goat meat or milk products [17]. With the increasingly important contribution of goats to Lao and regional food security, the zoonotic issues from Lao production systems will very likely become increasingly important. International aid groups and commercial farms are advised to serologically test goats prior to importing them into Laos, and work closely with Lao veterinary services to ensure limited pathogen spread occurs both between villages and from animals to humans.
|
10.1371/journal.pgen.1002136 | Multiple Regulatory Mechanisms to Inhibit Untimely Initiation of DNA Replication Are Important for Stable Genome Maintenance | Genomic instability is a hallmark of human cancer cells. To prevent genomic instability, chromosomal DNA is faithfully duplicated in every cell division cycle, and eukaryotic cells have complex regulatory mechanisms to achieve this goal. Here, we show that untimely activation of replication origins during the G1 phase is genotoxic and induces genomic instability in the budding yeast Saccharomyces cerevisiae. Our data indicate that cells preserve a low level of the initiation factor Sld2 to prevent untimely initiation during the normal cell cycle in addition to controlling the phosphorylation of Sld2 and Sld3 by cyclin-dependent kinase. Although untimely activation of origin is inhibited on multiple levels, we show that deregulation of a single pathway can cause genomic instability, such as gross chromosome rearrangements (GCRs). Furthermore, simultaneous deregulation of multiple pathways causes an even more severe phenotype. These findings highlight the importance of having multiple inhibitory mechanisms to prevent the untimely initiation of chromosome replication to preserve stable genome maintenance over generations in eukaryotes.
| Chromosomal DNA replication occurs as a two-step reaction in eukaryotes. In the first reaction, called licensing, the replicative helicase is loaded onto replication origin in an inactive form during the G1 phase of the cell cycle. In the second reaction, called initiation, the replicative helicase is activated, and replication forks are established. Because of this two-step mechanism, licensing and initiation must occur at different times in the cell cycle. Failure of this two-step regulation will cause heterogeneous re-replication of chromosomal DNA, and genome integrity will be lost. Although previous works have established that multiple regulatory pathways regulate licensing, much less is known about how untimely (premature) initiation is prevented during the G1 phase. In this paper, we show that untimely activation of replication origins during the G1 phase is inhibited on multiple levels. Notably, deregulation of a single pathway can cause genomic instability; simultaneous deregulation of multiple pathways causes a more severe phenotype, such as aneuploidy. Therefore, these findings not only indicate the importance of having multiple inhibitory mechanisms to prevent untimely initiation of chromosome replication but also should help us understand how replication might be deregulated in human cancer cells, in which the genome is frequently destabilized.
| When eukaryotic cells proliferate, their chromosomes must be precisely duplicated and segregated to daughter cells to maintain genome stability over generations. Failure of these processes is directly connected to lethality and severe disease, such as cancer. Genome instability is a hallmark of human cancer cells. To duplicate chromosomal DNA precisely, DNA replication must be restricted to occur exactly once per cell cycle. Chromosomal DNA replication in eukaryotes initiates from multiple specific regions of chromosomal DNA, called origins of replication. Therefore, it is important to regulate the activation of replication origins to only once per cell cycle because multiple rounds of origin activation per cell cycle will cause over-replication. Such re-replication might cause copy number heterogeneity throughout the genome, and genome integrity will be lost.
In eukaryotes, replication origin activation occurs as a conserved two-step reaction (for reviews, see [1]–[3]). In the first reaction, known as licensing, a specific protein-origin DNA complex, called the pre-replicative complex (pre-RC), is assembled at origins during the G1 phase of the cell cycle by the loading of an inactive form of the Mcm2-7 helicase complex. In the second reaction, called initiation, the pre-RC is activated, and bidirectional replication forks are established for DNA synthesis. Because of this two-step DNA replication initiation mechanism, these two reactions must occur in separate cell cycle phases. Therefore, initiation does not occur when cells assemble pre-RCs in G1 phase, and pre-RC assembly is inhibited when initiation can occur from S to M phase. Thus, DNA replication is limited to once per cell cycle.
Eukaryotes have multiple mechanisms to prevent pre-RC re-assembly at activated origins, [4]. For example, in the budding yeast Saccharomyces cerevisiae, all components of the pre-RC, including ORC, Cdc6, Cdt1, and Mcm2-7, are inhibited by the master cell cycle regulator, cyclin-dependent kinase (CDK). ORC is inhibited by CDK through phosphorylation of Orc2 and Orc6, and the S-phase cyclin Clb5 binds directly to an RXL motif in Orc6 [5], [6]. Cdc6 is also inhibited by CDK in three ways: transcription, proteolysis and direct association with mitotic CDK [7]–[9]. Finally, nuclear accumulation of Mcm2-7 and Cdt1 is inhibited by CDK activity [10]–[12]. Each of inhibitory reaction contributes to prevent pre-RC formation. Because of these multiple down-regulatory mechanisms, the deregulation of any one mechanism does not cause a severe phenotype. However, the simultaneous deregulation of more than one mechanism causes a more severe phenotype. Finally, the simultaneous deregulation of all of the mechanisms causes robust DNA re-replication [5], [9]. Because any one mechanism is insufficient to inhibit pre-RC formation completely, it is important to have multiple mechanisms to strictly enforce once per cell cycle DNA replication. Indeed, multiple inhibitory pathways for the formation of the pre-RC are common in model eukaryotes, although the specific mechanisms are different between organisms [4].
Untimely activation of the pre-RC during G1 phase must also be prevented because origin firing in G1 results in the reformation of pre-RC at replicated origin DNA, leading to multiple rounds of replication of the region [13], [14]. CDK and DDK (Dbf4-dependent kinase, which consists of Cdc7 and Dbf4) are conserved protein kinases and are required for activation of the pre-RC in eukaryotes [3]. Until recently, budding yeast was the only organism in which the essential targets of CDK in initiation had been identified. In this organism, S phase-specific CDKs (S-CDKs: Clb5- and Clb6-Cdc28) phosphorylate two essential replication proteins, Sld2 and Sld3, to promote DNA replication [13]–[15]. Phosphorylation of Sld2 and Sld3 enhances the interaction between Sld2 or Sld3 and a third protein Dpb11, respectively [13]–[16]. These interactions are not only essential for initiation but are also sufficient to bypass the requirement for CDK in the initiation of DNA replication. Combinations of mutations that can bypass the CDK phosphorylation of Sld2 and Sld3 allow cells to promote DNA replication [13], [14]. For example, CDK phosphorylation of Sld2 and Sld3 can be bypassed by phosphomimetic form of Sld2 and the Cdc45Jet1-1, respectively [13]. When phosphomimetic Sld2 (Sld2-11D) is induced from a galactose-inducible promoter (GALp) in the CDC45JET1-1 background, DNA replication occurs even in G1-arrested CDK inactive cells. Under this condition, DNA re-replication occurs as expected, indicating that repeated formation and activation of the pre-RC is occurring [13].
This “CDK-bypass” DNA replication, surprisingly, requires neither bypass of DDK nor artificial expression of Dbf4 [13]; however, it is inhibited by inactivation of DDK. Although DDK's regulatory subunit, Dbf4, is degraded via anaphase promoting complex/cyclosome (APC/c) in G1 phase [17]–[19], these observations suggest that G1 cells have residual DDK activity and that activity is sufficient to induce DNA replication [13]. This further suggests that CDK activity is crucial for the initiation of DNA replication during G1 phase. We thus examined a CDK-bypass strain to elucidate the consequences of untimely initiation and how it is prevented in wild-type yeast cells. Our results show that untimely initiation in G1 causes a severe loss of viability and the genomes of surviving cells are very frequently destabilized. To prevent untimely initiation and to maintain genome stability, multiple mechanisms are employed.
Untimely activation of the pre-RC in G1 phase causes multiple rounds of origin activation. To understand the effect of untimely DNA replication in G1 on cell viability, we induced untimely DNA replication in G1 through the high-level expression of phosphomimetic Sld2 in CDC45JET1-1 cells and monitored cell viability. Cells were arrested and kept in G1 phase with alpha factor, and then, various Sld2 derivatives were expressed. Only when phosphomimetic Sld2s (Sld2-11D: all potential 11 CDK phosphorylation sites are substituted by aspartic acid and Sld2-T84D: only essential threonine 84 is substituted by aspartic acid) were expressed did DNA replication occur as previously shown (Figure S1A). Because cells were kept in alpha factor-containing medium, S-CDK was not activated, and Orc6 protein, a phosphorylation target of S-CDK [5], was maintained in a hypophosphorylated fast-migrating form during the experiment (Figure S1C). When DNA replication occurred in phosphomimetic Sld2-expressing cells, their viability was rapidly diminished (Figure S1B, S1C). For example, one hour after Sld2-11D and Sld2-T84D induction, the increased DNA content estimated from flow cytometry was only approximately 10% and 1.5% (see Materials and Methods for details), but 93% and 85% of cells lost viability, respectively (Figure S1A, S1B). When origin firing is induced by phosphomimetic Sld2 and Jet1-1, pre-RCs re-assemble again at origins because in alpha factor-arrested cells, low CDK activity allows pre-RC formation in budding yeast. Therefore, untimely replicated portions of chromosomes will be replicated repeatedly in the same G1 phase or in the following S phase. It has been shown that multiple rounds of DNA replication (re-replication) occur during CDK-bypass replication [13]. Therefore, our results imply that cell viability is very sensitive to re-replication.
To further confirm that untimely replication causes a loss in viability, we utilized a cdc7-4 temperature-sensitive allele of CDC7 that is defective in the catalytic subunit of DDK. Previously, it has been shown that DDK activity is required for CDK-bypass DNA replication in CDC45JET1-1 GALp-sld2-11D cells [13]. At the restrictive temperature, cdc7-4 inhibits CDK-bypass DNA replication [13]. cdc7-4 CDC45JET1-1 GALp-sld2-11D cells were first arrested in G1 with alpha factor and then kept in alpha factor-containing medium until the end of the experiment to maintain low CDK activity. After G1 synchronization, the temperature was shifted to the restrictive temperature (37°C) for cdc7-4, and then Sld2-11D was expressed by the addition of galactose. High-temperature incubation not only prevented untimely DNA replication but also inhibited the loss of viability (Figure S2). In contrast, at 25°C, DNA replication occurred, and in addition, viability was lost when galactose was added (Figure S2). Therefore, we conclude that the loss of viability observed here is caused by untimely DNA replication during G1 phase rather than the high level of phosphomimetic Sld2 itself.
CDK-bypass DNA replication does not require artificial expression of Dbf4 because G1 cells have residual DDK activity as described above [13]. Moreover, ectopic expression of Dbf4 enhances the extent of DNA replication [13], which further predicts that the expression of Dbf4 in G1-arrested cells leads to more re-replication and a greater loss in viability. To test this possibility, we simultaneously induced Dbf4 from a galactose-inducible promoter with Sld2-11D in CDC45JET1-1 GALp-sld2-11D cells (Figure S3). Although the effect was not strong, Dbf4 expression resulted in more DNA replication and loss of viability (Figure S3). These results further indicate that the loss in viability observed here is a direct consequence of re-replication.
Although untimely initiation induced in G1-arrested cells killed most of CDC45JET1-1 GALp-sld2-11D cells (Figure S1B), small numbers of surviving cells were recovered on glucose-containing plates. Similar to the original cells, these surviving cells were unable to grow on the galactose-containing plate. To examine the effect of DNA re-replication on genome stability, we analyzed the chromosomes of survivors by pulsed-field gel electrophoresis (Figure 1A). In the wild-type control cells, of 10 clones examined, no survivors showed gross chromosome abnormalities except for fluctuation in the length of chromosome XII (Figure S4A). Chromosome XII harbors the rDNA repeats, and the rDNA copy number fluctuates naturally [20], [21]. Therefore, we omitted chromosome XII from the analysis. In contrast to wild-type cells, CDC45JET1-1 GALp-sld2-11D survivors had an abnormal chromosome composition. For example, in survivors 6, 11 and 17 (Figure 1, lanes 6, 11 and 17), the band intensity ratio of chromosome III was almost two times higher than that of the control (lane C at both ends), indicating a duplicated chromosome III. In addition, much aneuploidy was observed (Figure 1, filled arrowheads). Moreover, some other survivors appeared to have chromosomes with an atypical length (Figure 1, lanes 7, 8, 9, 15 and 17, open arrowheads). Of 17 clones examined, 14 showed chromosome abnormalities (Figure 1). These data show that untimely DNA replication induces abnormal chromosome composition. CDC45JET1-1 GALp-sld2-11D cells maintained high viability when Sld2-11D was not induced (Figure S1B). However, four of 12 clones obtained from this condition showed an abnormal chromosome composition (Figure S4B, data not shown). This result suggests that chromosome composition is frequently altered in CDC45JET1-1 GALp-sld2-11D cells, even when the cells retain high viability.
As shown above, the expression of phosphomimetic Sld2, Sld2-T84D or Sld2-11D is a key to the induction of untimely DNA replication in the CDC45JET1-1 background. To further understand the conditions for untimely DNA replication, we next tried to replace the genomic copy of SLD2 with sld2-11D in the CDC45JET1-1 background. We first expected that the sld2-11D CDC45JET1-1 strain could not be isolated because of the re-replication phenotype. Surprisingly, we were able to isolate the sld2-11D CDC45JET1-1 strain, although the cells grew very slowly (Figure 2A). Moreover, the cells arrested in G1 did not replicate DNA (Figure 2B) or lose viability (Figure 2C), although a significant amount of Sld2-11D protein was observed (Figure 2D).
The transcript level of SLD2 fluctuates during the cell cycle and peaks at the G1/S boundary, leading to fluctuation in the Sld2 protein level and its accumulation in S phase [15], [22]. We thus examined whether or not the S-phase level of Sld2 efficiently induces untimely DNA replication. To obtain the S-phase level of Sld2 in the absence of S-CDK activity, we expressed a stable form of Sic1, Sic1ΔNT [23], in G1 phase cells released from alpha factor arrest. Expression of Sic1 inhibits S-CDK activity but not G1-specific CDK (G1-CDK: Cln-CDK in the budding yeast) activity, which induces SLD2 transcription at the G1/S boundary. As a consequence, SLD2 is expressed at the normal level seen in S-phase, even in the absence of S-CDK activity. sld2-11D CDC45JET1-1 GALp-SIC1ΔNT cells were first arrested in G1 with alpha factor, and Sic1ΔNT was then induced before transfer into fresh medium lacking alpha factor. Control cells harboring wild-type SLD2 arrested at the G1/S boundary because of the high level of Sic1ΔNT, and DNA replication did not occur (Figure 3A). In contrast, DNA replication occurred in cells harboring sld2-11D (Figure 3A). The percentage of budded cells, which is indicative of G1-CDK activity, increased 30 to 60 minutes after release from the G1 block in all strains (Figure 3B). Reflecting this result, the Sld2-protein level increased in all strains (Figure 3C), whereas DNA replication occurred only in the sld2-11D cells (Figure 3A). Therefore, these results suggest that untimely DNA replication requires a higher level of Sld2-11D than is normally seen during early G1 phase.
To further explore how the protein level of Sld2-11D affects untimely replication, we controlled the protein level of Sld2-11D in GALp-sld2-11D CDC45JET1-1 cells arrested in G1 phase with alpha factor by adding various concentrations of galactose to the medium and monitored DNA replication (Figure 4). The protein level of Sld2-11D increased as the galactose concentration increased (Figure 4B). The level of induced Sld2-11D protein was similar to endogenous Sld2, even after one hour treatment with 0.01% galactose, which was not enough to induce obvious DNA replication. However, the addition of more than 0.025% galactose induced DNA replication, and the induction of a higher protein level of Sld2-11D caused more DNA replication (Figure 4). Thus, the protein level of Sld2-11D is crucial for efficient induction of DNA replication.
The data shown above suggests that the Sld2 protein level in G1 phase is a limiting factor for the initiation of DNA replication. The limited level of Sld2 may contribute to the inhibition of untimely DNA initiation in G1 phase during normal cell proliferation. Because constitutive expression of Sld2 and Sld2-11D from the GAL promoter does not affect the overall rate of cell growth (data not shown), we hypothesized that this might cause a slight enhancement of initiation that does not confer slow cell growth or cell death. To detect such inefficient initiation, we employed the gross chromosome rearrangement (GCR) assay [24], which efficiently detects abnormal chromosomal transactions, such as untimely initiation. GCRs are chromosomal abnormalities, such as translocations, deletion of a chromosome arm, and interstitial deletions or inversions. In this assay, by measuring the loss rate of two counter-selectable markers, URA3 and CAN1 on chromosome V, the rate for GCR generation of the strain can be calculated [24]. When control cells harboring the empty GALp vector were grown in galactose-containing media, the GCR rate was 0.79×10−10/cell division (Table 1). When Sld2 was expressed, a GCR of 13 times higher was observed (Table 1, Figure 5A). This result suggests that high levels of Sld2 affect genome stability. Interestingly, more than 700 times higher GCR rate was observed when Sld2-11D was expressed (Table 1, Figure 5A). Because the expression level of Sld2 and Sld2-11D was similar (Figure 5B), the much higher GCR rate for Sld2-11D was possibly induced by the phosphomimetic property of the Sld2-11D protein. These results suggest that increased expression of Sld2 causes untimely initiation, and this occurs more frequently in Sld2-11D-expressing cells.
When untimely origin activation occurs through high levels of Sld2, re-assembly of the pre-RC is the next event required for re-replication to occur. Although the G1 phase is a window of time during which cells can form the pre-RC, the potential for pre-RC formation is limited by the fact that Cdc6 is unstable even during G1 [25]. Therefore, when untimely pre-RC activation is induced by high levels of Sld2 in the G1 phase, simultaneous expression of Cdc6 should cause a further increase in the GCR rate by increasing the potential for re-assembly of the pre-RC at activated origins. To test this idea, we combined GALp-CDC6 and GALp-SLD2 for simultaneous expression. In control cells, in which Cdc6 alone is expressed, the GCR rate changed very little (2.0 times higher than that of GALp vector). This result is expected because pre-RC assembly itself does not cause untimely activation of the pre-RC. In contrast, simultaneous expression of Cdc6 and Sld2 resulted in a highly elevated GCR rate. The GCR rate of GALp-CDC6 GALp-SLD2 cells was 430 times higher than that of the control vector and was 31 times higher than that of GALp-SLD2 cells (430×10−10/11×10−10 = 31), in which SLD2 alone was expressed (Figure 5A and Table 1). Even with the GALp-sld2-11D background, enhancement of the GCR rate by simultaneous Cdc6 expression was observed (from 730 to 1100 times higher than control, Figure 5A and Table 1). These data further support the possibility that high levels of Sld2 in G1 phase cause untimely initiation.
To further investigate whether untimely initiation in G1 phase is the reason why high level of Sld2 induced a higher GCR rate, we modulated the Sld2 expression pattern during the cell cycle. For this purpose, two types of cell cycle-dependent degron tags were attached to the Sld2 N-terminus to control the accumulation pattern of Sld2 during the cell cycle. One tag was the destruction box (Db) of Clb2, which is unstable in G1 and is responsible for the degradation of Clb2 from late M to G1 phase [26]. The other tag was the N-terminal 100 amino acids of Sic1 (Sic1N), which is degraded when CDK is active [27]. Sic1N lacks the CDK inhibitory domain, and thus, its expression does not affect cell cycle progression ([28], data not shown).
When untagged Sld2s were expressed from the GAL promoter, their protein levels were high throughout the cell cycle (Figure 5B). In contrast, Db-tagged Sld2 specifically disappeared from G1 cell extracts, although it accumulated at a high level in S or G2/M extracts (Figure 5C). The GCR for the GALp-Db-SLD2 strain was decreased to one-sixth of the GALp-SLD2 strain (from 13- to 2.1-fold) and was only two-fold higher than that of the control GALp-Db strain (from 2.1- to 0.91-fold) (Table 1 and Figure 5A). The GCR rate of the GALp-Db-sld2-11D strain was also decreased to less than one-fifth of the GALp-sld2-11D strain (from 730- to 140-fold) (Table 1 and Figure 5A). When cells were not expressing Sld2, GCR rates did not increase in any case (Table 1 and Figure 5A). Therefore, destabilization of Sld2 in G1 phase by the Db-tag suppressed the increase in GCR rates in GALp-SLD2 cells.
On the contrary, Sic1N-tagged Sld2 was destabilized specifically from S to M phase, and importantly, its protein level was high in G1 phase (Figure 5D). The GCR rate of the GALp-sic1N-SLD2 strain did not decrease, but rather, it increased from 13- to 190-fold (Table 1 and Figure 5A), although the exact reason for this increase is not clear. The GCR rate of the GALp-sic1N-sld2-11D strain was only affected modestly by protein destabilization from S to M phase (reduced from 730- to 460-fold) (Table 1 and Figure 5A). Overall, these data indicate that high level of Sld2 in G1 phase is the primary reason for the elevated GCR.
As described above, high levels of Sld2 increase the GCR rate, and the effect is enhanced by simultaneous expression of Cdc6. These results strongly suggest that untimely replication in G1 induced by high level of Sld2 is the primary reason for the elevated GCR. However, another mechanism is possible that high level of Sld2 in G1 might titer away other replication factors and compromise either pre-RC formation or fork progression, which would then cause the GCRs, although this mechanism is not very likely because there are no known roles of Sld2 in pre-RC assembly and fork elongation. To test the possibility, we inserted the efficient replication origin ARS306 at the YEL062w locus, which is proximal to the GCR marker CAN1 (YEL063c) on chromosome V. We deleted original ARS306 on chromosome III to avoid the duplication of the ARS306 sequence over two different chromosomes. If elevated GCR in GALp-SLD2 cells is caused by repetitive origin firing, yel062w::ARS306 cells would show a higher GCR rate, while defects in pre-RC assembly or fork progression would be rescued by ARS insertion and GCR would be repressed. In the control (GALp vector yel062w::ARS306) cells, ARS306 insertion did not affect the GCR rate (Table 1 and Figure 5A). In contrast, the insertion increased the GCR rate of GALp-SLD2 cells approximately ten fold when Sld2 was induced (from 13- to 130-fold), although GALp-sld2-11D cells did not show a significant change (from 730- to 500-fold) (Table 1 and Figure 5A). These results exclude the possibility that high level of Sld2 causes defective origin activation or fork progression.
As shown in Figure 2 and Figure 3, in the CDC45JET1-1 background, untimely DNA replication cannot be induced by G1-level Sld2-11D. This result suggests that even Sld2-11D would not affect the overall genome stability, such as the GCR rate, if it is expressed at an endogenous level. As expected, when Sld2-11D was expressed as a sole genomic copy of Sld2 by replacing the genomic copy of SLD2 with sld2-11D, the GCR rate was almost the same as that of the wild type (Figure 6A and Table 2). The Sld2-11D protein level in G1 was higher than that of the wild type in G1 but lower than that of the wild type in S phase (Figure 6B, compare lanes 9, 10, and 16).
In contrast, because S phase-level Sld2-11D can induce untimely DNA replication in G1/S phase-arrested cells (Figure 2 and Figure 3), we asked whether the S phase-level Sld2-11D in G1 phase can induce a higher GCR rate or not. For this purpose, SLD2 promoter-regulated SLD2, sld2-11D, sic1N-SLD2, or sic1N-sld2-11D were inserted at the LEU2 locus. Cells with extra copies with SLD2 or sld2-11D accumulated Sld2 or Sld2-11D protein at a level higher than that of wild type in G1 and is similar to that in S phase (Figure 6B, compare lanes 9, 11–13, 16, and 18). When Sic1N-Sld2 or Sic1N-Sld2-11D was expressed, they were observed only in G1 cells (Figure 6B lanes 14, 15, 21 and 22, and data not shown). Therefore, Sld2 amount was increased only in G1 phase in these cells. The GCR rate was increased when cells had an excess amount of Sld2-11D (+sld2-11D cells: 82-fold, +sic1N-sld2-11D cells: 11-fold (Figure 6A and Table 2)), and even the cells with an excess amount of Sld2 tended to increase GCR (+Sld2 cells: 1.3-fold, +sic1N-Sld2 cells: 1.4-fold (Figure 6A and Table 2)). These data indicate that it is important to keep the Sld2 level low to prevent untimely replication in G1 cells.
In addition to Sld2, Dpb11 and Sld3 are required for initiation, and these proteins all form a complex regulated by CDK, which is crucial for the initiation of DNA replication. The expression of Dpb11 and Sld3 are constant throughout the cell cycle, and their expression levels are relatively low ([29], data not shown). This fact also raises the possibility that high-level expression of Sld3 or Dpb11 might cause untimely initiation and hence a higher GCR rate. We tested this possibility with a GALp-DPB11 strain. A High level of expression of Dpb11 resulted in an approximately 60 times higher GCR rate (Table 1). This result suggests that limiting the level of proteins involved in the initiation reaction is important to prevent untimely initiation and hence is important for genome stability.
As shown in Figure 1, re-replication induced by untimely DNA replication in G1 causes abnormal chromosome composition. Relicensing experiments using Xenopus egg extracts suggested that multiple rounds of initiation from the same origin would generate consecutive replication forks travelling in the same direction and that they may finally collide [30]. Such collisions may generate extruded DNA strands, which will be recognized by DNA damage response machinery [30], [31]. A similar situation would occur in re-replicating DNA induced by untimely initiation in G1. To address this possibility, we have monitored the phosphorylation of Rad53, an essential protein kinase required for cell cycle checkpoint function, and foci formation of Ddc1, a subunit of a PCNA-like complex required for DNA damage response (Figure 7). CDC45JET1-1 GALp-sld2-11D DDC1-GFP cells were arrested in G1 phase with alpha factor, Sld2-11D was expressed temporally by galactose addition, and then cells were synchronously released into glucose-containing medium. When Sld2-11D is not expressed [Raff (OFF)], cells can finish the cell cycle and enter into the next cell cycle after the release (Figure 7A). In contrast, when Sld2-11D was induced [Gal (ON)], untimely DNA replication occurred as in previous experiments (see Figure S1, S2, S3). Release from G1 arrest allowed cells to enter S phase within 30 min. After bulk DNA replication, cells were arrested in G2 with more than 2C DNA (Figure 7A). As cells pass through S phase, phosphorylated Rad53 and Ddc1 foci accumulated only when Sld2-11D was induced (Figure 7B–7D). This result indicates that a checkpoint pathway is activated in these cells and further suggests the occurrence DNA damages. Because Ddc1-foci formation is observed in many cells, this indicates that at least double-strand DNA breaks were generated as cells go through S phase.
Because DNA replication in eukaryotes occurs as a two-step reaction, activities for these steps, pre-RC assembly and activation, must be separated to prevent re-replication in the cell cycle. In this study, we have investigated how untimely activation of the pre-RC, the initiation of DNA replication, is prevented in G1 phase. When untimely pre-RC activation and consequent DNA replication are induced in G1-arrested cells, cells lose viability very quickly (Figure S1). Because the induction of untimely initiation of DNA replication itself does not immediately activate Rad53, a checkpoint kinase activated by DNA damage (Figure 7, data not shown), it is likely that abnormal chromosome structures and/or DNA damage generated later by re-replication are genotoxic rather than untimely initiation itself. In fact, the survivors recovered from re-replicating cells frequently have abnormal chromosome compositions (Figure 1). Although at least double-strand DNA breaks are occurring later in the cell cycle (Figure 7) as in the case of abrogation of the mechanism to inhibit relicensing [32], it is still unclear what types of structures cause such damage or whether other types of DNA damage are generated by this re-replication. Multiple rounds of initiation may generate multiple replication forks chasing one another along the same DNA template. For example, head-to-tail replication fork collision is suggested to occur when re-replication is induced by relicensing of activated origins in Xenopus egg extracts [30].
When the mechanism to inhibit relicensing is partially abrogated in Saccharomyces cerevisiae, specific chromosomal loci are preferentially re-replicated and potentially induce gene amplification [33], [34]. In Figure 1, three independent survivors showed that similar chromosome rearrangement, and this may suggest the existence of hot spot(s) for chromosome rearrangement (Figure 1A, lane 8–10). This site is different from that preferentially re-replicated when relicensing inhibition is abrogated. This difference might be caused by the difference in the order of activation of origins. When untimely initiation in G1 phase is induced, the temporal control of replication origins observed in the normal S phase is likely to be maintained [13], while re-replication in G2/M phase caused by relicensing primarily occurs at a subset of both active and latent origins [34]. Therefore, it would be intriguing to determine whether hotspots for chromosome rearrangement appear when untimely initiation is induced. It would also be interesting to analyze the chromosome context surrounding if such hot spots exist. Moreover, whole chromosome duplication is observed very frequently in survivors (Figure 1). The reason for this finding also should be addressed in a future study to understand the impact of untimely DNA replication on genome stability.
We have shown that the regulation of the protein-level of initiation factors such as Sld2 and Dpb11 is important to prevent untimely initiation in G1 phase, in addition to the previously described context of CDK phosphorylation of Sld2 and Sld3. The high-level expression of Sld2 or Dpb11 alone resulted in a higher GCR rate (Figure 5 and Table 1). This result is the first example indicating that the protein-level of initiation factors directly affects genome stability and further confirms the direct relationship between the G1/S regulatory machinery and genome stability, both of which are frequently deregulated in human cancer cells [31], [35]–[40]. In budding yeast cells, the protein levels of replication factors required for initiation, such as Sld3, Dpb11, Cdc45, Mcm2-7 and GINS, are constant throughout the cell cycle and only the protein level of Sld2 fluctuates ([15], [41]–[43], our unpublished data). Of these replication factors, Sld2, Sld3 and Dpb11 are only required for the initiation step. Although phosphorylation of Sld2 by S-CDK greatly enhances the interaction between Sld2 and Dpb11, in vitro analysis suggests that unphosphorylated Sld2 and Dpb11 can interact, although the interaction is inefficient [15], [16]. Because G1-level Sld2-11D is not sufficient to induce DNA replication but G1/S-level Sld2 is (Figure 2 and Figure 3), cell cycle-regulated Sld2 expression is important not only for the enhancement of initiation in S phase but also for the prevention of unfavorable Sld2-Dpb11 complex formation in G1 phase to inhibit untimely initiation. The concentration of other key initiation factors, Dpb11 and Sld3, is relatively low throughout the cell cycle (Tanaka and Araki, manuscripts in preparation), and this must also be important to prevent untimely initiation because a high level of expression of Dpb11 resulted in a higher GCR rate.
It is known that deletions of or defects in factors that play a role in the DNA damage response pathway elevate GCR in budding yeast [44]–[46]. Dpb11 is known to play a role in the intra S phase and DNA damage checkpoint pathway [29], [47]. Because Sld2 interacts with Dbp11, a high level of Sld2 could disturb Dpb11's checkpoint function and result in increased GCR, although this seems unlikely. In budding yeast, DNA damage is mainly recognized in S and G2 phase, and cells with DNA damage do not show a remarkable delay for S phase entry [48]. If a high level of Sld2 affects checkpoint function, a high level of Sld2 in S to M phase should cause elevated GCR, and a high level of Sld2 in G1 would not affect GCR generation. However, our results showed the opposite effect, which indicates that the higher GCR rate observed in Sld2-expressing cells is not caused by checkpoint failure.
Our results indicate that multiple regulatory mechanisms are employed to prevent untimely initiation in G1. At least three mechanisms, the hypophosphorylated status of essential initiation proteins Sld2 and Sld3 and the low level of Sld2, contribute to prevent untimely activation of replication origins. These mechanisms are independent; therefore, disruption of one regulatory mechanism, for example, the high level of expression of Sld2, is enough to elevate GCR, and the simultaneous deregulation of these mechanisms causes more severe phenotypes. For example, a high level of expression of Sld2-11D (both the protein level and phosphorylation of Sld2 are deregulated) resulted in a very high GCR rate (Table 1 and Figure 5), and the combination of CDC45JET1-1 and sld2-11D (both of the phosphorylations of Sld2 and Sld3 are bypassed) frequently generated chromosome rearrangements and aneuploidy (Figure S5). When all of the regulatory mechanisms are bypassed in CDC45JET1-1 GALp-SLD2-D cells, cells die (Figure S1 and Figure 2A). Although each inhibitory mechanism can mostly block the initiation of DNA replication, each of them alone is not sufficient. Therefore, multiple mechanisms are important to the robustness of the system to prevent untimely initiation and hence for stable genome maintenance over generations.
Although budding yeast is so far the only system in which the untimely initiation of DNA replication can be artificially induced [13], [14], regulation of initiation by combining CDK phosphorylation and dosage control of initiation factors might also be employed in other eukaryotes. The CDK requirement for the initiation of DNA replication is a highly conserved feature in eukaryotes. In vertebrates, TopBP1/Cut5/Mus101 and RecQL4 are thought to be the orthologues of Dpb11 and Sld2, respectively, because of sequence similarities and their roles in DNA replication (reviewed in [3], [49]). In Xenopus egg extracts, both Cut5 and RecQL4 bind chromatin before initiation and interact each other, and the Sld2-homology domain of RecQL4 contains many potential CDK phosphorylation sites [50]–[53], although whether or not phosphorylation of RecQL4 is required for these functions is unclear. Recently, novel factors called Treslin/Ticrr, GemC1 and DUE-B were reported as essential factors for initiation [54]–[57]. Treslin is distantly related to Sld3 [58], and like budding yeast Sld3, phosphorylated Treslin interacts with N-terminal tandem BRCTs of TopBP1 in Xenopus egg extracts [56]. GemC1 also has multiple CDK phosphorylation sites that are important for the initiation of DNA replication and interacts with TopBP1 [54]. Therefore, these proteins are possible functional analogues of Sld3. Of these orthologues/analogues of Sld2, Sld3 and Dpb11, the expression of TopBP1 is under the control of E2F, a G1-S specific transcription factor [59].
Genomic instability is a hallmark of human cancer cells [35], [38]. Based on our results, the deregulation of initiation factors in human cells may be important in the induction of genomic instability and cancer. The retinoblastoma/E2F pathway is known to be deregulated frequently in cancer cells [60], [61]. As described above, TopBP1 is one of its targets, although the expression level of TopBP1 has not been precisely determined. Interestingly, in osteosarcoma, chromosomal rearrangements and genomic imbalances affecting 8q24 in which the RECQL4 gene maps are frequent and the increased expression of RecQL4 are correlated to some type of chromosome instability [62]. Therefore, it is possible that untimely initiation is occurring in these cells by the high-level expression of initiation factors, perhaps triggering genomic instability. In summary, our data in budding yeast provide a good model to understand how untimely initiation is prevented and hence how stable genome maintenance is achieved in eukaryotic cells.
The strains used in this study are listed in Table 3. All strains used in the GCR assay are derived from RDKY3615. All others are derived from W303-1a. Cells were grown in rich medium YPA (1% yeast extract, 2% Bacto-peptone and 40 µg/ml adenine) or Synthetic Complete (SC: 0.67% Yeast Nitrogen Base, supplemented with amino acids) supplemented with 2% sugar (glucose, galactose, raffinose, or sucrose). Cells were arrested in G1 with 30 (if release was required) or 100 ng/ml alpha factor for Δbar1 strains or 10 µg/ml alpha factor for BAR1 (wild type) strains. For the cell cycle block in S and G2/M phase, 200 mM hydroxyurea (HU) and 5 µg/ml nocodazole were added to the medium.
Endogenous Sld2 and Myc-tagged Sld2 were detected with anti-Sld2 polyclonal antibodies [16]. Orc6 was detected with the SB49 monoclonal antibody [63]. FLAG-tagged Orc6 was detected with the M2 monoclonal antibody (Sigma). Myc-tagged Sld2-11D, Db-tag and Sic1-tag were detected with the 9E10 monoclonal antibody. Rad53 was detected with the anti-Rad53 serum [48]. Flow cytometry was performed as described elsewhere [11]. To quantify the increase in DNA content, the average of DNA contents was calculated with the CellQuestPro program (Beckton-Dickinson).
Yeast chromosomes were separated with the CHEF-DRII (Bio-Rad) in a 0.8% agarose gel with 0.5× TBE buffer. Gel images were acquired and analyzed with the LAS-4000 mini and the Multi Gauge software (GE Healthcare).
The GCR assay was performed as previously described [24], [64]. Briefly, in the typical experiment, five independent colonies were grown in the appropriate medium and then spread onto synthetic medium containing canavanine (CAN) and 5-fluoroorotic acid (5FOA). The number of colonies formed on CAN+5FOA plates was counted, and the median was used to calculate the GCR rate. In the case of *2 in Table 1, colonies appeared on less than the half of plates, and the GCR rate was calculated from the highest colony number.
|
10.1371/journal.pcbi.1000299 | Alignment and Prediction of cis-Regulatory Modules Based on a Probabilistic Model of Evolution | Cross-species comparison has emerged as a powerful paradigm for predicting cis-regulatory modules (CRMs) and understanding their evolution. The comparison requires reliable sequence alignment, which remains a challenging task for less conserved noncoding sequences. Furthermore, the existing models of DNA sequence evolution generally do not explicitly treat the special properties of CRM sequences. To address these limitations, we propose a model of CRM evolution that captures different modes of evolution of functional transcription factor binding sites (TFBSs) and the background sequences. A particularly novel aspect of our work is a probabilistic model of gains and losses of TFBSs, a process being recognized as an important part of regulatory sequence evolution. We present a computational framework that uses this model to solve the problems of CRM alignment and prediction. Our alignment method is similar to existing methods of statistical alignment but uses the conserved binding sites to improve alignment. Our CRM prediction method deals with the inherent uncertainties of binding site annotations and sequence alignment in a probabilistic framework. In simulated as well as real data, we demonstrate that our program is able to improve both alignment and prediction of CRM sequences over several state-of-the-art methods. Finally, we used alignments produced by our program to study binding site conservation in genome-wide binding data of key transcription factors in the Drosophila blastoderm, with two intriguing results: (i) the factor-bound sequences are under strong evolutionary constraints even if their neighboring genes are not expressed in the blastoderm and (ii) binding sites in distal bound sequences (relative to transcription start sites) tend to be more conserved than those in proximal regions. Our approach is implemented as software, EMMA (Evolutionary Model-based cis-regulatory Module Analysis), ready to be applied in a broad biological context.
| Comparison of noncoding DNA sequences across species has the potential to significantly improve our understanding of gene regulation and our ability to annotate regulatory regions of the genome. This potential is evident from recent publications analyzing 12 Drosophila genomes for regulatory annotation. However, because noncoding sequences are much less structured than coding sequences, their interspecies comparison presents technical challenges, such as ambiguity about how to align them and how to predict transcription factor binding sites, which are the fundamental units that make up regulatory sequences. This article describes how to build an integrated probabilistic framework that performs alignment and binding site prediction simultaneously, in the process improving the accuracy of both tasks. It defines a stochastic model for the evolution of entire “cis-regulatory modules,” with its highlight being a novel theoretical treatment of the commonly observed loss and gain of binding sites during evolution. This new evolutionary model forms the backbone of newly developed software for the prediction of new cis-regulatory modules, alignment of known modules to elucidate general principles of cis-regulatory evolution, or both. The new software is demonstrated to provide benefits in performance of these two crucial genomics tasks.
| The spatial-temporal expression pattern of a gene is controlled by its regulatory sequences, sometimes called a cis-regulatory module (CRM). A CRM contains a number of transcription factor binding sites (TFBSs), which read the expression level of the cognate transcription factors (TFs) and drive the appropriate expression pattern through the combinatorial interactions among TFs, their co-factors and the basal transcriptional machinery [1]. Cross-species comparison plays a central role in various problems involving cis-regulatory sequences, including computational prediction of CRM sequences [2]–[4], discovery of novel sequence motifs [5],[6] and exploration of the principles of regulatory sequence evolution [7]. For these different types of analysis, the standard procedure almost always starts with an alignment of these sequences, followed by an analysis of the conservation pattern of sequences as suited for the specific task.
The first major limitation of this two-step procedure arises from errors in alignment. It has been shown that alignment procedure may seriously affect the results of comparative genomic analysis such as reconstruction of phylogenetic trees and inference of positive selection [8]. Most alignment tools are not customized to regulatory sequences, and thus cannot take advantage of their specific structural and evolutionary properties. A second shortcoming of many current methods for regulatory sequence comparison is their heuristic nature. It is difficult to assess the significance of the results if appropriate statistical models have not been specified and used. While there are indeed a number of successful programs based on sound statistical models of DNA sequence evolution [9],[10], few of them incorporate the CRM structure. Finally, it is commonly assumed that a TFBS is conserved across all species being studied [5],[11]. However, there is strong evidence that functional noncoding sequences in general, and TFBSs in particular, are not always conserved in an alignable sequence even in relatively close species [12],[13]. This process of TFBS change has been recognized as an important source of evolution of phenotypes [14].
Several approaches have been proposed to address one or more of the problems discussed above. The programs Stubb [3], EvoPromoter [15] and PhylCRM [16] predict CRMs as significant clusters of TFBSs, which are detected by comparing orthologous sequences using an evolutionary model of binding sites. However, all methods require a fixed alignment as input and do not model the binding site gains and losses. The programs CONREAL [17], EEL [18] and SimAnn [19] align putative CRM sequences with the explicit goal of aligning the sites matching known TF profiles. None of these methods use rigorous statistical or evolutionary models, and they all assume the complete conservation of TFBSs. Moses et al. [13] deal with alignment uncertainty in their analysis of binding site turnover, but this is done as a post-hoc analysis rather than being integrated with the inference step. Our recent work, Morph [20], tries to solve all the above problems in a single framework with a pair-HMM model. However, the Morph model does not accurately capture the evolutionary dynamics of CRMs. Lineage specific TFBSs are treated not as gain or loss events in evolutionary time, but merely as HMM “emissions” from one sequence, and not the other. Another recent work, SAPF [21], aims to combine probabilistic model-based alignment with “phylogenetic footprinting”, which refers to the identification of evolutionarily constrained sequences based on their lower substitution rates. However, TFBSs are not explicitly represented in the SAPF model, and the program is not designed to predict targets of specific transcription factor(s). Our goal is not to detect constrained sequences per se, but the target sequences of specific transcription factors, whose binding motifs are known a priori.
Our philosophy of doing cross-species sequence analysis is: firstly, the method should be based on an explicit model of sequence evolution, as expressed in [10] – “the study of biological sequence data should not be divorced from the process that created it”; secondly, the problem should be solved in a single, integrative framework, instead of being split into multiple steps. Specifically, this means that to predict a CRM, one should take into account the uncertainty of alignment and TFBS annotation by summing over them from a combined statistical model. The above philosophy has been adopted previously in the area of statistical alignment [10],[22],[23], where stochastic models are used to describe the evolution of indels and the alignment task is often integrated with the ultimate goals, most notably, the reconstruction of phylogenetic tree [24]. Models that describe one or more aspect of regulatory sequence evolution have been proposed recently [25]–[31], but none of these methods offers a complete evolutionary model of CRM sequences that can be directly used for bioinformatic tasks such as CRM alignment and prediction. We propose an expressive and biologically realistic model of CRM evolution where (i) stochastic models of substitution and indels are used to characterize the evolution of background sequences (non-TFBS sequences inside a CRM); (ii) TFBSs evolve according to a population genetic model developed previously; (iii) functional switching between a non-TFBS and TFBS can occur in a manner dependent on the binding energy of the evolving site. We implement an efficient inference machinery and apply it to the tasks of CRM alignment and prediction.
We used alignments produced by our program to analyze the regulatory sequences involved in early development of Drosphila melanogaster. We took advantage of the recent genome-wide binding data on key TFs involved in blastoderm-stage gene regulation (obtained using ChIP-chip technology [32]), and tested two important hypothesis. First, we investigated previously published claims that there is a high level of non-functional binding in such genome-wide TF binding studies [33],[34]. The prime candidates of such non-functional binding sites are those that are not adjacent to genes expressed in blastoderm. If the claim is true, we expect that these sites will be less conserved than binding sites adjacent to appropriately expressed genes. We found statistical evidence to the contrary, opening up the possibility of functional binding at a larger scale than previously thought. Second, the positions of CRMs relative to the coding sequences, may have a large impact on their functions. For example, computationally predicted CRM sequences enriched with TFBSs are much more likely to drive expression of reporter genes if they are located close to transcription start site (TSS) [2]. Wray has suggested an interesting hypothesis that the CRMs near TSS are likely “control modules”, while those distal ones may be “booster modules” that are less essential [35]. We tested this hypothesis by comparing the conservation level of TFBSs in proximal bound regions and in distal ones. We find no support for this hypothesis, and in fact distal bound regions seem to have a greater conservation of binding sites than proximal regions, contrary to expectation.
In this section, we present the details of our model, which first captures the salient properties of a CRM's content and then lays out the evolutionary forces acting upon its different components. The model prescribes the joint likelihood of a set of orthologous CRMs that are related by a given phylogenetic tree.
We begin with a model of CRM composition and assume that the ancestral CRM is generated from this model. We use a generalized HMM of zero order, similar to the ones used in [3],[15],[36]. The binding specificities (motifs) of TFs are represented by position weight matrices (PWMs), and the nucleotide frequencies of the background sequence are denoted by . At each step, the background state or the motif is sampled with probability and , respectively. If the motif is chosen, the actual site is sampled from the PWM; otherwise, a single nucleotide is sampled from . The HMM transition probability, , can be interpreted as the average number of binding sites of this motif per nucleotide at equilibrium, or simply binding site density.
Our evolutionary model of the background sequences is adapted from the models developed earlier for “statistical alignment” [9],[22],[37]. Substitutions are described by the standard HKY model [38], with equilibrium distribution and transition-transversion bias . Insertions and deletions follow Poisson processes with rates and respectively. The length of an indel follows the geometric distribution with the probability of extension . Following this model, the joint probability of the sequences and in the example below, under a two-species phylogenetic tree with branch lengths and , is approximately (# stands for any nucleotide):(1)where . The terms , and are the probabilities of not seeing an indel event in , of seeing a gap in the second sequence and of seeing a gap in the first sequence, respectively. The derivation of these probabilities can be found in Text S1.
We use the population genetics-based Halpern-Bruno (HB) model for TFBS evolution [39],[40]. This model captures the fact that the evolutionary constraints at different positions of a TFBS may be different: less degenerate positions in the PWM generally have lower substitution rates. Let be the substitution rate matrix of the background sequences, and be the PWM of the motif being evolved, the rate of substitution of a nucleotide to at position is:(2)The transition probability of to in time is thus the entry of the matrix . Since HB model is time reversible, the joint probability: is simply .
Gain and loss of TFBSs are commonly observed across a large evolutionary spectrum: e.g. fungi [41], insects [13] and vertebrates [12]. There are two different scenarios in which these events may occur. In the first scenario, the expression pattern of the target gene is under adaptive change, which “demands” a change in the composition of the controlling CRM, causing binding site gains and losses. In the second scenario, the expression of the gene is under stabilizing selection, but the selection on individual TFBSs may be weak, and as a result, a TFBS may be lost during evolution due to random drift. New TFBSs may also be created in the background sequences simply by mutations and random drift, due to the fact that TFBSs are often short and degenerate. The two processes may be linked to each other: the loss of one TFBS could make gain of a TFBS in the background more beneficial so as to compensate for the loss; likewise the gain of a new site could make existing sites redundant, thus relax the constraints and speed up the loss process.
The main difference between the two scenarios is: in the former, the changes of TFBSs are driven by external selection forces while in the latter, the changes are mainly dominated by the stochastic forces of mutation and random drift, with selection being weak. In our model, we adopt the second scenario as it is a more “parsimonious” explanation of TFBS gain and loss, and is more consistent with our current knowledge about the Drosophila early developmental CRMs [7],[42], which are among the most well-characterized available today.
Our specific model formulates the ideas discussed above. We follow the usual definition of binding energy of a TFBS, for example [28]. We assume that there is a threshold for the binding energy of a site, , above which a site is not functional. We use and to denote the evolutionary models of a TFBS and non-TFBS respectively. Our basic idea for modeling gain and loss is: switching of a site between TFBS and non-TFBS states is a switch between the models that govern the evolution of this site. Under the model , mutations that change the energy of a site above may occasionally be fixed due to random drift. After that point, natural selection will not be able to perceive this site (switch to ). Likewise, under the model , a background site could occasionally reach by mutation and random drift. This site will then be visible to its cognate TF and will be subjected to natural selection (switch to ). We note that indel events may happen inside TFBSs, albeit with a much lower rate than in background sequences, and we denote by the relative rate of intra-TFBS indels. Interaction between gain and loss events as explained above is not explicitly modeled, to avoid creating dependencies that make the computational task much more difficult.
We illustrate our model of TFBS loss in Figure 1: starting with a functional site , a substitution or indel event disrupts this site at time ; the background model then governs the evolution of this site, which eventually becomes sequence . Let and be the sequences preceding and following the loss event respectively, then:(3)where is the instantaneous rate of substitution (given by Eq. 2) or indel (given by the product of and the background indel rate) under the model , the evolutionary model of TFBSs; and must satisfy the energy constraint: and , and the neighborhood constraint: they differ by a single mutation event. The probability of TFBS gain can be calculate in a similar way. Joint probability under a two-species phylogenetic tree can be found in Text S1 and Figure S1. For computational efficiency, we make the parsimony assumption: suppose is an intermediate site between and , then the symbol at any position of is either the symbol of or of at that position. We also note that, even though we rely on a threshold for determining when binding site gain or loss happens, this parameter is not directly used for classifying a site as functional or not. Instead, the annotation of a site depends upon an examination of the site and its orthologous sequences, and their probability under different histories: background, conserved or lineage-specific.
We solve the following computational problems: given two orthologous sequences (that are roughly alignable so that they could be identified in the first place) and a set of TF motifs, (1) align the two sequences and annotate the TFBSs; (2) predict if the sequence is a CRM targeted by the given motifs. We use dynamic programming to simultaneously find the optimal alignment and TFBS annotation. For the second task, we use a likelihood-ratio test of two models: the CRM evolutionary model and the background evolutionary model where no motif is used. Computation under each model is also done by dynamic programming, summing over all possible alignment paths and annotations of TFBSs. The details of the algorithms can be found in Materials and Methods. We allow all parameters to be learned automatically from the data while allowing certain parameters to be specified by users (Text S1). Our computational framework is implemented as a program called EMMA (Evolutionary Model-based cis-regulatory Module Analysis).
Simulation of sequence evolution is a useful strategy for assessing computational methods, as the true evolutionary history is often unknown for real data. In addition, simulation is an important way to understand how various factors, such as divergence time or the presence of TFBSs, affect the performance of a computational procedure, since such questions are generally difficult to answer analytically. We developed a simulation program that can generate orthologous CRM sequences according to our evolutionary model. The binding site densities, branch lengths, indel rates, etc., are all user-specified parameters. Our simulator captures a richer biology of regulatory evolution than many other sequence simulators [43],[44] and can be used as a general tool for the study of cis-regulatory evolution.
We first used simulation to study alignment methods, similar to what was done previously by Pollard et al. [44] and Huang et al. [45]. We implemented several versions of EMMA so that we could study the effect of each of its features. We denote by EMMA0 the version that uses only the background model, but not the motifs (thus EMMA0 is equivalent to the traditional Needleman-Wunsch alignment with affine gap penalty); EMMA1 is the version that considers only conserved TFBSs; and EMMA2 models both conserved and lineage-specific TFBSs. In addition, we tested a widely-used general-purpose alignment tool, Lagan [46], and our recently developed program for aligning regulatory sequences, Morph [20]. Morph uses a pair-HMM to model the alignment of two sequence, where the HMM contains several motif states to encode the presence of TFBSs in one or both sequences. We note that Morph does not model binding site turnover, so a non-conserved site will be aligned with gaps, instead of its true orthologous sequence. All programs were run under the default settings. The alignment performance was measured using specificity and sensitivity, as in [20]. In addition, we defined a measure called “TFBS conservation sensitivity” as the percentage of all positions in conserved TFBSs that are correctly aligned. For CRMs, this is clearly a more relevant measure of alignment quality [44].
The specificities and sensitivities of all programs are similar (see Tables S2 and Table S3 in Text S1) because different programs differ mostly in treating TFBSs, which occupy a small fraction of the total sequence length. The results with the TFBS conservation measure are shown in Figure 2. EMMA0 and Lagan have similar performance with all three measures. This suggests that the values of alignment parameters have relatively small effect on the alignment quality. At moderate to high divergences, both EMMA1 and EMMA2 significantly outperform EMMA0 and Lagan in terms of TFBS conservation sensitivity (e.g. EMMA1 is better than Lagan by 12% and 13% respectively at divergence 0.7 and 0.8) and are slightly better with the other two measures, suggesting that modeling conserved TFBSs is beneficial to alignment of divergent sequences. Modeling lineage-specific TFBSs does not seem to help alignment, as EMMA1 is slightly better than EMMA2 at high divergence levels. This somewhat counter-intuitive observation may be explained by the fact that in pairwise comparison, lineage-specific TFBSs will not help alignment by serving as “anchors”; on the other hand, a truly-conserved TFBS may occasionally be treated as two lineage-specific sites in EMMA2. Morph is superior to Lagan and EMMA0 in terms of aligning conserved TFBSs, but not as good as EMMA1 and EMMA2. In addition, the higher TFBS conservation of Morph is achieved at the cost of significantly lower overall alignment sensitivity (more than 6% lower than all other programs at divergence greater than 0.5, see Tables S2 and S3 in Text S1).
Pollard and colleagues also studied the problem of how alignment affects binding site detection, through simulation [44],[47]. Their recent program, CisEvolver [44], is similar to our simulator in that both treat sequences as a mixture of background and TFBSs, and both use the Halpern-Bruno model for binding site evolution. The main differences include: CisEvolver uses empirical indel frequencies to parameterize the evolution of indels while we use a simpler geometric distribution; and CisEvolver does not incorporate the possible gain and loss of functional sites. Nevertheless, to test the robustness of our conclusions, we generated the test data using CisEvolver and repeated the comparisons described above. The results are in broad agreement with our previous results (see Figure S2, and Tables S4 and S5 in Text S1), suggesting that EMMA is robust to the treatment of indels and that our gain and loss model will perform well even if there are actually no or few gain or loss events (i.e., it will not introduce such events artificially). The PSPE program [45] is also capable of simulating the evolution of CRMs. It goes beyond modeling the gain and loss of individual binding sites, and captures features not available in either our simulator or CisEvolver; for example, it allows “global” fitness constraints such as “the total number of sites in a CRM must fall in some range”. However, its semantics of sequence alignment is different from the conventional notion of alignment (i.e., nucleotide-level orthology), making its benchmarks unsuitable for studying our program's performance. Furthermore, these alignment benchmarks were obtained under the assumption that each TF had one and only one site in a CRM, an assumption that is overly stringent given that many factors are known for homotypic clustering [48]. Also, the key idea of the PSPE model is “replacement turnover”, where one site loss exactly matches a gain in another site and vice versa, but it is not clear if this is a general evolutionary process. One recent study did not find that such change is important in explaining the observed patterns of binding site turnover [13]. Based on these considerations, we chose not to test EMMA on the benchmark data of PSPE.
We investigated the problem of predicting CRM sequences given a set of motifs, again using simulated sequences. We implemented another program called EMMA-ANN, which scores a sequence by its conserved TFBSs using a fixed, non-motif alignment (we used EMMA0 alignment). The only difference between EMMA-ANN and EMMA1 is that a fixed alignment is used in the former, so the comparison between the two should suggest how important it is to treat alignment uncertainty. In addition to the EMMA family programs and Morph, we tested the program Stubb [3], which scores a sequence by its binding site cluster while favoring the conserved sites. Similar to EMMA-ANN, Stubb is also based on a fixed alignment. Here, a site that matches the PWM, but does not fall inside an aligned region, is allowed to contribute to CRM scoring; in contrast, only conserved sites are allowed to make contributions in EMMA-ANN. Thus in terms of identifying binding sites, Stubb is likely to be more sensitive but less accurate, a feature that has implications on CRM prediction. We generated a positive set of sequences simulated under the CRM evolutionary model, and a negative set under the background model. A program is tested by its ability to discriminate positive and negative sequences: all the sequences will be scored by the program; as the score threshold varies, the specificity and sensitivity will be computed. The overall performance of the program is measured by the area under curve (AUC) of the ROC curve, i.e., the plot of sensitivity vs (1 - specificity).
The results are shown in Figure 3. The first observation is that high accuracy of prediction can be achieved even at relatively large divergence. Our explanation is that at higher divergence, conserved TFBSs will be more significant, i.e., less likely to be explained by the chance conservation of neutral sequences. The implication is that CRM prediction is sensitive to the correct alignment of conserved TFBSs, but not to the overall alignment quality. Comparisons between EMMA1 and EMMA-ANN suggest that simultaneous alignment and CRM scoring consistently improve prediction at almost all levels of divergence examined. The effect of modeling lineage-specific TFBSs, as seen from the comparison between EMMA1 and EMMA2, is somewhat mixed: at lower divergence (<0.3), it reduces the performance by 5–8%; at higher divergence, it improves by 3–8%. The intuition is: TFBS gain and loss will become common and thus important for the algorithms only at relatively high level of divergence. EMMA2 is consistently better than both Stubb and Morph which are based on phenomenological models (e.g. at divergence 0.8, AUC of EMMA2 is 0.95, while AUCs of Stubb and Morph are 0.87 and 0.82 respectively), suggesting the importance of correct models. Morph unexpectedly shows poorer performance than all other methods including Stubb. The underlying model of Morph is quite different from the evolutionary model used here, and it is likely that Morph will interpret any weak match to a PWM, even if not conserved, as a TFBS (often a spurious site), thus inflating the scores of all sequences and making discrimination between the two sets more difficult. Overall, the best performance is obtained by EMMA2 at moderate to high divergence levels. This, combined with the fact that in practice the divergence level for pairwise comparison often lies in this range (e.g. human-mouse divergence is estimated to be 0.6–0.8 [49]), justifies the use of our full model, EMMA2, for practical tasks of CRM prediction. We also repeated the same comparisons using data simulated under CisEvolver, and obtained similar conclusions (data not shown).
In summary, we made several findings through simulation that are relevant to choosing and developing the computational tools for CRM prediction. Generally, one should use relatively divergent sequences, as long as they are alignable. The performance depends on the alignment of TFBSs, but less on the overall alignment quality, or the exact alignment parameters such as gap penalties. When the binding site turnover events are common, it is important to model the lineage specific binding sites. Finally, because of the inherent uncertainty of alignment, simultaneous inference and alignment taking advantage of the special properties of the sequences will help both tasks. All these observations support our efforts in building a comprehensive model of CRM evolution and using it as a basis for related inference tasks.
To test if our model truly brings benefits in real-world applications over existing programs, we start with the task of alignment, and compare our program EMMA (its full version, EMMA2) with Lagan and Morph. We study the set of blastoderm CRMs from the RedFly database [44] in D. melanogaster and D. pseudoobscura. Since it is not possible to know the true alignments of the real data, we follow the earlier approach [50] of evaluating an alignment by how often a TFBS appears to be conserved in this alignment: a correct alignment should contain more conserved TFBSs on average than an incorrect alignment. We call a TFBS conserved in an alignment if it appears as a gapless block, and both orthologous sites have binding energy above some threshold ( 0.002, where is defined through a standard likelihood ratio score [13],[51]).
In our first experiment, we use the known TFBSs [52] of seven motifs important in the blastoderm stage of development. Among the total of 188 known sites in 65 CRMs, 80 are conserved in the Lagan alignment, while 91 and 103 are conserved in EMMA and Morph alignments, respectively. We further manually examined some sequences on which the alignment programs disagree and show one such example in Figure 4. Three patterns of possible mis-alignment are revealed in Lagan alignment. For the first Hb site, the orthologous site is shifted by two nucleotides likely because the Hb motif has a repeat structure ( in its consensus sequence). For two Bcd sites in the middle row, the nucleotides at the boundaries are not aligned. In particular for the first one, the gap in D. melanogaster can be moved by one position without changing the Lagan score, suggesting that arbitrary resolution of ambiguous alignments can contribute to small-scale alignment errors that may be important for binding sites. Finally, the last Bcd site in D. melanogaster is close to, but does not align to a potentially orthologous Bcd site in D. pseudoobscura. In EMMA alignment, all four sites in D. melanogaster are aligned with their functional orthologs in D. pseudoobscura.
We next use predicted sites for further evaluation, since the number of known TFBSs is small. For each of the seven motifs, we constructed alignments with Lagan, EMMA and Morph, using only one motif a time. The results were evaluated by the number of predicted sites ( 0.002) that appear conserved in the alignments. The results are shown in Table 1. Similar to what we have found above, the number of conserved sites under EMMA is significantly higher than that under Lagan, for all motifs but Kni and Tll. The performance of Morph is intermediate between EMMA and Lagan.
Though our alignment evaluation is not perfect, all the evidence taken together strongly suggests that by utilizing the knowledge of binding motifs, EMMA can significantly improve detection of TFBSs over general purpose tools by overcoming the alignment problems such as arbitrary gap placement. Morph can also improve TFBS detection, but because lineage-specific TFBSs have to be gap-aligned, Morph results do not capture the true evolutionary history of orthologous sequences.
In this experiment, we tested different programs for predicting regulatory target sequences of a given TF. Each program was made to score test sequences with a single known PWM. We did not follow the previous procedure [3],[15] of classifying CRM and non-CRM sequences based on sets of known motifs, because we believe that our setting will make the task more challenging and thus make it easier to see the differences of various methods. Furthermore, this experimental setting is particularly relevant to the problem of reconstructing transcriptional regulatory networks, since knowing regulatory relations is often the goal, rather than knowing whether a sequence is a CRM per se [53]. In addition to EMMA, Stubb and Morph, we tested Cluster-Buster, a popular CRM finding program [54]. Cluster-Buster uses a HMM to search for binding site clusters in a given sequence and may therefore be used to discover such clusters for individual TFs. Unlike other methods we are testing, Cluster-Buster does not directly use information in orthologous sequences. For each of the seven blastoderm TFs, we constructed a positive set of sequences: those that contain at least one known binding site of this TF in FlyReg; and we used a common set of random noncoding sequences as the negative set. Again, the D. melanogaster-D. pseudoobscura comparison is used for this experiment. Our evaluation is based on, first, the same AUC measure used for synthetic data; and second, the average sensitivity of programs at high specificity levels. The latter measure is more relevant in practice than AUC because the score threshold is typically chosen to reduce false positive rate to a satisfiable level.
EMMA substantially outperforms all other three programs with the AUC measure (Figure 5A). Averaging over seven TFs, the improvements of EMMA over Cluster-Buster, Stubb, Morph are 9%, 9% and 17% respectively. Measured by the average sensitivity corresponding to the specificity levels above 80% (Figure 5B), the improvements of EMMA are even more convincing: 15%, 21%, 42%, over the three programs respectively. These results support the key ideas of EMMA: dealing with uncertainty of alignment and explicit modeling TFBS evolution will greatly assist the prediction of regulatory sequences. Interestingly, even though Cluster-Buster uses only sequences in D. melanogaster, it is comparable to or even better in some cases than Stubb and Morph, which are based on somewhat similar HMM models and use extra information in the orthologous sequences. Since unlike Cluster-Buster, neither Stubb nor Morph applies a threshold for determining a TFBS, it is likely that they are more sensitive to false positive sites. The problem seems particular serious for Morph because Morph allows a site to be emitted from only one sequence and thus may be overly tolerant to lineage-specific sites matching a PWM (also likely false positive sites). We also note that the experimental setting in this paper is different from the one in [20], where multiple motifs are used simultaneously to classify a sequence. In that setting where there is more motif information and the relative importance of conservation may be reduced, the ability of Morph to score non-conserved weak sites may become an advantage.
One common procedure to enhance the performance of a program running on single-species data, such as Cluster-Buster, is to filter out the sequences that are not very conserved before running the program. We combined this conservation filtering (percent identity greater than 70%, other values of threshold gave similar or worse results) with Cluster-Buster. However, the new results are only slightly better than the original Cluster-Buster, and still lag far behind EMMA (data not shown). This can be probably explained by the fact that a large fraction of Drosophila genome is under constraint [55],[56],[57], thus simple conservation measure is not very discriminative of CRM sequences. To test if our results are robust to PWMs, we also repeated the same experiment with PWMs of the same TFs obtained from bacterial one-hybrid experiments [58] and found similar trends (Figure S3).
In this experiment, we used EMMA to study the evolutionary pattern of TFBSs in sequences involved in gene regulation in blastoderm-stage development of Drosophila melanogaster. Such analysis depends on the accurate alignment of TFBSs, a task that EMMA has been shown to perform better than general purpose sequence alignment tools. We took the sequences bound by each transcription factor (except Gt and Kni, see Materials and Methods), as per ChIP-chip assays in Li et al. [32]. As a “negative control”, we took the intronic sequences that were not bound by the corresponding TFs. These control sequences are presumably neutral or close to neutral [32]. Each “bound sequence” was associated with its nearest gene. We grouped sequences based on whether their associated genes are expressed in the blastoderm or not. (The expression information was obtained from Berkeley Drosophila Genome Project (BDGP) [59]). The two groups were compared based on the level of conservation of predicted binding sites, defined as the percentage of binding sites in D. melanogaster that are conserved in D. pseudoobscura. We expect that the sequences in the expressed group are more conserved than those in the non-expressed group, because binding in the latter group is much more likely to be non-functional. Contrary to our expectation, the non-expressed group appears to be have slightly more binding site conservation than the expressed group (Figure 6A), though the difference is not significant (data not shown). Compared with the control sequences, sequences in both groups have much greater binding site conservation, suggesting functional constraint. We next compared the bound sequences that are proximal to TSS (defined as less than 2 kb distant) and those that are distal (defined as greater than 10 kb). Our expectation is that the proximal sequences overall are more functionally important than the distal sequences, as suggested by others [2],[35], and have more conserved binding sites. The results, however, show the opposite pattern (Figure 6B): binding sites in distal sequences tend to be more conserved than those in the proximal ones, and the differences are statistically significant ( value <10−4 for Bcd, Cad, Hb, and <0.005 for Kr, by hypergeometric test).
We have proposed an integrative framework for cross-species analysis of cis-regulatory sequences. At the heart of our approach is a probabilistic model covering important aspects of CRM evolution, including substitutions and indels in background sequences, and constraints and turnover of TFBSs. The dynamic programming algorithm allows us to efficiently carry out likelihood-based statistical inference. This framework solves the problems of the existing approaches discussed earlier. It aligns regulatory sequences by taking advantage of the tendency of conservation of TFBSs. The TFBS gain and loss model allows us to use information present in lineage-specific TFBSs. Most importantly, when used for predicting CRMs, our method treats alignment and annotation of TFBSs as random variables, summing over them and thus minimizing the impact of an uncertain alignment and TFBS annotation. Our previous programs Stubb and Morph have similar aims, but as shown in our experiments, EMMA significantly outperforms both, strongly suggesting that correct evolutionary modeling is essential to fully utilize the sequence information.
Our model is related to existing models of regulatory sequence evolution, but different from them in several key aspects. Our idea of generation of a new binding site is similar to [27],[31], but their work is limited to simulation studies. Lassig and colleagues [25],[28] have developed population genetic models where a binding site evolves under a fitness function that depends on the edit distance (to the consensus site) or the energy of the site. Their models are the most detailed and perhaps realistic existing models of binding site evolution; however, they cannot be easily used for computational inference since likelihood computation under these models is very expensive (see below). Mustonen and Lassig [28] also proposed ways to model the gain and loss events of TFBSs, but their model is different from ours in that these events are caused by external selection forces, whose rates of occurrences are independent of the actual sequences. A similar model of TFBS turnover has been used to discover lineage-specific TFBSs [30], where the gain and loss of binding sites are modeled by a two-state Markov chain, similar to the Jukes-Cantor model of nucleotide evolution. Again, the rates of change between functional and neutral sites are external parameters that do not depend on the sequences themselves. Durrett and Schmidt [26] studied binding site evolution from the perspective of time needed for a specific word to appear and be fixed in a population, according to population genetic models of mutation and drift. Their study assumes neutral evolution and points out that selective forces will take over if the specific word thus evolved is close to being a binding site; this is the view we have adopted in modeling binding site gain. Recently, Raijman et al. [29] developed a model of CRM evolution, based on the idea that any mutation that creates a new TFBS or destroys an existing one is penalized, i.e., fixed with a smaller probability. Their representation of TFBSs is based on the consensus sequence, instead of the more realistic PWM. Their treatment of TFBS gain is also different from ours: the possibility of TFBS gain from adaptive selection [13] is missing in their model, where all occurrences of new TFBSs will be selected against. Finally, we note that none of the above models integrates the binding site evolution model with the model of insertions and deletions, a feature that is essential to simultaneous alignment and regulatory sequence inference.
Our model is also an extension of statistical alignment (reviewed in [22]) to the analysis of cis-regulatory sequences. Our method shares key features with statistical alignment: explicit modeling of indel evolution; and a probabilistic treatment of alignment uncertainty. Statistical alignment started with the pioneering work of Thorne et al. on pairwise sequence alignment [10], commonly named TKF91 model, where insertions and deletions were treated as single nucleotide events. It was later extended to more realistic indel models, where the indels were treated as multi-nucleotide blocks that followed a geometric length distribution, emulating the commonly used affine gap penalty [23],[37],[60], or an arbitrary length distribution estimated empirically [61]. In other work, the TFK91 model has been applied to multiple alignment, and an MCMC approach developed to sample alignment from a phylogenetic tree [62]. To make the evolutionary model more realistic, some researchers have attempted to capture the heterogeneity of substitution and indel rates and used it to infer slowly-evolving DNA sequences [21],[23],[63]. More recently, the “transducer” model has provided a computational framework for multiple alignment, using TKF91 and other indel models [64],[65]. Our work, especially its alignment functionality, belonging to the category of statistical alignment; however, it is designed specifically for the alignment of cis-regulatory modules. Thus the modeling of substitution and indels, the characteristic feature of statistical alignment, has to be integrated with a model-based treatment of binding site evolution.
One main limitation of our model is that under Halpern-Bruno model, the nucleotides of a TFBS evolve independently while in reality, the TFBS as a whole should be a unit for natural selection [28]. Also, our model of TFBS gain and loss does not parameterize the fitness function of a TFBS, which will be required for correct modeling based on principles of population genetics[28],[39]. So our model can be viewed only as an approximation. Our model choice was based on: (i) avoidance of additional free parameters, which will be difficult to estimate given only an individual CRM sequence; (ii) computational complexity, since modeling a TFBS as a unit is very expensive [28],[29]. One consequence of our simplifications is: any new site created by evolution of background sequences will be selected afterward. A better model should reflect the variability of the rate of TFBS gain in different CRM sequences. Despite these simplifications, we found through simulation that the gain and loss rates under our model with a realistic parameter setting agreed broadly with the empirically estimated values in Drosophila [13],[66] (data not shown).
The relationship between TF binding and target gene expression is an important, but not straightforward, issue. Earlier studies suggested a high level of non-functional binding in ChIP-chip experiments. Gao et al. estimated that more than 40% TF binding are not functional, based on the correlation of binding and mRNA expression [33]. More recently, Hu et al. found that only a small percentage of genes whose promoters bind to some TF changed expression level when that TF was knocked out in yeast [34]. Our analysis based on binding site conservation provides a new way of studying binding-expression relationship. We find that sequences whose associated genes are not expressed, and thus most likely non-functional, are at least as conserved as the sequences close to expressed genes. This suggests that the extent of non-functional binding may be very low, at least when we restrict ourselves to strong binding events (1% FDR). This immediately raises the following question: if strong binding sites near non-expressed genes are indeed functional (as their evolutionary conservation would reveal), what is this function? We speculate several answers. The function of these sites may be to control expression of more distant genes. (Recall that we annotated only the nearest gene as being the target of each site.) Alternatively, these sites may not directly activate or inactive expression, rather, they help attract TF molecules to DNA, and thus help direct the TF molecules to their true target sequences. Another possibility is that these sites function in regulating the nearby gene in a different developmental stage (i.e., not in blastoderm).
Very little is known about the difference between proximal and distal regulatory sequences. It is likely that the two types of sequences work through different mechanisms (for example, the distal sequences may need specific mechanisms such as DNA looping, to communicate with the core promoter sequences of the target genes [67]) and that they play different functional roles, as hypothesized by [35]. Our results suggest that binding sites are more conserved in the distal regions than in proximal regions. One possible explanation is that the proximal sequences are under more adaptive selection than distal sequences, perhaps because it is easier to achieve a different expression pattern by changing the binding sites in the proximal sequences. This increased adaptive selection has been demonstrated in Drosophila in 5′ UTR sequences [55]. Another possibility is that because it is more difficult for distal regulatory sequences to target the promoters, they will be more sensitive to minor changes of binding sites, and thus will be more evolutionarily constrained.
We believe that our proposed framework opens up possibilities for a few major applications. The immediate task is to extend the current work to comparison of more than two species. In pairwise comparison, a TFBS is either conserved or not and it is difficult to distinguish a non-conserved but functional TFBS from a spurious site. In the case of multi-species comparison, there is a wide spectrum of partial conservation, which could be effectively used by a program, as shown in earlier studies [68]. We therefore anticipate that our improved evolutionary model and methods will make a crucial difference to the accuracy of multi-species analysis. Our method takes a set of TF motifs as input; however, which TFs may cooperate while binding is often unknown. Our framework itself offers a way of learning such regulatory rules: the probability of sequences under different TF combinations could suggest how well a particular combination explains the data. Finally, it is possible to learn motifs de novo by treating PWMs as unknown parameters. This approach to motif finding will introduce several benefits over existing programs, e.g., PhyloGibbs [6], such as correcting the alignment errors and using information in partially-conserved TFBSs.
We use to denote the background evolutionary model (both substitutions and indels), and for the evolutionary model for binding sites of the TF (HB model for substitution and reduced indel rate ). The joint probability of the orthologous sites and under a model (background or TFBS) is represented as: , where , are branch lengths of the 2 sequences. In the case of TFBS gain or loss, the probability of a functional site of TF being present in the first sequence but not in the second one is denoted by ; similarly we use for the opposite case.
For a pair of sequences and , we wish to compute the joint probability of the two sequences under the CRM evolutionary model, given the parameters. We define the recurrence variables for dynamic programming as: , the probability of sub-sequences and where the last site is either TFBS (if ) or background (if ) with the “state variable” as explained below and in Figure 7. We then define:(4)(5)Then the probability of the sequences is where and are their respective lengths. In the first case in Figure 7, the last site of the sequence and is a matched background column:(6)where . In the second case, the last column is a gap in the second sequence. If the previous column is also a gap, this should be treated as extension of an existing indel, otherwise as a new indel:(7)The third case, , is handled similarly. In the fourth case, the last sites are a conserved pair of TFBSs of the motif, whose length is :(8)In the fifth case, the last site is a TFBS in but a non-site in . Note that in this case, the length of the non-conserved site may not be , since there could be insertion or deletion in non-TFBS. We will denote it as . We use to denote the probability that the site in is TFBS, but the site in is background with length , then:(9)To make the computation tractable, we will limit to the range of for some user-specified parameter Thus, for the new recurrence variable, we have:(10)The treatment of the last case is similar. One complication is: if the orthologous sites have a gap at the beginning, it may be an extension of an existing indel. In other words, we may have multiplied the probability of an indel event twice: one during the computation of TFBS switching (the second term in Eq. 10), and the other during the computation of the earlier sequences (the first term in Eq. 10). We correct for this as described in Text S1.
For alignment, we simply need to replace the sum operator in the algorithm with the max operator, as is standard in computations involving HMM. The parameters can be estimated by the standard maximum-likelihood approach. In practice, we estimate or fix some parameters, such as transition-transversion bias in HKY model, through external data. The details can be found in Text S1.
For our simulation, we first sampled ancestral sequences from the CRM model described in the main text and evolved the sequences in two branches independently for specified lengths of the two branches. Only the two descendant sequences will be used for alignment input. We used PWMs of Drosophila TFs from the webpage maintained by Dan Pollard (http://rana.lbl.gov/̃dan/matrices.html), which are based on footprinted binding sites [52]. For the alignment experiment, we used Bcd, Kr and Hb with densities equal to 0.008, 0.009 and 0.005 respectively. The motif thresholds were chosen so that the expected rate of TFBS gain equals to the expected rate of loss through simulating evolution of individual TFBS. For the CRM discrimination experiment using simulated data, we used only Bcd and Hb motifs with lower densities 0.004 for both. We generated 50 pairs of sequences with ancestral sequence length equal to 500 bp at each divergence time for the alignment experiment; and 100 pairs of positive and negative sequences with the same ancestral sequence length at each divergence time for the CRM discrimination experiment. A total of 8 divergence time from 0.1 to 0.8 were sampled. For both experiments, the other parameters took values estimated from earlier studies involving Drosophila genomes. The distribution of nucleotides in the background sequences was 0.3, 0.2, 0.2 and 0.3 for A, C, G, T respectively [13]. And from the same study, the transition-transversion bias was 2.0. For the ratio of indels vs substitutions, we used the value (0.225 ) estimated from two close Drosophila species, sechellia and simulans [61], which was evenly split between insertions and deletion in simulation. The length of indels followed geometric distribution with the probability of adding one more nucleotide equal to 0.87 [56]. The rate of indel within TFBS relative to the rate within background sequence was 0.25, from manually inspecting the alignment of eve-stripe 2 CRM in [7].
For the simulation under CisEvolver, we used the same parameters except that the indel lengths are specified by their empirical frequencies [61] instead of approximation by the geometric distribution.
We took 67 D. melanogaster blastoderm CRMs from RedFly database [69], and extracted their orthologous sequences in D. pseudoobscura by using the LiftOver tool from UCSC Genome Browser [70]. Two CRMs without D. pseudoobscura orthologs were discarded. We used seven TFs important for early development in our analysis: Bcd, Cad, Gt, Hb, Kr, Kni and Tll and the PWMs of these TFs were taken from the same source we used for simulation. Bona fide binding sites were collected from FlyReg [52], after some preprocessing: the sites in FlyReg frequently contain some sequence flanking the true binding sites, so we scanned each FlyReg site with the corresponding PWM and extracted the best match to be used as the “known” binding site for evaluation purposes, rather than the original FlyReg site.
In running EMMA, we set the evolutionary parameters according to the values estimated from earlier studies: divergence between the two species is about 1.5 measured by synonymous substitution (http://rana.lbl.gov/̃dan/trees.html), since the noncoding sequences in general are under a high level of constraint (on average, the intergenic sequences evolve 50–60% more slowly than neutral ones [57]), we rescale the divergence to be 1.5*(1−0.6) = 0.6; the indel parameters, the transition-transversion bias and the equilibrium distribution of nucleotides are all set by the values used in simulation. All other parameters are either default or estimated from data by the program itself. Both Stubb and Morph were run under the default parameters except that the divergence was set at the same value.
For each of the seven TFs, we took all RedFly CRM sequences that contain at least one FlyReg site of this TF as the set of positive sequences. Each positive sequence was expanded or truncated so that the length was 1000 bp. 500 random sequences of length 1000 bp each were chosen randomly from the D. melanogaster genome as the negative set. The orthologous sequences in D. pseudoobscura genome were extracted similarly using the LiftOver tool. EMMA was run under the same parameter setting as in the alignment experiment. Stubb was run under the same divergence value, 0.6 and Morph used the automatically estimated value of divergence (similar performance was obtained if using 0.6 as the divergence). Cluster-Buster was run under the default setting, as we do not have extra data for training parameters of Cluster-Buster.
The genome-wide binding data is taken from [32]. We only looked at four factors in this experiment: Bcd, Cad, Hb and Kr. Gt is ignored in this analysis because the PWM of its binding motif is not very specific, and Kni is also ignored because only 35 peaks are identified at 1% FDR level. A bound region is defined as a peak plus 250 bp flanking sequences both upstream and downstream. For the control sequences, we used an equal number of non-first introns, randomly chosen from the D. melanogaster genome. The alignment of the sequences with their orthologs in D. pseudoobscura were constructed using EMMA. The binding sites in both species were then predicted following the same procedure we used before for the experiment of evaluating alignment performance of EMMA. Similarly, we define a binding site as being conserved if both orthologous sites have scores greater than the threshold ( value 0.001). To define expressed and unexpressed group, we used the annotations in BDGP of the expression patterns of genes measured by in situ hybridization (http://www.fruitfly.org/cgi-bin/ex/insitu.pl). A sequence belongs to the expressed group, if its associated gene is classfied as being expressed in stage 4–6 according to BDGP, and similarly for the unexpressed group (using the term blastoderm instead of stage 4–6 will give similar results, but we want to be more conservative when defining the unexpressed group).
The EMMA program, an evolution simulator, and the dataset used in this paper are all available at http://veda.cs.uiuc.edu/emma/.
|
10.1371/journal.pcbi.1002265 | Speed, Sensitivity, and Bistability in Auto-activating Signaling Circuits | Cells employ a myriad of signaling circuits to detect environmental signals and drive specific gene expression responses. A common motif in these circuits is inducible auto-activation: a transcription factor that activates its own transcription upon activation by a ligand or by post-transcriptional modification. Examples range from the two-component signaling systems in bacteria and plants to the genetic circuits of animal viruses such as HIV. We here present a theoretical study of such circuits, based on analytical calculations, numerical computations, and simulation. Our results reveal several surprising characteristics. They show that auto-activation can drastically enhance the sensitivity of the circuit's response to input signals: even without molecular cooperativity, an ultra-sensitive threshold response can be obtained. However, the increased sensitivity comes at a cost: auto-activation tends to severely slow down the speed of induction, a stochastic effect that was strongly underestimated by earlier deterministic models. This slow-induction effect again requires no molecular cooperativity and is intimately related to the bimodality recently observed in non-cooperative auto-activation circuits. These phenomena pose strong constraints on the use of auto-activation in signaling networks. To achieve both a high sensitivity and a rapid induction, an inducible auto-activation circuit is predicted to acquire low cooperativity and low fold-induction. Examples from Escherichia coli's two-component signaling systems support these predictions.
| Different times call for different measures. Therefore, cells adjust their protein levels depending on their environment. Upon the detection of certain environmental signals, transcription factors are activated, which activate or inhibit the production of specific sets of proteins. As it turns out, these transcription factors often also stimulate their own production. Indeed, such self-regulation is a common motif in signal–response systems of many organisms, including bacteria, animals, plants and viruses–but its function is not well understood. We have used mathematical models to study its benefits and drawbacks. On the one hand, calculations show that self-regulation can be a very useful tool if the cell needs to respond in a sensitive way to changes in its environment, or if it is supposed to respond only if the signal exceeds a threshold level. On the other hand, these benefits come at a cost: self-regulation severely slows down the cell's response to changes in the environment. We have analyzed how the cell can benefit from the advantages of self-regulation, while mitigating the drawbacks. This leads to strict design constraints that examples from the bacterium E. coli indeed seem to obey.
| Biological organisms employ a variety of signaling networks to respond to changes in environmental conditions. An interesting class of examples is given by the two-component signaling (TCS) systems, which are ubiquitous in bacteria and plants [1]. TCS systems typically consist of two proteins: a sensor histidine kinase (HK) and a response regulator (RR). The HK is a transmembrane protein that auto-phosphorylates in response to a “signal”. The phosphate group of the HK is subsequently transferred to the RR, which in its phosphorylated form usually acts as a transcription factor. As a result, the RR activates its target genes only when the signal is present. TCS systems are the predominant signaling motifs in bacteria; E. coli, for instance, features about 30 TCS systems [1]. Interestingly, in about half of the cases, the RR also activates its own expression. The functions of this positive feedback are not well understood [2], [3].
Fig. 1 illustrates the transcriptional circuit of TCS systems. It consists of a transcription factor (the RR) that has to be modified post-transcriptionally in order to regulate its target genes; in addition, it may activate its own transcription. Gene networks of this type do not only occur in TCS systems, but are in fact a common motif in many organisms, including eubacteria, archaea, eukaryotes, and viruses [2], [4]. While in TCS systems the RR is modified by phosphorylation, many other transcription factors (TFs) are activated by other covalent modifications or by the binding of a ligand. Here, we use mathematical models to study the characteristics of such inducible auto-activation circuits.
Intuitively, the auto-activation and open-loop circuits each possess their distinct advantages [2], [5], [6]. In the open-loop circuit, the TF is expressed constitutively. As a result, the circuit can be induced quickly, because the post-transcriptional processes that activate the TF are rapid, typically occurring in seconds or less. In contrast, the full induction of the auto-activation circuit involves transcription and translation of the TF, which takes minutes [7]. On the other hand, in the open-loop circuit, the TF is produced even if the signal is absent for a long time. The constitutive presence of numerous TFs in high copy numbers could lead to cross-talk or noise, for instance due to spontaneous phosphorylation. These problems are alleviated in the auto-activation circuit, in which the TF level is reduced in the absence of the signal. In addition, positive feedback is generally expected to increase the sensitivity of the response. However, auto-activation can also lead to bistability and hysteresis [8]. While in some circuits bistability can perhaps be beneficial, in signaling circuits that are supposed to provide a well-defined output to a given input level, bistability and strong hysteresis should presumably be avoided. The significance of each of these effects clearly depends on the parameters of the circuits. Below, we examine the above effects using quantitative models and determine which parameter range could combine the benefits of auto-activation while minimizing its drawbacks.
Our results show several surprises. First of all, they demonstrate that an inducible auto-activation circuit can generate an ultra-sensitive threshold response, even if the activation mechanism is non-cooperative. This is surprising, because in open-loop systems sensitivity is associated with molecular cooperativity, either through the cooperative binding of TFs to multiple binding sites at the promoter, or by cooperativity in the activation of TFs. These new results emphasize that auto-activation is an excellent tool for signaling circuits that require a threshold or switch-like response. However, this benefit comes at a cost: stochastic models reveal that the induction speed is strongly affected by auto-activation–in fact, much more so than previously estimated based on deterministic rate equations [2], [3], [7], [9]. The discrepancy between deterministic and stochastic models is most pronounced when the basal transcription rate of the TF is low, in which case rate equations dramatically underestimate the induction time. Moreover, in this regime the induction time also becomes very unpredictable. We show that these effects are expected to occur under conditions that are fairly typical for bacterial circuits. These novel findings demonstrate that the need for a rapid and reliable induction severely constrains the use of auto-activation in response circuits.
Below, we first introduce the model used in this study. Next, we discuss results regarding bistability, sensitivity, and induction speed. We finally combine these results to explore how these characteristics may restrict the designs actually found in nature.
We consider the inducible circuits illustrated in Fig. 1, consisting of a TF that must be activated to function and possibly activates the transcription of its own gene. To keep the analysis general, we do not specify the nature of the modification, nor the environmental signal triggering the TF's activation. Instead, we assume that in steady state, at a given signal level, a fraction of the TFs will be activated. Thus, can be considered the input of the circuit. If , the signal is completely absent, while if the signal is saturating.
We use a simple, deterministic model to derive our first results [10]. The dynamics of the TF concentration are described by the following ordinary differential equation:(1)Here is the degradation rate constant of the TF; in growing cells, also accounts for dilution due to growth. The TF's transcription rate is a function of the concentration of modified TFs, , because only the modified TFs can activate transcription. We assume has the following Hill-type form:(2)Parameter is the dissociation constant of the modified TF binding to its operators, and is the Hill coefficient. In this notation, the basal transcription rate is , and the maximal transcription rate at full activation is , showing that is the maximal fold change of the promoter. If , the auto-regulation is eliminated and the model describes the open-loop circuit. Lastly, we assume that each mRNA transcribed from the promoter is instantly translated times (the “burst size”) [11]. This results in an increase in the TF concentration by an amount , where is the volume of the cell. Note that for simplicity we do not explicitly include the dynamics of the mRNAs and we neglect time delays due to transcription, translation and protein folding.
For a given input , the dynamics of Eq. 1 define a steady state TF concentration that is at most . The function therefore describes the response of the total TF concentration to the signal . The expression level of genes encoded in the same operon as the TF, such as tg2 in Fig. 1, is expected to be proportional to , too.
Another important quantity is the steady state concentration of modified TF, . Because only the modified TF regulates the target genes, we consider to be the output of the circuit; we will call the response function of the circuit. The shape of the response function is determined only by parameters and (see Supporting Text S1). We therefore focus on the role of these parameters.
For our analysis of the induction speed, a stochastic version of the above model is required; it will be introduced below, in the section “Achieving rapid induction”.
It is well known that auto-activation can lead to bistability [4], [12]–[15]. When an auto-activating circuit is bistable, it has two stable steady states: one in which the TF concentration is high and stays high because the TF activates the transcription of its own gene, and one in which the TF concentration is low and stays low because the TF's gene is not activated.
Fig. 2 summarizes which parameter values yield bistability in our model. As long as the fold change and the Hill coefficient of auto-activation are low, bistability cannot occur (regime I in the diagram). More precisely, bistability does not occur as long as (see Supporting Text S1)(3)( is the solid, blue line in Fig. 2.) In this regime, for any input , the circuit has a unique steady-state TF level (see the insets in Fig. 2). When exceeds the critical value , bistability sets in, but only for an intermediate range of values. For high values of and , this regime extends all the way up to (regime III, to the right of the dotted line). In this regime, the circuit is still bistable even when the signal is saturating.
Regime III seems inappropriate for a signaling circuit. Even at saturating signal levels (), the circuit will not be induced, because the low-expression state remains stable. (In stochastic models the system will eventually turn on by a random fluctuation, but the induction will be very slow, as demonstrated below.) In regime II the circuit is bistable for a range of values. There, the output is not uniquely determined by the input, because two output levels are compatible with a given . This behavior leads to hysteresis [16]. Bistability and hysteresis can presumably be beneficial in some systems [2], in particular if the circuit has to be cautious in turning on or off, or in the context of bet hedging [17]. However, in signaling systems, assuming that there is an optimal expression level for any given signal level, bistability and hysteresis will tend to trap the circuits in a non-optimal state. We argue therefore that a bistable response is often not desired. Based on these considerations, we expect that the circuit parameters should usually be in regime I (also see the discussion in Ref. [18]).
An advantage of positive feedback is that it can increase the sensitivity of a circuit [5], [6]. In the context of dose-response curves, a high sensitivity can be beneficial. In particular, highly sensitive signaling circuits allow the cell to ignore low-level signals, below a certain threshold value, which may be due to noise or cross-talk. We here quantify how much the sensitivity of signal-response circuits can benefit from auto-activation.
The sensitivity of response curves can be defined in various ways [19]. A common approach is to define the sensitivity of a response function as the maximum slope of that function in a log–log plot, that is, as . This definition has many desirable properties: a high indeed indicates that increases rapidly in some domain, the measure is invariant under scaling (), and convenient for mathematical analysis. For those reasons, we will use this measure below. However, a pitfall is that, in order for to be deemed sensitive, a high log–log slope in a single point is sufficient. As a result, a high does not guarantee that the circuit behavior resembles a binary switch [19]. Therefore, we also report results for the measure , defined as the average slope of in a log–log plot, calculated over the domain in which it switches from low to high. (In other words: in the switching domain.) The switching domain is defined heuristically as the domain in which increases from 10% to 90% of its maximum value.
Ultimately, the most relevant quantity is the sensitivity with which the expression of the target genes responds to changes in the signal level. This sensitivity is shaped by each step in the response network: the detection of the signal, the signal transduction, the modification of the TF, and the promoters of the target genes. Consequently, a sensitive response can be implemented at different places in the response network. Here we study the sensitivity that is contributed by the auto-activation circuit and therefore focus on the sensitivity of .
That auto-activation can strongly improve the sensitivity can be understood by revisiting Fig. 2. If the parameters are chosen near the border between regions I and II, the response is almost bistable, leading to a high log–log slope (see inset in Fig. 2). Indeed, for the model in Eq. 1 and 2, can be calculated exactly (see derivation in Supporting Text S1):(4)( was defined in Eq. 3.) This confirms that the maximal log–log derivative diverges when approaches the critical value . We conclude that an arbitrarily high can be obtained for any Hill coefficient by properly choosing the fold change, and vice versa.
In particular, if the auto-activation is non-cooperative (), Eq. 4 reduces to(5)proving that, even in the absence of molecular cooperativity, increases without bound when is increased. Indeed, in the limit of large a strict threshold response is obtained. We illustrate this in Fig. 3A and B. Assuming is large, the expression of the TF, , is fully inhibited when is below the threshold ; for it takes the form of a translated (shifted) Michaelis–Menten curve (Fig. 3A). In the same limit of high , the concentration of activated TF is threshold-linear (Fig. 3B). (Mathematical derivations are provided in the Supplementary Text S1.)
Fig. 3C graphically explains the origin of this behavior. It plots the rate of production of TFs, , as a function of the TF concentration, for four signal levels; the degradation rate is shown too. The steady state value of is the value at which production and degradation balance, i.e., where the production and degradation lines cross (indicated with dots). If is large, the steady state level is large. But if is reduced enough, the production line shifts below the degradation line and the steady state becomes . The threshold value is and can be varied biochemically by tuning .
To illustrate how remarkable the values of are that can be achieved using auto-activation, we reexamine the open-loop case (see Fig. 1). The open-loop circuit itself is insensitive (), but this can be compensated by choosing a sensitive promoter for the target genes [20]. This requires that the TF binds cooperatively to multiple binding sites on the target promoters. If the TF binds fully cooperatively to binding sites and achieves a fold change (again assuming the Hill form of Eq. 2), the expression of the target gene responds with the following sensitivity to changes in :(6)(We provide a derivation in Supplementary Text S1.) Because , this shows that in open-loop circuits the sensitivity at best equals the number of cooperative binding sites at the promoter. However, this maximal is obtained only if the fold change is large, so that . The following numerical example illustrates this point. If the fold change in the non-cooperative auto-activation circuit is , the resulting is 5.5. To obtain the same value in the open-loop circuit, at , more than 7 cooperative TF binding sites are required at the target promoter. Clearly, if a threshold response is desired, auto-activation can be an excellent tool.
We repeat, however, that even though for the non-cooperative auto-activation circuit diverges in the limit of large , the response function does not converge to a step function (as Hill functions do in the limit of high ), but to the threshold-linear response in Fig. 3B. While this is obviously an excellent threshold response, its quality as a switch is better represented by the measure . Unlike , does not diverge in the large- limit, but it nevertheless acquires large values. For instance, in the example above (, , and ), we find . For comparison, for Hill functions,(7)where converges to 1 from below as the fold change increases (see derivation in Supplementary Text S1). From this it follows that, to obtain the value in the open-loop circuit (again assuming ), the promoter of the target gene should have a Hill coefficient . This shows that non-cooperative auto-activation circuits can constitute an excellent switch.
Fig. 3D summarizes the regions of parameter space that lead to a high sensitivity. The figure is based on Eq. 4 for . The line where diverges is shown–the same line that marks the boundary between mono- and bistable regimes in Fig. 2. If the circuit is to be sensitive, the parameters have to be close to that line, within the shaded region (where we chose a somewhat arbitrary cutoff ). An analogous figure based on is presented in the Supplementary Text S1 and leads to very similar conclusions.
Another important characteristic of a signaling circuit is its induction speed. A circuit that can be induced rapidly in a changing environment is expected to have a fitness advantage [7], [9]. We therefore study the induction time of the circuit.
As we explain below, the deterministic model is not adequate to describe the induction time of the circuit. We therefore introduce a stochastic model. This model is based on the following Master equation, describing the evolution of the probability distribution for the TF copy number at time :(8)All parameters are analogous to the deterministic model. As in the deterministic case, the mRNA level is not included explicitly but TFs are produced in bursts of size . We do not model the binding and unbinding of the TF to the DNA explicitly, but assume that the binding kinetics of the TF at the promoter are fast compared to the time scale of transcription initiation [21]. These assumptions simplify the analysis but are not critical: similar results can be obtained with other models previously presented in the literature [21]–[27].
To introduce the induction time, we imagine that the signal has been absent () for a period long enough to ensure that the circuit is in steady state. Then, at time , the signal is introduced at a saturating level (). We then define the induction time as the waiting time before the expression of the TF arrives at 50% of its steady state level. In the deterministic model, this can be calculated by solving the differential equation 1. In the stochastic model, the waiting time is actually a random variable; therefore, we report the mean waiting time, which can be calculated exactly from the Master equation 8. In the Supporting Text S1 we describe the method used; there we also discuss the full induction time probability distributions.
Fig. 4A shows results from both the deterministic and the stochastic model. Plotted is the induction time as a function of the fold change. We assume that the maximal expression level of the circuit is prescribed by the functional context of the circuit; therefore, we vary but keep fixed. The deterministic model predicts that the induction time increases mildly–less than two-fold–as is increased from 10 to 1000. Based on such results, one might conclude that the effect of auto-activation on the induction time is mild. The stochastic model, however, reveals a different picture. When is low, both models are in agreement. But when (as we explain below), the stochastic model deviates dramatically from the trend predicted by the deterministic theory, demonstrating that the induction time is much more strongly affected by the auto-activation than expected from deterministic rate equations [2], [3], [7], [9].
What causes the discrepancy between the deterministic model and its stochastic counterpart? When the fold change is increased while the maximal transcription rate is kept fixed, the basal transcription rate is reduced. As a result, the expected number of TFs present in the cell at time –just before the signal arrives–becomes small. This means that, when the signal is introduced, the TF concentration is initially too low to activate the TF's transcription significantly, so that the transcription rate remains of the order . Crucially, this means that the expected waiting time before the first transcription event occurs is close to . Using realistic parameters for bacteria, in which the maximal transcription rate is of the order , this waiting time can easily become large. Indeed, Fig. 4A shows that for large the induction time is of order , indicating that the first transcription events become the limiting step of the induction. This effect is not accounted for by the deterministic model, which disregards the discreteness of the molecular events.
To visualize the process, Fig. 4B shows five representative time traces of the induction, obtained by kinetic Monte Carlo simulations, for a circuit with a large fold change . Before the signal is switched on, the copy number fluctuates around the average value . The signal is introduced at time . The traces clearly show that the induction time is dominated by the waiting time before the first transcription event; next, the circuit usually switches on rapidly. This has another important consequence: because transcription is (modeled as) a Poisson process, the probability distribution of induction times is approximately exponential (see Supplementary Text S1). The standard deviation of the induction times is therefore as large as the mean. This means that, in this parameter regime, the induction process is not only slow, but also unpredictable.
The anomalous induction time ultimately results from a low basal transcription rate . However, assuming that the maximal expression level of the TF is set by its biological function, so that can be considered given, this directly leads to constraints on . Fig. 4C is a contour plot of the induction time according to the stochastic model. It clearly shows that large fold changes lead to long induction times. Also, unless the fold change is small, increasing the Hill coefficient strongly slows down induction; comparison to Fig. 2 shows that this is because the circuit then approaches and eventually enters the (deterministically) bistable regime. If we arbitrarily decide that the induction time should not exceed 50 min, the admissible values of and are limited to the small shaded region.
To be precise, we distinguish between bistability and bimodality. We call a circuit bistable if the deterministic model predicts two stable steady states, and bimodal if the stochastic model predicts a steady state probability distribution with two peaks. Naively, one might expect that, in order for a circuit to be bimodal, it should be bistable. Theoretical work has shown, however, that this correspondence does not always hold [24], [26], [28], [29]. In particular, stochastic models of auto-activation circuits can produce bimodality even if the auto-regulation is non-cooperative (), in which case the circuit cannot be bistable (see Fig. 2) [24], [26]. Recently, this theoretical result was verified experimentally using a synthetic auto-activation circuit in Saccharomyces cerevisiae [27]. As we will explain, this phenomenon is directly related to the slow induction time we just described.
We explained that the slow induction was due to the fact that, if the circuit is prepared in a state with no TFs (), it on average lingers in that state for a time , which becomes large if is large. Obviously, the same lingering time holds if the circuit arrives in a state with by a rare random fluctuation. Therefore, as is increased, the stochastic circuit will spend an increasing fraction of its time in the state, which for large enough results in a peak in the steady-state distribution at . (In fact, in the limit the state becomes an absorbing state, so that the steady-state probability distribution becomes concentrated entirely on .) If another peak is present at non-zero expression levels, a bimodal distribution results. In other words, the anomalous bimodality and the slow induction are different manifestations of the same underlying characteristics of the circuit.
It can readily be derived from the Master equation 8 that the bimodality for non-cooperative auto-activating circuits requires that (see Supporting Text S1). In other words, in the non-cooperative circuit, bimodality is obtained only if the basal transcription rate is smaller than the protein degradation/dilution rate . This holds independent of , or ; moreover, more detailed stochastic models yield the same result (see supplementary text of Ref. [27]). Given the relation between the anomalous induction time and bimodality, this explains why the induction time of the stochastic model deviates noticeably from the deterministic one when ≳ α/β, as we indeed observed in Fig. 4A.
The above interpretation also indicates that a large burst size is not required to obtain bimodality at [27]. Bimodality can occur at any burst size, provided the fold change is large enough, such that (or, equivalently, provided the basal expression level is low enough such that ). Yet, the burst size can be important in an indirect way: to maintain a fixed steady state expression level, an increased burst size has to be compensated by a decreased or an increased . In both cases the requirement is relaxed.
In models that explicitly treat the binding of the TF to its own promoter, bimodality can also occur due to a different mechanism [21], [25]. If the binding kinetics of the TF are slow, the steady state distribution can have two peaks: one corresponding to the expression when a TF is bound to the promoter, and one corresponding to the expression when the promoter is unbound. For this type of bimodality the duality with an anomalous induction does not necessarily hold.
We have assumed that signaling circuits generally care about speed and sensitivity, and should avoid bistability and hysteresis. Each of these properties imposes constraints on the auto-regulation by the TF, as shown in the phase diagrams Figs. 2, 3D and 4C. In Fig. 5 we combine these results to analyze which parameters are compatible with all these constraints. To eliminate bistability, the operating point of the circuit should be below the black solid line. To obtain sensitivity, it should be close to this line. To avoid a slow induction, the fold change should not be too large; yet, to have any benefit from the auto-activation, it should not be too small. These constraints restrict the system to the small parameter region indicated in the plot. In this region, the fold change is at most moderate (≲ 40), and the Hill coefficient is roughly in the range 1 to 2. Such a low Hill coefficient can be achieved using auto-regulation by a single TF dimer.
We analyzed the properties of inducible auto-activation circuits to find parameter regions that are compatible with the requirements of signalling systems. For that reason, we studied bistability, sensitivity and induction speed, and discovered several new phenomena.
First of all, we found that auto-activation circuits can create an ultra-sensitive threshold response, even in the absence of molecular cooperativity. This conclusion holds for both definitions of sensitivity that we studied. To achieve similar levels of sensitivity in open-loop circuits, the promoters of all target genes have to be very sensitive, requiring many cooperative TF binding sites. As this is not feasible for every TF, positive feedback seems an excellent way to greatly increase the sensitivity of the response of all target promoters through the construction of a single binding site.
However, auto-regulation comes at a cost. We demonstrated that the induction time is much more severely affected by auto-activation than previously appreciated. When the basal transcription rate of the TFs promoter is small, ordinary rate equations fail and stochastic models should be used. In this regime, the waiting time until the first transcription event takes place becomes rate limiting, which makes the induction both slow and unpredictable. This effect imposes strong constraints on the use of auto-regulation for signaling. As a rule of thumb, the basal transcription rate should be such that the average waiting time for the first transcription event, , is safely below the required induction time. To indicate the severity of this limitation we note that if and , the response time will be more than . We therefore expect that this new stochastic effect will be relevant under fairly typical bacterial conditions.
Together, the constraints imposed by speed, sensitivity and bistability restrict the parameters to the small shaded area delineated in Fig. 5. Of course, the exact position of the borders of this area depend on somewhat arbitrary choices. If a narrow domain of bistability can be tolerated, somewhat higher Hill coefficients can be admitted. The induction time required in real circuits presumably varies, and therefore the restrictions on the fold change will vary too. Nevertheless, Fig. 5 clearly illustrates the trade-offs that may shape the parameters of response circuits.
For simplicity, we have discussed the sensitivity and the bistability of the circuit in terms of the deterministic model. However, similar results can be obtained in the framework of the stochastic model. In the Supporting Text S1, we demonstrate that the threshold response obtained in the deterministic model is preserved in the stochastic one. There, we also show that the bistable region of parameter space is virtually identical to the region that produces bimodality in the stochastic model–that is, apart from the anomalous bimodality at that we discussed above. (How the stability of the deterministic steady states is affected by noise has been discussed in detail in Ref. [10].)
With these ideas in mind, we revisit the TCS systems of E. coli, which we discussed briefly in the introduction. In order to model any specific TCS system quantitatively, the general models that we presented have to be extended to include the complex details of the signal transduction and the expression of the sensor kinase [3], [18], [30]. Nevertheless, the constraints that we derived should also affect TCS systems. To our knowledge, the E. coli genome contains 26 RRs, 14 of which are believed to auto-regulate. Table 1 lists these auto-regulating RRs. From this table, it is clear that a significant fraction of RRs activate their own expression level. Table 1 also mentions the number of binding sites found for the RR at its own promoter. Indeed, it appears that, in the well-studied cases, the auto-activation is typically mediated by a single (usually dimeric) binding site; examples include BaeR, CusR, PhoB, PhoP and ZraR. This is compatible with our predictions, because a dimeric binding site should yield a Hill coefficient between 1 and 2, depending on the dimerization kinetics. Note that the RRs that inhibit their own expression have more binding sites.
Our other prediction is that the maximal fold-change of the auto-activating signaling circuits should not be high. Unfortunately, the fold changes of most TCS systems are not accurately known. One problem is that the native signals of many TCS systems are unknown. In addition, auto-activation may only be observed at a high signal level [30]. Also, if the introduction of the signal affects bacterial growth, global physiological effects have to be accounted for [31]. A subtle point is that the fold change , defined in terms of the regulatory function , is generally not equal to the relative change in the steady state expression levels at and , because even at the circuit generally does not operate at the maximal transcription level . We can therefore provide only rough estimates. For PhoP a modest 10-fold increase in expression is reported between stimulated and unstimulated conditions [32]. Between phosphate-rich and phosphate-poor conditions, PhoB changes 12-fold according to Ref. [33], and 40-fold according to Ref. [34]. BaeR regulates its own operon, mdtABCD-baeSR, but the effect of BaeR amplification on baeS is only 10-fold [35], [36]; in addition, the putative inducer indole increases BaeR expression only 1.6 fold [36]. ZraR expression increases “significantly” in the presence of , but a quantitative measurement of the fold change is, to our knowledge, not available [37]. In all these cases the basal expression is non-negligible and the fold change seems to be low to moderate, as expected from our analysis.
Few experiments have directly measured the induction time of signaling systems. An exception is the PhoP/PhoQ two-component system of Salmonella [38]–[40]. The PhoP response regulator auto-activates by binding to a single dimeric binding site. After induction, the mRNA level of PhoP needs to reach its maximum value, which is about 20–30-fold higher than its basal level (depending on the signal); however, it takes about 30–40 min for the concentration of the protein PhoP to reach 50% of its maximum value [40]. These numbers agree well with our prediction in Fig. 5.
Induction times have also been measured for an entirely different class of auto-activating circuits: the type II restriction-modification (R-M) systems. R-M systems function as a defense against bacteriophages and are pervasive in bacteria; several thousands of putative R-M systems have been found [41]. These plasmid-borne systems consist of a restriction endonuclease (REase) that specifically cleaves DNA, and a methyltransferase (MTase) that methylates the same sequence and thereby protects it against the REase. Some R-M systems contain a third gene, called the C gene, which codes for a TF. C is co-transcribed with the REase, and in many examples regulates its own expression. For example, in the PvuII system, C binds to two dimeric binding sites, and , from which it respectively activates and represses its own expression [42]–[44]. The repressor site is much weaker than the activator site; it becomes relevant only at high C concentrations [44]. The auto-activation is believed to be important for the horizontal transfer of the plasmid between bacteria [42], [45]. Upon entering a cell, it is crucial that the MTase is expressed before the REase, to prevent the REase from damaging the new host's DNA. Indeed, auto-activation by C can provide such a delay. Experiments show that the induction of the C and REase proteins after entering the cell takes about 30 min [43]–a modest 10 min longer than the constitutively expressed MTase–and that the fold-change of auto-activation is ≳ 25 [43]. Despite the fact that a short delay is actually beneficial in this system, these numbers are compatible with the predictions in Fig. 5. The basal expression that is required to control the delay is provided by a separate, constitutive promoter [45], which suggest that natural selection has favored a short and predictable delay.
How can a cell reduce the induction time of an auto-activation circuit? Obviously, for a fixed fold change, the induction time can be decreased by using a high maximum transcription rate . However, this would also result in a high expression level at full activation. Using typical numbers for E. coli: at a burst size and a degradation/dilution rate of , a high transcription rate would lead to a steady state TF copy number of , which is exceptionally high for a TF. To compensate, a low burst size (weak ribosomal binding site) could be used, and/or active degradation of the TF. Another option is to cap the maximal expression level by implementing auto-repression at high TF concentrations, on top of auto-activation at lower TF levels [46] (as is often found in type II restriction-modification systems, discussed above [42]). However, because auto-repression reduces the steady state expression at full induction, it also limits the effective fold-change achieved. Each of the above measures obviously introduces additional overhead, the cost of which should be weighted against a higher basal transcription level or a slower induction.
Finally, our analysis provides several testable predictions. In particular, the strong dependence of the induction time on the basal expression level could be tested experimentally in modified versions of the PvuII circuit, or in a synthetic system, similar to the one presented in Ref. [27]. One way to vary the basal expression level of a synthetic auto-activation circuit is to express the TF from two independent promoters, one of which can be controlled. By tuning the expression from this promoter, its impact on the induction time can be studied. We hope that our analysis inspires such experiments to characterize the importance of stochasticity in constraining the design of signaling circuits.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.